Jensen Huang – Will Nvidia’s moat persist?

Full Transcript

https://www.youtube.com/watch?v=Hrbq66XqtCo

[00:00] We've seen the valuations of a bunch of software companies crash because people are expecting AI to commoditize software.
[00:08] And there's a a potentially naive way of thinking about things which is like look Nvidia sends a GDS2 file to TSMC.
[00:13] TSMC builds the logic dies.
[00:16] It builds the switches.
[00:18] Um then it packages them with the HBM that SK Highex and Micron and Samsung make.
[00:22] Then it sends it to an ODM in Taiwan where they assemble the racks.
[00:26] And so Nvidia is fundamentally making software that other people are manufacturing.
[00:28] And if software gets commoditized, does Nvidia get commoditized?
[00:32] Well, in the end, something has to transform electrons to tokens.
[00:38] That transformation um there's no the transformation of electrons to tokens uh and making those tokens more valuable over time.
[00:48] I I I don't I think that that that's hard to hard to um completely commoditize the transformation from electrons to
[01:00] The transformation from electrons to tokens is such an such an incredible journey and and making that token.
[01:07] You know, it's like making a one molecule more valuable than another molecule.
[01:11] Making one token more valuable than another.
[01:15] The amount of artistry, engineering, science, invention that goes into making that token valuable.
[01:21] Obviously, we're we're watching it happening in real time.
[01:22] And so, so the the the the transformation, the manufacturing, um all of the science that goes in there is far from un deeply understood and it's far from the journey is far from far from over.
[01:38] And so, so I I doubt that it will happen.
[01:41] Um we're going to make it more efficient, of course.
[01:42] I mean the whole the whole thing about Nvidia in fact the way that you frame the question is is my mental model of our company.
[01:49] The input is electron the output is tokens.
[01:55] That is in the middle Nvidia and our job is to to do as much as necessary as
[02:02] is to to do as much as necessary as little as possible to enable that.
[02:04] little as possible to enable that transformation to be done at incredible capabilities.
[02:07] transformation to be done at incredible capabilities and and what I mean by as little as possible.
[02:09] capabilities and and what I mean by as little as possible whatever I don't need to.
[02:11] little as possible whatever I don't need to.
[02:13] to I partner with somebody and I make it part of my ecosystem to do.
[02:14] I partner with somebody and I make it part of my ecosystem to do.
[02:16] part of my ecosystem to do. And if you look at Nvidia today, we probably have the largest ecosystem of partners both in supply chain upstream, supply chain downstream.
[02:18] look at Nvidia today, we probably have the largest ecosystem of partners both in supply chain upstream, supply chain downstream.
[02:21] the largest ecosystem of partners both in supply chain upstream, supply chain downstream.
[02:22] in supply chain upstream, supply chain downstream. all of the computers, computer companies and all the application developers and all the model makers and all the you know AI is a five five layer cake if you will and and we have ecosystems across the entire five layers and and so we try to do as little as possible but the part that we have to do as it turns out is insanely hard and and um.
[02:25] downstream. all of the computers, computer companies and all the application developers and all the model makers and all the you know AI is a five five layer cake if you will and and we have ecosystems across the entire five layers and and so we try to do as little as possible but the part that we have to do as it turns out is insanely hard and and um.
[02:26] computer companies and all the application developers and all the model makers and all the you know AI is a five five layer cake if you will and and we have ecosystems across the entire five layers and and so we try to do as little as possible but the part that we have to do as it turns out is insanely hard and and um.
[02:28] application developers and all the model makers and all the you know AI is a five five layer cake if you will and and we have ecosystems across the entire five layers and and so we try to do as little as possible but the part that we have to do as it turns out is insanely hard and and um.
[02:31] makers and all the you know AI is a five five layer cake if you will and and we have ecosystems across the entire five layers and and so we try to do as little as possible but the part that we have to do as it turns out is insanely hard and and um.
[02:34] five layer cake if you will and and we have ecosystems across the entire five layers and and so we try to do as little as possible but the part that we have to do as it turns out is insanely hard and and um.
[02:36] have ecosystems across the entire five layers and and so we try to do as little as possible but the part that we have to do as it turns out is insanely hard and and um.
[02:39] layers and and so we try to do as little as possible but the part that we have to do as it turns out is insanely hard and and um.
[02:42] as possible but the part that we have to do as it turns out is insanely hard and and um.
[02:45] do as it turns out is insanely hard and and um.
[02:46] and um >> I I don't think that that gets commoditized in fact in fact um.
[02:47] >>> I I don't think that that gets commoditized in fact in fact um.
[02:50] commoditized in fact in fact um >> uh I also don't think that the the enterprise software ware companies uh the tools makers you know most of the software companies today are tools makers um some of them are not um but.
[02:52] uh I also don't think that the the enterprise software ware companies uh the tools makers you know most of the software companies today are tools makers um some of them are not um but.
[02:55] enterprise software ware companies uh the tools makers you know most of the software companies today are tools makers um some of them are not um but.
[02:57] the tools makers you know most of the software companies today are tools makers um some of them are not um but.
[02:59] software companies today are tools makers um some of them are not um but.
[03:02] Makers, um, some of them are not, um, but are some of them are workflow.
[03:05] Are some of them are workflow, um, codification.
[03:07] Um, codification, you know, systems, um, but for a lot of companies, they're tool makers.
[03:10] You know, systems, um, but for a lot of companies, they're tool makers.
[03:11] For example, you know, Excel is a tool.
[03:13] PowerPoint's a tool, uh, Cadence makes tools.
[03:15] Synopsis makes tools.
[03:18] I, I actually see the opposite of what people see.
[03:21] I think the number of agents are going to grow exponentially.
[03:24] The number of agents are going to grow exponentially.
[03:27] The number of tool users are going to grow exponentially and it's very likely that.
[03:28] Exponentially and it's very likely that the number of instances of all these tools are going to skyrocket.
[03:32] The number of instances of all these tools are going to skyrocket.
[03:37] It is very likely the number of instances of Synopsys Design Compiler is going to skyrocket and the number of.
[03:39] It is very likely the number of instances of Synopsys Design Compiler is going to skyrocket and the number of number of agents that are going to be.
[03:41] Number of agents that are going to be using the floor planners and all of our layout tools and our design design rule checkers.
[03:45] Using the floor planners and all of our layout tools and our design design rule checkers.
[03:48] The number of agents that are today we're limited by the number of engineers.
[03:49] Today we're limited by the number of engineers.
[03:52] Tomorrow those engineers are going to be supported by a bunch of agents.
[03:54] We're going to be exploring out.
[04:04] Agents. We're going to be exploring out the design space like you've never seen.
[04:05] The design space like you've never seen explore before and want to use the tools that we use today.
[04:10] And so, so I think I think tool use is going to cause cause these software companies to skyrocket.
[04:14] The reason why it hasn't happened yet is because the agents aren't good enough at using their tools yet.
[04:20] And so either these companies are going to build the agents themselves or agents are going to get good enough to be able to use those tools.
[04:25] And I think it's going to be a combination of both.
[04:30] Um I think in your latest filings it was you had almost hundred billion dollars in purchase commitments with people foundries memory packaging and then uh semi analysis has reported that you will have $250 billion of these kinds of purchase commitments.
[04:45] And so one interpretation is Nvidia's mode is really that you've locked up many years of these scarce components that are uh you know somebody else might have an accelerator but can they actually get the memory to build it?
[04:56] Can they actually get the logic to build it?
[04:59] And this is really Nvidia's big mode for the next few years.
[05:01] Well, it it's one it's one of the things that we can do that is hard for someone
[05:04] that we can do that is hard for someone else to do.
[05:07] The reason why we could we we've made enormous commitments upstream.
[05:12] Um some of it is explicit.
[05:14] These commitments that you mentioned, some of it is implicit.
[05:17] Um, for example, a lot of the investments that are upstream are made by our our supply chain because I said to the CEOs, "Let me tell you how big this industry is going to be and let me explain to you why and let me reason through it with you and let me show you what I see."
[05:33] And so as a result of that that process of of uh informing inspiring um aligning with CEOs of all different industries upstream they're willing to make the investments.
[05:47] Now why are they willing to make the investments for me and not someone else and the reason for that is because they know that I have the capacity to buy it buy their supply and sell it through my downstream.
[06:01] the fact that Nvidia's downstream supply chain and our downstream demand is so large,
[06:07] and our downstream demand is so large, they're willing to make the investment upstream.
[06:13] And so if you look at GTC um and and uh you know, people are marveled by the scale of GTC and the people that go, it's a 360° that the entire universe of AI all in one place.
[06:24] and they they're all in one place because they need to see each other.
[06:28] I bring them together so that the the downstream could see the upstream.
[06:31] The upstream could see the downstream and all of them could see all the advances in AI and very importantly they can all meet the AI natives and all the AI startups that are all you know being being built and all the amazing things that are happening so that they could see firsthand all the things that I tell them.
[06:49] And so I spend a lot of my time informing directly or indirectly um our supply chain and our partners and our ecosystem about the opportunity that's that's in front of us.
[06:59] You know, most of my keynotes, you know, some some people always say, you know, Jensen in most keynotes, it's like one announcement after another announcement.
[07:08] Announcement after another announcement after another announcement after another.
[07:12] Announcement. Our keynotes are there's always a part.
[07:15] Our keynotes are there's always a part of it that's a little torturous in the sense that it's almost comes across like an ed like education and and in in fact.
[07:22] That's exactly on my mind. I need to make sure that the entire supply chain.
[07:27] Make sure that the entire supply chain upstream and downstream the ecosystem understands.
[07:32] Understands what is coming at us, why it's coming, when it's coming, how big is it going to be, and be able to reason about it systematically just like I reason about it.
[07:41] It. And and so so I think the the the the mode as you you describe it we're able to of course um build for a future.
[07:52] Able to of course um build for a future uh if our next next several years is a trillion dollars in in scale we have the supply chain to do it without our reach.
[08:02] Supply chain to do it without our reach the velocity of our business you know just as there's cash flow there's supply chain flow there turns uh nobody's going.
[08:10] chain flow there turns uh nobody's going to build a supply chain for an AR architecture if the architecture the architecture the business turns is low.
[08:17] And so our ability to sustain the scale is only because our downstream demand is so great and they see it and they all hear about it.
[08:24] They they see it all coming.
[08:26] And so that's it allows us to do the things that we're able to do at the scale we're able to do.
[08:32] I do want to understand more concretely whether the upstream can keep up.
[08:37] Um for many years now you guys have been 2xing revenue year-over-year.
[08:40] You guys have been more than tripling the amount of flops you're providing to the world year over year.
[08:44] and 2xing at the scale now is really incredible.
[08:47] Exactly.
[08:47] So then you look at logic say you're the biggest customer on TSMC's N3 node and um you're one of the biggest on uh AI as a whole this year is going to be 60% of N3.
[08:59] It's going to be 86% next year according to some analysis.
[09:01] How how do you 2x if you're the majority?
[09:07] Um and how do you do that year-over-year?
[09:09] So are we are we in a regime now where the
[09:11] are we are we in a regime now where the growth rate in AI compute has to slow?
[09:13] growth rate in AI compute has to slow because of upstream?
[09:15] Do you see a way to get around these you know you how do we build 2x more fabs year-over-year ultimately?
[09:21] ultimately?
[09:26] Yeah, at some at some level um the the instantaneous demand uh is greater than the supply upstream and downstream uh in the world.
[09:37] And and it could be at any instant any instance we could be limited by the number of plumbers.
[09:43] plumbers.
[09:44] Mhm.
[09:46] Which which actually happens.
[09:47] The plumbers are invited to next year's GTC.
[09:51] Yeah. You know, by the way, great idea.
[09:53] But that's a good condition. You you want you want you want a market you want an industry where the instantaneous demand is greater than the total supply of the industry.
[10:01] Um the opposite is obviously less good.
[10:03] If we're too far apart, uh if one particular item, one particular component is too far too far away, um obviously obviously the
[10:14] away, um obviously obviously the industry swarms it.
[10:17] So for example, notice people aren't talking very much about co-ass anymore.
[10:20] Yeah.
[10:22] And the reason for that is because for two years we swarmed a living daylights out of it and we double double double on on several doubles and and now I think we're in a fairly good shape.
[10:31] And TSMC now knows that co-ass supply has to keep up with the rest of the logic demand and the memory demand and and so so they're scaling co-ass um and their scaling uh you know future packaging technologies at the same level as a scale logic which is terrific because for a long time co-ass was rather specialty and um uh HBM was rather specialty but they're not specialties anymore people now realize they're mainstream computing technology.
[11:02] Um and and then and of course uh we're now much more able to influence a larger scope of our supply chain.
[11:10] In the past in the past um you know in the beginning of the AI revolution all the things that
[11:15] of the AI revolution all the things that I say now I was saying 5 years ago and I say now I was saying 5 years ago and some people believed in it and invested some people believed in it and invested in it.
[11:20] for example, uh, Sanjay and and in it.
[11:23] for example, uh, Sanjay and and the Micron team.
[11:25] the Micron team.
[11:25] I still remember the meeting really well where where I I was clear about exactly what's going to happen and why it's going to happen and and the predictions the predictions that that um of today and they they really doubled down on it and we partnered with them and uh across LPDDR across you know HBM memories uh they really invested in it and and it it it obviously has been tremendous for the company.
[11:28] meeting really well where where I I was clear about exactly what's going to happen and why it's going to happen and and the predictions the predictions that that um of today and they they really doubled down on it and we partnered with them and uh across LPDDR across you know HBM memories uh they really invested in it and and it it it obviously has been tremendous for the company.
[11:29] clear about exactly what's going to happen and why it's going to happen and and the predictions the predictions that that um of today and they they really doubled down on it and we partnered with them and uh across LPDDR across you know HBM memories uh they really invested in it and and it it it obviously has been tremendous for the company.
[11:31] happen and why it's going to happen and and the predictions the predictions that that um of today and they they really doubled down on it and we partnered with them and uh across LPDDR across you know HBM memories uh they really invested in it and and it it it obviously has been tremendous for the company.
[11:33] and the predictions the predictions that that um of today and they they really doubled down on it and we partnered with them and uh across LPDDR across you know HBM memories uh they really invested in it and and it it it obviously has been tremendous for the company.
[11:37] that um of today and they they really doubled down on it and we partnered with them and uh across LPDDR across you know HBM memories uh they really invested in it and and it it it obviously has been tremendous for the company.
[11:39] doubled down on it and we partnered with them and uh across LPDDR across you know HBM memories uh they really invested in it and and it it it obviously has been tremendous for the company.
[11:42] them and uh across LPDDR across you know HBM memories uh they really invested in it and and it it it obviously has been tremendous for the company.
[11:44] HBM memories uh they really invested in it and and it it it obviously has been tremendous for the company.
[11:47] it and and it it it obviously has been tremendous for the company.
[11:47] tremendous for the company.
[11:50] uh some some people came a little bit later and uh but they now they're all here and so I I think the each one of these generation each one of these bottlenecks gets a great deal of attention um and now we're we're prefetching the bottlenecks uh years in advance.
[11:52] people came a little bit later and uh but they now they're all here and so I I think the each one of these generation each one of these bottlenecks gets a great deal of attention um and now we're we're prefetching the bottlenecks uh years in advance.
[11:54] but they now they're all here and so I I think the each one of these generation each one of these bottlenecks gets a great deal of attention um and now we're we're prefetching the bottlenecks uh years in advance.
[11:57] think the each one of these generation each one of these bottlenecks gets a great deal of attention um and now we're we're prefetching the bottlenecks uh years in advance.
[11:59] each one of these bottlenecks gets a great deal of attention um and now we're we're prefetching the bottlenecks uh years in advance.
[12:02] gets a great deal of attention um and now we're we're prefetching the bottlenecks uh years in advance.
[12:04] now we're we're prefetching the bottlenecks uh years in advance.
[12:06] bottlenecks uh years in advance.
[12:06] So for example uh the the the investments that we've done uh with uh with Lum and Coherent and um all of the silicon photonix ecosystem uh the last several
[12:09] example uh the the the investments that we've done uh with uh with Lum and Coherent and um all of the silicon photonix ecosystem uh the last several
[12:12] we've done uh with uh with Lum and Coherent and um all of the silicon photonix ecosystem uh the last several
[12:15] Coherent and um all of the silicon photonix ecosystem uh the last several
[12:18] Photonix ecosystem, uh, the last several years, we really reshaped the ecosystem.
[12:20] Years, we really reshaped the ecosystem and the supply chain, silicon photonix.
[12:23] And the supply chain, silicon photonix.
[12:25] We, we, uh, built up an entire supply chain around TSMC.
[12:28] We partnered with them on coupe, uh, invented a whole bunch of technology.
[12:30] Technology, we licensed, uh, those patents to the supply chain.
[12:33] To the supply chain, keep it nice and open.
[12:35] Um, and so we're preparing the supply chain through invention of new technologies, new workflows, uh, new test, new testing equipment.
[12:37] New testing equipment, double-sided probing, um, investing in companies, helping them scale up their capacity.
[12:39] Companies, helping them scale up their capacity.
[12:42] Um, and so, so you could see that we're trying to shape the ecosystem so that it's ready, the supply chain so that it's ready to support the scale.
[12:43] It seems like some bottlenecks are easier than others.
[12:46] And so scaling up co-ass versus scaling up.
[12:48] I went to the hardest one by the way.
[12:51] Which is plumbers.
[12:53] Yeah, it's true.
[12:55] Yeah.
[12:57] Yeah, I actually went to the hardest one.
[12:58] Yeah.
[13:01] Yeah, plumbers and electricians.
[13:02] And the reason for that is because.
[13:04] Because and this is one of the concerns that I have about of all the doom the doomers, describing the end of end of.
[13:20] Doomers, um, describing the end of, end of work and killing of jobs.
[13:23] And you know, one of the things that that that um if we discourage people from being software engineers, we're going to run out of software engineers.
[13:30] And and uh the same prediction 10 years ago, some of the doomers were were uh uh saying that we're telling people whatever you do, don't be a radiologist.
[13:40] And you might hear some of those some of those videos are still on the web.
[13:44] You know, radiology is is going to be the first career to go.
[13:48] Nobody's the world's not going to need any more radiologists.
[13:50] Guess what?
[13:51] But we're short of radiologists.
[13:54] Oh, but okay.
[13:55] So, going back to this point about well some things you scale other things like how do you actually get how do you actually manufacture 2x the amount of logic a year?
[14:02] Ultimately that's bottleneck by memory and logic are bottleneck by UV.
[14:05] How do you get to 2x as many UV machines a year?
[14:09] Yeah.
[14:10] Year over year.
[14:10] None of that none of that's impossible to scale quickly.
[14:13] You just need to you you could do all of that is easy to do within two or three years.
[14:17] You just need a demand signal that it's
[14:21] You just need a demand signal that it's not it once you once you can build one.
[14:24] Not it once you once you can build one, you can build 10 and once you can build.
[14:25] You can build 10 and once you can build build 10, you can build a million and so.
[14:28] Build 10, you can build a million and so these things are not not hard to.
[14:29] These things are not not hard to replicate.
[14:31] How far down the supply chain do you go where you do you go to ASML?
[14:34] Do you go where you do you go to ASML and say hey, if I look out three years.
[14:35] And say hey, if I look out three years from now for me to for Nvidia to be generating two trillion in a year in.
[14:38] Generating two trillion in a year in revenue, we need way more AUV machines.
[14:40] Revenue, we need way more AUV machines and.
[14:41] And >> some of them I have to directly uh some.
[14:44] >> some of them I have to directly uh some of them are indirectly and some of them.
[14:46] Of them are indirectly and some of them um if I can convince TSMC as ASML will.
[14:49] Um if I can convince TSMC as ASML will be convinced and so that's that you know.
[14:51] Be convinced and so that's that you know we have to think about the critical.
[14:53] we have to think about the critical critical pinch points and uh but if TSMC.
[14:56] Is convinced uh you'll have plenty of EV machines in a few years.
[15:00] Machines in a few years. And so none of.
[15:03] That my point is that none of the bottlenecks last longer than a couple 2.
[15:05] Bottlenecks last longer than a couple 2 three years. None of them. And meanwhile.
[15:07] Three years. None of them. And meanwhile meanwhile we're uh improving computing.
[15:11] Meanwhile we're uh improving computing efficiency by 10x 20x in the case of.
[15:13] Efficiency by 10x 20x in the case of Hopper to Blackwell some 30 50x um we're.
[15:16] Hopper to Blackwell some 30 50x um we're coming up with new algorithms because.
[15:20] Coming up with new algorithms because.
[15:22] coming up with new algorithms because CUDA is so flexible.
[15:25] CUDA is so flexible.
[15:27] Uh we're we're developing all kinds of new techniques so that we drive efficiency.
[15:29] uh in addition to increasing capacity.
[15:31] Yeah.
[15:33] And so so there those those are those are things that that none of that worry me.
[15:36] It's the stuff that's downstream from us.
[15:38] Um energy policies that prevent energy from from you know you can't grow you can't create you can't create an industry without energy.
[15:44] You can't create a whole new manufacturing industry without energy.
[15:47] Uh we want to re-industrialize the United States.
[15:49] We want to bring back uh chip manufacturing and computer manufacturing and packaging and we want to build new things like EVs and robots and we want to build AI factories and you you can't build any of these things without energy and those things take a long time but more chip capacity that's a two threeear problem.
[16:11] more coass capacity 2 three year problem.
[16:13] interesting I I feel like I have guests tell me the exact opposite thing sometimes and I don't in this case I just don't have the technical knowledge to adjudicate but.
[16:20] well the beautiful thing is you're talking to the expert.
[16:21] Yeah, true, true.
[16:25] Talking to the expert, yeah, true, true.
[16:27] Um, okay. I want to ask about um your competitors.
[16:28] Competitors.
[16:28] Yeah.
[16:28] So, if you look at TPU, arguably two out of the top three models in the world, Claude and Gemini, were trained on TPU.
[16:39] What does that mean for Nvidia going forward?
[16:41] Forward?
[16:43] Um, well, we have we have a very different, we built a very different thing.
[16:44] Um, you know, what what Nvidia built is accelerated computing.
[16:51] Not a tensor processing unit.
[16:55] And uh accelerated computing is used for all kinds of things.
[16:57] You know, molecular dynamics and quantum chromodynamics and it's used for data processing, data frames, structured data, unstructured data.
[17:07] It's used for um fluid dynamics, particle physics, you know, and in addition, we use it for AI.
[17:17] And so accelerated computing is is um much more diverse and and although AI is the conversation today is obviously very
[17:25] The conversation today is obviously very important and impactful. Uh computing is important and impactful.
[17:29] Uh computing is much broader than that and what Nvidia has done is reinventing reinvented the way computing is done from general purpose computing to accelerate computing.
[17:34] Our market reach is far greater than any any TPU can any ASA can possibly have.
[17:46] And so if you look at our position, uh we're the only company that that accelerates applications of all kinds.
[17:54] We have a gigantic ecosystem and so all kinds of frameworks and algorithms all run on Nvidia.
[18:04] And because our computers are designed to be operated by other people, anyone who's an operator could buy our systems.
[18:13] Most of these homebuilt systems you have to be your own operator because it was never designed to be flexible enough for other people to operate.
[18:20] And so as a result of the fact that anybody can operate our systems, we're in every
[18:26] operate our systems, we're in every cloud including Google and Amazon and cloud including Google and Amazon and you know Azure and OCI and right and so whether you want to operate it to rent or operate it if you want to operate to rent you better have large ecosystem of customers in many industries that be the offtakers.
[18:42] if you're operating it if you if you want to operate it for yourself um we you know we obviously have the ability to help you operate yourself like for example for Elon with XAI and uh because we could we could enable operators uh in any any company in any industry you could use it uh to build a supercomput for uh scientific research and drug discovery at Lily and so we can help them operate their own supercomputer and and use it for the entire diversity of drug discovery and biological sciences um that that we accelerate >>
[19:20] and so so there there just you know a whole bunch of applications that we can address that you can't do so with TPUs
[19:28] address that you can't do so with TPUs because Nvidia's built CUDA as a fantastic tensor processing unit as well.
[19:34] but it does you know it does every every life cycle of data processing and computing and AI and so on so forth.
[19:41] and so I our our market opportunity is just a lot larger.
[19:43] Our reach is a lot greater and because we have such a large um we basically support every application in the world.
[19:53] now you could build Nvidia systems anywhere and know that there will be customers for it.
[19:58] and so it's a very different thing.
[20:00] Uh this is going to be sort of a long question but you know you have spectacular revenue um and this revenue is mostly you're not making 60 billion a quarter from uh pharma and um quantum.
[20:10] you're making it because AI is unprecedented technology that is growing unprecedentedly fast.
[20:14] and so then the question is what is best for AI specifically and I'm not in the details.
[20:18] but I talked to my AI researcher friends and they say look when I use a TPU it's this big systolic array that's perfect for doing matrix multiplies whereas a GPU is very flexible.
[20:25] It's great when you
[20:30] GPU is very flexible It's great when you have lots of branching when you have um
[20:33] have lots of branching when you have um irregular memory access but with these
[20:36] irregular memory access but with these you know what what is AI just like these
[20:37] you know what what is AI just like these very predictable matrix multiplies again
[20:39] very predictable matrix multiplies again and again and again and you don't have
[20:41] and again and again and you don't have to give up any die area for warp
[20:43] to give up any die area for warp schedulers for you know switches between
[20:45] schedulers for you know switches between threads and memory banks and so the TPU
[20:48] threads and memory banks and so the TPU is really optimized for the majority the
[20:50] is really optimized for the majority the bulk of this growth in revenue and use
[20:52] bulk of this growth in revenue and use case for uh compute that is coming
[20:54] case for uh compute that is coming online right now um yeah I I wonder how
[20:57] online right now um yeah I I wonder how you react to
[20:59] you react to um
[21:01] um matrix multiplies is an important part
[21:03] matrix multiplies is an important part of AI but it's not the only part of AI
[21:06] of AI but it's not the only part of AI and if you want to come up with a new
[21:08] and if you want to come up with a new attention mechanism or if you want to
[21:10] attention mechanism or if you want to disagregate in a different way if you
[21:13] disagregate in a different way if you want to come up with a whole new type of
[21:17] want to come up with a whole new type of architecture altogether for example you
[21:20] architecture altogether for example you know a hybrid SSM uh if you want to use
[21:23] know a hybrid SSM uh if you want to use a you want to create a model that that
[21:26] a you want to create a model that that um that fuses diffusion and auto
[21:30] um that fuses diffusion and auto reggressive somehow. Uh you you want an
[21:33] reggressive somehow. Uh you you want an architecture that's just generally
[21:35] architecture that's just generally programmable
[21:36] programmable and and we run everything you can
[21:40] and and we run everything you can imagine. And so that's the advantage. It
[21:42] imagine. And so that's the advantage. It allows for invention of new algorithms a
[21:45] allows for invention of new algorithms a lot more a lot a lot more easily.
[21:48] lot more a lot a lot more easily. >> And so because it's a programmable
[21:50] >> And so because it's a programmable system and and the ability to invent new
[21:53] system and and the ability to invent new algorithms is really what makes AI
[21:56] algorithms is really what makes AI advance. So quickly, you know,
[22:00] advance. So quickly, you know, TPUs like anything else is impacted by
[22:03] TPUs like anything else is impacted by Moore's law. And we know that Moore's
[22:05] Moore's law. And we know that Moore's law is increasing about 25% per year.
[22:08] law is increasing about 25% per year. And so the only way to really get 10x
[22:12] And so the only way to really get 10x leaps, 100x leaps,
[22:15] leaps, 100x leaps, is to fundamentally change the algorithm
[22:19] is to fundamentally change the algorithm and how it's computed every single year.
[22:22] and how it's computed every single year. >> And that's Nvidia's fundamental
[22:23] >> And that's Nvidia's fundamental advantage. The only reason why we were
[22:27] advantage. The only reason why we were able to make black well to hopper 50
[22:29] able to make black well to hopper 50 times, you know, I said it was 35 times
[22:32] times, you know, I said it was 35 times and and and when I first announced it
[22:34] and and and when I first announced it was going to black wall is going to be
[22:35] was going to black wall is going to be 35 times more energy efficient than
[22:38] 35 times more energy efficient than hopper. Uh nobody believed it and and uh
[22:42] hopper. Uh nobody believed it and and uh and then and then Dylan wrote an
[22:43] and then and then Dylan wrote an article. He said he said in fact in fact
[22:45] article. He said he said in fact in fact I sandbagged it's actually 50 times. And
[22:49] I sandbagged it's actually 50 times. And you can't reasonably do that with just
[22:50] you can't reasonably do that with just Moore's law. And so the the way that we
[22:54] Moore's law. And so the the way that we solve that problem is new out new models
[22:59] solve that problem is new out new models um uh parallelized and disagregated and
[23:02] um uh parallelized and disagregated and and distributed uh uh across a computing
[23:06] and distributed uh uh across a computing system uh and without the ability to
[23:10] system uh and without the ability to really get down and come up with new
[23:12] really get down and come up with new kernels with CUDA, it's really hard to
[23:14] kernels with CUDA, it's really hard to do and and so the combination of the
[23:18] do and and so the combination of the programmability of our of our
[23:20] programmability of our of our architecture
[23:21] architecture uh the the fact that Nvidia is an
[23:24] uh the the fact that Nvidia is an extreme codeesign company where we could
[23:27] extreme codeesign company where we could even offload some of the computation
[23:29] even offload some of the computation into the fabric itself, MVLink for
[23:31] into the fabric itself, MVLink for example into the network spectrum X um
[23:35] example into the network spectrum X um uh and that we could affect change
[23:38] uh and that we could affect change across the processors, the system, the
[23:43] across the processors, the system, the fabric, the libraries, the algorithm.
[23:47] fabric, the libraries, the algorithm. All of that was done simultaneously.
[23:49] All of that was done simultaneously. Without CUDA to do that, I wouldn't even
[23:51] Without CUDA to do that, I wouldn't even know where to start.
[23:53] know where to start. >> My sponsor Cruso was among the first
[23:54] >> My sponsor Cruso was among the first clouds to offer Nvidia's Blackwell and
[23:56] clouds to offer Nvidia's Blackwell and Blackwell Ultra platforms, and they just
[23:58] Blackwell Ultra platforms, and they just announced their Nvidia Vera Rubin
[24:00] announced their Nvidia Vera Rubin deployment scheduled for later this
[24:01] deployment scheduled for later this year. But access to state-of-the-art
[24:03] year. But access to state-of-the-art hardware is only part of the story. For
[24:05] hardware is only part of the story. For example, most inference engines already
[24:07] example, most inference engines already do KV caching for a single user's
[24:08] do KV caching for a single user's forward passes, but Cruso does it across
[24:11] forward passes, but Cruso does it across users and GPUs. So if a thousand agents
[24:13] users and GPUs. So if a thousand agents are running on the same system prompt,
[24:14] are running on the same system prompt, Cruso only has to compute the KV cache
[24:16] Cruso only has to compute the KV cache once for it to become available to every
[24:18] once for it to become available to every single GPU in the cluster. This is
[24:20] single GPU in the cluster. This is especially important as systems get more
[24:21] especially important as systems get more agendic and require much longer prefixes
[24:24] agendic and require much longer prefixes in order to use tools and access files.
[24:27] in order to use tools and access files. In a recent benchmark, Crusoe was able
[24:28] In a recent benchmark, Crusoe was able to deliver up to 10 times faster time to
[24:32] to deliver up to 10 times faster time to first token and up to five times better
[24:33] first token and up to five times better throughput than VLM. This is just one
[24:36] throughput than VLM. This is just one among many reasons that you should run
[24:37] among many reasons that you should run your inference workload with Cruso. And
[24:39] your inference workload with Cruso. And if you need GPUs for training, you don't
[24:41] if you need GPUs for training, you don't need to switch clouds. Cruso's got you
[24:42] need to switch clouds. Cruso's got you covered there, too. Go to
[24:43] covered there, too. Go to cruso.ai/torcashe
[24:45] cruso.ai/torcashe to learn more.
[24:47] to learn more. >> So, this gets at a interesting question
[24:49] >> So, this gets at a interesting question about um Nvidia's
[24:52] about um Nvidia's clientele where if 60% of your revenue
[24:55] clientele where if 60% of your revenue is coming from these big five
[24:58] is coming from these big five hyperscalers, you know, in a in in a
[25:01] hyperscalers, you know, in a in in a different era where different customers,
[25:02] different era where different customers, let's say it's professors who are
[25:03] let's say it's professors who are running experiments and they are helped
[25:05] running experiments and they are helped a bunch by they need CUDA. um they can't
[25:08] a bunch by they need CUDA. um they can't use another accelerator. They need to
[25:10] use another accelerator. They need to just run PyTorch with CUDA and have
[25:12] just run PyTorch with CUDA and have everything optimized. But if you got
[25:14] everything optimized. But if you got these hyperscalers, they have the
[25:15] these hyperscalers, they have the resources to write their own kernels. In
[25:17] resources to write their own kernels. In fact, they have to to get that extra
[25:18] fact, they have to to get that extra last 5% that they need for their
[25:21] last 5% that they need for their specific architecture. Um Anthropic,
[25:24] specific architecture. Um Anthropic, Google are mostly running their own
[25:26] Google are mostly running their own accelerators or running TPUs um and
[25:29] accelerators or running TPUs um and Tranium, but even OpenAI using GPUs has
[25:32] Tranium, but even OpenAI using GPUs has um has Triton which they're like we need
[25:35] um has Triton which they're like we need our own kernels. So they've um down to
[25:38] our own kernels. So they've um down to CUDA C++ they've instead of using Kublas
[25:41] CUDA C++ they've instead of using Kublas and Nickel and everything they've got
[25:43] and Nickel and everything they've got their own stack which compiles to other
[25:45] their own stack which compiles to other accelerators as well. Um and so if most
[25:47] accelerators as well. Um and so if most of your customers can can and do make
[25:51] of your customers can can and do make replacements for CUDA to what extent is
[25:53] replacements for CUDA to what extent is CUDA really the thing that is going to
[25:55] CUDA really the thing that is going to make Frontier AI happen on Nvidia? CUDA.
[25:59] make Frontier AI happen on Nvidia? CUDA. CUDA is um is a a rich ecosystem and so
[26:04] CUDA is um is a a rich ecosystem and so if you want to build on any computer
[26:06] if you want to build on any computer first, building on CUDA first is
[26:09] first, building on CUDA first is incredibly smart and because the
[26:12] incredibly smart and because the ecosystem is so rich uh we support every
[26:15] ecosystem is so rich uh we support every framework. uh if you want to create
[26:17] framework. uh if you want to create custom kernels uh if you need for
[26:20] custom kernels uh if you need for example we contribute enormously to
[26:22] example we contribute enormously to Triton and so the back end of Triton um
[26:25] Triton and so the back end of Triton um huge amounts of NVIDIA technology
[26:28] huge amounts of NVIDIA technology we're delighted to help every framework
[26:30] we're delighted to help every framework uh become as great as it can be and
[26:33] uh become as great as it can be and there's lots and lots of frameworks
[26:34] there's lots and lots of frameworks there's Triton there's VLM there's SG
[26:36] there's Triton there's VLM there's SG lang and then there's more right and now
[26:38] lang and then there's more right and now there's there's a whole bunch of new
[26:40] there's there's a whole bunch of new reinforcement learning frameworks coming
[26:42] reinforcement learning frameworks coming out you know you got Verl you got Nemo
[26:44] out you know you got Verl you got Nemo RL you got a whole bunch of new and then
[26:45] RL you got a whole bunch of new and then the the now with with with post-
[26:48] the the now with with with post- trainining and reinforcement learning
[26:50] trainining and reinforcement learning that entire area is just exploding right
[26:53] that entire area is just exploding right and so if you want to build on on an
[26:55] and so if you want to build on on an architecture building on a CUDA makes
[26:57] architecture building on a CUDA makes the most sense because you know that the
[26:58] the most sense because you know that the ecosystem is great you know that if
[27:01] ecosystem is great you know that if something happens it's more likely in
[27:03] something happens it's more likely in your code and not in the mountain of
[27:05] your code and not in the mountain of code underneath you know don't forget
[27:07] code underneath you know don't forget the amount of code that you're dealing
[27:08] the amount of code that you're dealing with when you're building these systems
[27:11] with when you're building these systems when something doesn't work was it you
[27:14] when something doesn't work was it you or was it the computer, you would like
[27:16] or was it the computer, you would like it always to be you and to to be able to
[27:19] it always to be you and to to be able to trust the computer and and you know,
[27:21] trust the computer and and you know, obviously we still have lots and lots of
[27:22] obviously we still have lots and lots of lots and lots of bugs ourselves, but but
[27:25] lots and lots of bugs ourselves, but but our system is so well rung out that you
[27:29] our system is so well rung out that you could at least build on top of the
[27:30] could at least build on top of the foundation. So that's number one is that
[27:32] foundation. So that's number one is that the richness of the ecosystem, the
[27:34] the richness of the ecosystem, the programmability of it, the capability of
[27:35] programmability of it, the capability of it. The second thing is is um if you
[27:38] it. The second thing is is um if you were a developer and you were building
[27:40] were a developer and you were building anything at all, the single most
[27:42] anything at all, the single most important thing you want more than
[27:43] important thing you want more than anything is install base. You want the
[27:45] anything is install base. You want the software that you run to run on a whole
[27:47] software that you run to run on a whole bunch of other computers. You don't want
[27:49] bunch of other computers. You don't want to build a software. You're not building
[27:50] to build a software. You're not building software just for yourself. You're
[27:52] software just for yourself. You're building software for your fleet or for
[27:54] building software for your fleet or for everybody else's fleet because you're a
[27:55] everybody else's fleet because you're a framework builder. And Nvidia's CUDA
[27:58] framework builder. And Nvidia's CUDA ecosystem is ultimately its great
[28:01] ecosystem is ultimately its great treasure. We are now I don't know how
[28:04] treasure. We are now I don't know how many several hundred million GPUs. Every
[28:07] many several hundred million GPUs. Every cloud has it goes back to A10, A100,
[28:11] cloud has it goes back to A10, A100, H100, H200,
[28:14] H100, H200, you know, the L series, the P series. I
[28:18] you know, the L series, the P series. I mean, there's a whole bunch of them and
[28:21] mean, there's a whole bunch of them and and they're they're they're in all kinds
[28:23] and they're they're they're in all kinds of sizes and shapes. And if you're a
[28:24] of sizes and shapes. And if you're a robotics company, you want that CUDA
[28:26] robotics company, you want that CUDA stack to actually run in the CUDA in the
[28:28] stack to actually run in the CUDA in the robot itself. We're literally
[28:29] robot itself. We're literally everywhere. And so the install base says
[28:32] everywhere. And so the install base says that once you develop the software, once
[28:34] that once you develop the software, once you develop the model, it's going to be
[28:36] you develop the model, it's going to be useful everywhere. And so the install
[28:38] useful everywhere. And so the install base is just too incredibly valuable.
[28:41] base is just too incredibly valuable. And then lastly, the fact that we're in
[28:43] And then lastly, the fact that we're in every single cloud makes us genuinely
[28:46] every single cloud makes us genuinely unique because you're an AI company and
[28:49] unique because you're an AI company and you're an AI developer. You're not
[28:51] you're an AI developer. You're not exactly sure which CSP you're going to
[28:53] exactly sure which CSP you're going to partner with and where you would like to
[28:54] partner with and where you would like to run it. And we'd run it everywhere,
[28:56] run it. And we'd run it everywhere, including on prem for you if you like.
[28:58] including on prem for you if you like. And so so I think that that the the
[29:02] And so so I think that that the the richness of the ecosystem, the
[29:05] richness of the ecosystem, the expansiveness of the of the of the
[29:08] expansiveness of the of the of the install base and the versatility of
[29:11] install base and the versatility of where where where we are, that
[29:13] where where where we are, that combination is is uh makes CUDA
[29:15] combination is is uh makes CUDA invaluable.
[29:16] invaluable. >> That makes a lot of sense. I guess I I
[29:17] >> That makes a lot of sense. I guess I I guess the thing I'm curious about is um
[29:20] guess the thing I'm curious about is um whether those advantages matter a lot to
[29:24] whether those advantages matter a lot to your main customers. um like there
[29:28] your main customers. um like there there's many people who who they might
[29:29] there's many people who who they might matter for for the kind of person who
[29:30] matter for for the kind of person who can actually build their own software
[29:32] can actually build their own software stack who are make up most of your
[29:33] stack who are make up most of your revenue um especially if you go to a
[29:35] revenue um especially if you go to a world where AI is getting especially
[29:37] world where AI is getting especially good at the things which have tight
[29:39] good at the things which have tight verification loops where you can RL on
[29:41] verification loops where you can RL on them and then this question of how do
[29:43] them and then this question of how do you write a kernel that does attention
[29:46] you write a kernel that does attention or MLP the most efficiently across a
[29:48] or MLP the most efficiently across a scale up it's a very verifiable sort of
[29:51] scale up it's a very verifiable sort of feedback loop and so oh can everybody
[29:54] feedback loop and so oh can everybody can all the hyperscalers write these
[29:55] can all the hyperscalers write these custom kernel for themselves. Um, and
[29:58] custom kernel for themselves. Um, and they might still Nvidia has uh still has
[30:01] they might still Nvidia has uh still has great price performance. So, they might
[30:02] great price performance. So, they might still prefer to use Nvidia. But then the
[30:04] still prefer to use Nvidia. But then the question is does it just become a
[30:05] question is does it just become a question of who is offering the best
[30:08] question of who is offering the best specs, the best um flops and memory and
[30:11] specs, the best um flops and memory and memory bandwidth for a given dollar
[30:13] memory bandwidth for a given dollar where historically Nvidia has just had
[30:14] where historically Nvidia has just had and still has you know the best margins
[30:17] and still has you know the best margins in all of AI across hardware and
[30:18] in all of AI across hardware and software 70% plus because of this CUDA
[30:21] software 70% plus because of this CUDA mode. And the question is, oh, can you
[30:23] mode. And the question is, oh, can you sustain those margins if for most of
[30:26] sustain those margins if for most of your customers they can actually afford
[30:28] your customers they can actually afford to build
[30:31] to build build instead of the CUDA mode. The
[30:34] build instead of the CUDA mode. The number of engineers we have assigned to
[30:35] number of engineers we have assigned to these AI labs is insane.
[30:38] these AI labs is insane. working with them, optimizing their
[30:39] working with them, optimizing their stack. And the reason for that is
[30:42] stack. And the reason for that is because because um nobody knows our
[30:44] because because um nobody knows our architecture better than we do. And
[30:46] architecture better than we do. And these architectures are not not as
[30:49] these architectures are not not as general purpose as a CPU. The reason the
[30:52] general purpose as a CPU. The reason the reason why a CPU is so, you know, a CPU
[30:54] reason why a CPU is so, you know, a CPU is kind of like like a Cadillac, you
[30:56] is kind of like like a Cadillac, you know, it's it just always, you know, it
[30:59] know, it's it just always, you know, it it's a nice cruiser. It never goes too
[31:01] it's a nice cruiser. It never goes too fast.
[31:03] fast. Everybody drives it pretty well. You
[31:05] Everybody drives it pretty well. You know, it's got cruise control. you know,
[31:08] know, it's got cruise control. you know, and everything is easy. But in a lot of
[31:11] and everything is easy. But in a lot of ways, Nvidia's GPUs are accelerators are
[31:14] ways, Nvidia's GPUs are accelerators are kind of like F1 racers. And yeah, I I
[31:18] kind of like F1 racers. And yeah, I I could imagine everybody's able to drive
[31:20] could imagine everybody's able to drive it at 100 100 miles an hour, but it
[31:23] it at 100 100 miles an hour, but it takes quite a bit of expertise to be
[31:24] takes quite a bit of expertise to be able to push it to the limit. And we use
[31:27] able to push it to the limit. And we use we use a ton of AI to create the kernels
[31:30] we use a ton of AI to create the kernels that we have. And um I'm pretty sure
[31:34] that we have. And um I'm pretty sure we're going to still be needed for quite
[31:35] we're going to still be needed for quite some time. And so our expertise um helps
[31:38] some time. And so our expertise um helps our our our um uh our AI labs partners
[31:43] our our our um uh our AI labs partners get another 2x out of their stack
[31:47] get another 2x out of their stack easily. Often times it's not unusual
[31:50] easily. Often times it's not unusual that we you know by the time that we're
[31:51] that we you know by the time that we're done optimizing their stack or
[31:53] done optimizing their stack or optimizing a particular kernel their
[31:55] optimizing a particular kernel their model sped up by 3x 2x 50%.
[32:00] model sped up by 3x 2x 50%. Um, that's a huge number, especially
[32:04] Um, that's a huge number, especially when you're talking about the installed
[32:05] when you're talking about the installed base of the fleet that they have of all
[32:07] base of the fleet that they have of all the hoppers and black walls that they
[32:09] the hoppers and black walls that they have. When you increase it by a factor
[32:11] have. When you increase it by a factor of two, that doubles the revenues. That
[32:15] of two, that doubles the revenues. That directly translates to revenues.
[32:17] directly translates to revenues. Nvidia's computing stack is the best
[32:20] Nvidia's computing stack is the best performance per TCO in the world, bar
[32:22] performance per TCO in the world, bar none.
[32:24] none. Nobody can demonstrate to me that any
[32:27] Nobody can demonstrate to me that any single platform in the world today has
[32:30] single platform in the world today has better performance TCO ratio. Not one
[32:33] better performance TCO ratio. Not one company. And in fact in fact the the uh
[32:36] company. And in fact in fact the the uh the benchmarks are out there uh Dylan's
[32:39] the benchmarks are out there uh Dylan's right inference max is sitting out there
[32:41] right inference max is sitting out there for everybody to to use and not one TPU
[32:44] for everybody to to use and not one TPU won't come trrenium won't come. I I
[32:47] won't come trrenium won't come. I I encourage them to
[32:50] encourage them to use inference max and demonstrate their
[32:53] use inference max and demonstrate their incredible
[32:54] incredible inference cost. It's really really hard.
[32:57] inference cost. It's really really hard. Uh not nobody wants to show up. Uh ML
[33:00] Uh not nobody wants to show up. Uh ML Perf I would I would welcome Trrenium to
[33:04] Perf I would I would welcome Trrenium to demonstrate their 40% that they claim
[33:06] demonstrate their 40% that they claim all the time. I would I would love to to
[33:08] all the time. I would I would love to to hear them demonstrate the the uh cost
[33:11] hear them demonstrate the the uh cost advantage of TPUs. It makes no sense in
[33:13] advantage of TPUs. It makes no sense in my mind. it makes absolutely zero sense
[33:16] my mind. it makes absolutely zero sense on first principles. It makes no sense.
[33:18] on first principles. It makes no sense. And so I I think the I think the the the
[33:21] And so I I think the I think the the the reason why we're so successful is simply
[33:24] reason why we're so successful is simply because our TCO is so great. There's a
[33:27] because our TCO is so great. There's a second you say um 60% of our customers
[33:31] second you say um 60% of our customers are the top five but most of that
[33:34] are the top five but most of that business is external. For example, most
[33:37] business is external. For example, most of AWS is most of Nvidia in AWS is for
[33:40] of AWS is most of Nvidia in AWS is for external customers not internal use.
[33:42] external customers not internal use. Most of our customers at Azure,
[33:44] Most of our customers at Azure, obviously all of our customers are
[33:45] obviously all of our customers are external. All of our customers at OCI
[33:47] external. All of our customers at OCI are external, not internal use. The
[33:49] are external, not internal use. The reason why they they favor us is because
[33:52] reason why they they favor us is because our reach is so great. We can bring them
[33:56] our reach is so great. We can bring them all of the great customers in the world.
[33:57] all of the great customers in the world. They're all built on Nvidia. And the
[33:59] They're all built on Nvidia. And the reason why all these C companies are
[34:00] reason why all these C companies are built on Nvidia is because our reach and
[34:02] built on Nvidia is because our reach and our versatility is so great. And so so I
[34:06] our versatility is so great. And so so I think I think the flywheel
[34:08] think I think the flywheel is is really install base the
[34:11] is is really install base the programmability of our architecture the
[34:14] programmability of our architecture the richness of our ecosystem and the fact
[34:16] richness of our ecosystem and the fact that there's so many AI companies in the
[34:18] that there's so many AI companies in the world there's tens of thousands of them
[34:20] world there's tens of thousands of them now
[34:21] now >> and if you were one of those AI startups
[34:24] >> and if you were one of those AI startups what architecture would you would you
[34:25] what architecture would you would you choose you would choose an architecture
[34:27] choose you would choose an architecture that's most abundant where the most
[34:29] that's most abundant where the most abundant in the world
[34:31] abundant in the world >> the one has the largest installed base
[34:33] >> the one has the largest installed base where the
[34:34] where the largest installed base and one that has
[34:36] largest installed base and one that has a rich ecosystem. And so that's the
[34:38] a rich ecosystem. And so that's the flywheel that that's the reason why
[34:39] flywheel that that's the reason why between the combination of one, our perf
[34:43] between the combination of one, our perf per dollar is so great um that that uh
[34:47] per dollar is so great um that that uh uh they have the lowest cost tokens.
[34:49] uh they have the lowest cost tokens. Second, our perf per watt is the highest
[34:52] Second, our perf per watt is the highest in the world. And so if if uh uh one of
[34:55] in the world. And so if if uh uh one of these companies if our partners built a
[34:58] these companies if our partners built a 1 gawatt data center that 1 gawatt data
[35:01] 1 gawatt data center that 1 gawatt data center better deliver the maximum amount
[35:04] center better deliver the maximum amount of revenues that and number of tokens
[35:07] of revenues that and number of tokens which directly translates to revenues
[35:09] which directly translates to revenues you wanted to generate as many tokens as
[35:10] you wanted to generate as many tokens as possible maximize the revenues for that
[35:12] possible maximize the revenues for that data center. We have the highest tokens
[35:15] data center. We have the highest tokens per watt architecture in the world. And
[35:17] per watt architecture in the world. And then lastly if your goal is to rent the
[35:19] then lastly if your goal is to rent the infrastructure we have the most
[35:21] infrastructure we have the most customers in the world. M and so that's
[35:23] customers in the world. M and so that's the reason why the flywheel works.
[35:25] the reason why the flywheel works. >> Interesting. I I I guess the question
[35:27] >> Interesting. I I I guess the question comes down to what is the actual market
[35:30] comes down to what is the actual market structure here because even if there's
[35:31] structure here because even if there's other companies there could have been a
[35:33] other companies there could have been a world where there's tens of thousands of
[35:34] world where there's tens of thousands of AI companies uh that have roughly equal
[35:37] AI companies uh that have roughly equal share of compute but if even through
[35:39] share of compute but if even through these five hyperscalers really the
[35:41] these five hyperscalers really the people on Amazon using the computer
[35:44] people on Amazon using the computer anthropic openai um and these big big
[35:47] anthropic openai um and these big big foundation labs who who can themselves
[35:49] foundation labs who who can themselves afford and have the ability to make
[35:52] afford and have the ability to make different accelerators work um
[35:54] different accelerators work um >> no I I I think your your your assumption
[35:57] >> no I I I think your your your assumption is is um premise is wrong.
[35:58] is is um premise is wrong. >> Maybe um let me let me let me ask you a
[36:00] >> Maybe um let me let me let me ask you a slightly different question which is
[36:01] slightly different question which is >> come back and make me correct your your
[36:03] >> come back and make me correct your your your um your premise.
[36:04] your um your premise. >> Okay, let me just ask a different
[36:06] >> Okay, let me just ask a different question which is okay if everything
[36:08] question which is okay if everything >> but still make sure that make me come
[36:09] >> but still make sure that make me come back and okay and fix because it's just
[36:11] back and okay and fix because it's just too important to AI it's too important
[36:14] too important to AI it's too important to the future of science is too
[36:16] to the future of science is too important to the future of the industry
[36:18] important to the future of the industry that that premise
[36:20] that that premise >> the premise look let me just finish the
[36:22] >> the premise look let me just finish the question and then we can address it
[36:24] question and then we can address it together. Yeah.
[36:25] together. Yeah. >> So what do you think if if all these
[36:29] >> So what do you think if if all these things are true about uh price
[36:31] things are true about uh price performance and performance per watt etc
[36:33] performance and performance per watt etc are true why why do you think it is the
[36:35] are true why why do you think it is the case that say um anthropic for example
[36:38] case that say um anthropic for example just announced a couple days ago they
[36:40] just announced a couple days ago they have a multi- gigawatt deal with
[36:41] have a multi- gigawatt deal with Broadcom and uh Google for TPUs and
[36:44] Broadcom and uh Google for TPUs and majority of their compute obviously for
[36:46] majority of their compute obviously for Google it's um TPU majority comput so if
[36:48] Google it's um TPU majority comput so if I look at these big AI companies it
[36:50] I look at these big AI companies it seems like a lot of their there was some
[36:52] seems like a lot of their there was some point where it was all Nvidia
[36:54] point where it was all Nvidia and now it's not. And so I'm curious how
[36:58] and now it's not. And so I'm curious how to square
[37:00] to square if these things are true on paper, why
[37:01] if these things are true on paper, why are they going with other accelerators?
[37:03] are they going with other accelerators? >> Yeah, anthropic is is an is a unique
[37:06] >> Yeah, anthropic is is an is a unique instance um and not a trend. Uh without
[37:09] instance um and not a trend. Uh without an anthropic, why would there be any TPU
[37:12] an anthropic, why would there be any TPU growth at all?
[37:14] growth at all? It's 100% anthropic. Without anthropic,
[37:17] It's 100% anthropic. Without anthropic, why would there be any tranium growth at
[37:19] why would there be any tranium growth at all? It's 100% anthropic. And I think
[37:21] all? It's 100% anthropic. And I think that's fairly wellnown and well
[37:23] that's fairly wellnown and well understood. It's not that it's not that
[37:25] understood. It's not that it's not that there's an abundance of ASIC
[37:27] there's an abundance of ASIC opportunities.
[37:29] opportunities. There's only one anthropic,
[37:31] There's only one anthropic, >> but OpenAI deals with AMD. They're
[37:33] >> but OpenAI deals with AMD. They're building their own Titan accelerator.
[37:35] building their own Titan accelerator. >> Yeah. But they're mostly I we could all
[37:36] >> Yeah. But they're mostly I we could all acknowledge they're vastly Nvidia and
[37:39] acknowledge they're vastly Nvidia and and we're going to still do a lot of
[37:41] and we're going to still do a lot of work together.
[37:42] work together. >> Yeah. And we're not we're not I'm not
[37:46] >> Yeah. And we're not we're not I'm not offended by other people using something
[37:48] offended by other people using something else and trying things. If they don't
[37:51] else and trying things. If they don't try these other things, how would they
[37:52] try these other things, how would they know how good ours is, you know? And
[37:55] know how good ours is, you know? And sometimes you got to be reminded of it
[37:57] sometimes you got to be reminded of it and and um we we got to and we have to
[38:00] and and um we we got to and we have to continuously earn earn um uh the
[38:02] continuously earn earn um uh the position that we're in. Uh you there
[38:05] position that we're in. Uh you there always claims and look at the number of
[38:07] always claims and look at the number of AS6 that have been cancelled. Just
[38:10] AS6 that have been cancelled. Just because you're going to build an ASIC,
[38:11] because you're going to build an ASIC, you still have to build something
[38:12] you still have to build something better. than Nvidia.
[38:14] better. than Nvidia. And it's not that easy building
[38:16] And it's not that easy building something better than Nvidia. It's not
[38:17] something better than Nvidia. It's not sensible actually, you know. It's we
[38:20] sensible actually, you know. It's we Nvidia's got to be missing something.
[38:22] Nvidia's got to be missing something. Seriously, you know, and because our our
[38:24] Seriously, you know, and because our our scale, our velocity, we're the only
[38:27] scale, our velocity, we're the only company in the world that's cranking it
[38:28] company in the world that's cranking it out every single year. Big leaps every
[38:31] out every single year. Big leaps every single year.
[38:32] single year. >> I guess their logic is that, hey, it
[38:33] >> I guess their logic is that, hey, it doesn't need to be better. It just needs
[38:34] doesn't need to be better. It just needs to be not more than 70% worse because
[38:37] to be not more than 70% worse because they're paying you 70% margins.
[38:39] they're paying you 70% margins. >> No, no, no. Don't forget uh even an AS6
[38:42] >> No, no, no. Don't forget uh even an AS6 margin is really quite high. Nvidia's
[38:44] margin is really quite high. Nvidia's margin 6 70% let's say but an ASIC
[38:47] margin 6 70% let's say but an ASIC margin is 65.
[38:49] margin is 65. What are you really saving?
[38:51] What are you really saving? >> Oh, you mean from Broadcom or something?
[38:52] >> Oh, you mean from Broadcom or something? >> Yeah, sure.
[38:54] >> Yeah, sure. >> You got to pay somebody.
[38:55] >> You got to pay somebody. >> Yeah.
[38:56] >> Yeah. >> And so so I think the the ASIC margins
[38:58] >> And so so I think the the ASIC margins are are incredibly good from what I can
[39:01] are are incredibly good from what I can tell and and they believe it. They
[39:03] tell and and they believe it. They believe it so too. And so they're
[39:05] believe it so too. And so they're they're quite proud of their their
[39:07] they're quite proud of their their incredible ASIC margins. And so you ask
[39:10] incredible ASIC margins. And so you ask the question why.
[39:12] the question why. A long time ago we just didn't have the
[39:14] A long time ago we just didn't have the ability to do it.
[39:17] ability to do it. And and this is this is this is and at
[39:19] And and this is this is this is and at the time I at the time I didn't deeply
[39:23] the time I at the time I didn't deeply internalize how difficult it would be to
[39:27] internalize how difficult it would be to build a a foundation AI lab
[39:30] build a a foundation AI lab >> like OpenAI and Anthropic.
[39:33] >> like OpenAI and Anthropic. uh and the the fact that they needed
[39:37] uh and the the fact that they needed huge investments from the supplier
[39:39] huge investments from the supplier themselves. Uh we just weren't in a
[39:42] themselves. Uh we just weren't in a position to make the multi-billion
[39:43] position to make the multi-billion dollar investment into anthropic so that
[39:46] dollar investment into anthropic so that they could use our use our compute but
[39:49] they could use our use our compute but Google and and AWS were and they put in
[39:52] Google and and AWS were and they put in huge investments in the beginning so
[39:54] huge investments in the beginning so that anthropic um in return use their
[39:57] that anthropic um in return use their compute. uh we we just weren't in a
[39:59] compute. uh we we just weren't in a position to do so uh at the time. Nor
[40:02] position to do so uh at the time. Nor nor did I I would say my mistake is I
[40:06] nor did I I would say my mistake is I didn't deeply internalize that they they
[40:08] didn't deeply internalize that they they really had no other options that that
[40:11] really had no other options that that that a VC would never put in 510 billion
[40:15] that a VC would never put in 510 billion of investment into an AI lab with the
[40:18] of investment into an AI lab with the with the hopes of it turning out to be
[40:20] with the hopes of it turning out to be anthropic. And so that was my miss. Uh
[40:24] anthropic. And so that was my miss. Uh but even if I understood it, I don't
[40:26] but even if I understood it, I don't think we would have been in a position
[40:27] think we would have been in a position to do that at the time. But um I'm not
[40:30] to do that at the time. But um I'm not going to make that same mistake again.
[40:32] going to make that same mistake again. And and um uh I'm delighted to invest in
[40:35] And and um uh I'm delighted to invest in OpenAI and and um I'm delighted to to uh
[40:39] OpenAI and and um I'm delighted to to uh help them scale and I believe it's
[40:41] help them scale and I believe it's essential to do so. And then and then
[40:44] essential to do so. And then and then when um uh when I was able to uh anth
[40:47] when um uh when I was able to uh anth when Anthropic came to us, I'm delighted
[40:49] when Anthropic came to us, I'm delighted to be an investor, delighted to help
[40:52] to be an investor, delighted to help them scale and um uh but we just weren't
[40:55] them scale and um uh but we just weren't at at the time able to do so.
[40:57] at at the time able to do so. >> If I if I could uh rewind everything, uh
[41:01] >> If I if I could uh rewind everything, uh Nvid Nvidia could have been as big back
[41:03] Nvid Nvidia could have been as big back then as we are now, I would have been
[41:05] then as we are now, I would have been more than happy to do it. This is this
[41:07] more than happy to do it. This is this is actually quite interesting which is
[41:09] is actually quite interesting which is um for many years Nvidia has been this
[41:12] um for many years Nvidia has been this um the company in AI making money making
[41:16] um the company in AI making money making lots of money and um now you're
[41:19] lots of money and um now you're investing it it's been reported that
[41:21] investing it it's been reported that you've done up to 30 billion in open AI
[41:23] you've done up to 30 billion in open AI and 10 billion in um anthropic um but
[41:27] and 10 billion in um anthropic um but now their valuations have increased and
[41:28] now their valuations have increased and I'm sure they'll continue to increase um
[41:30] I'm sure they'll continue to increase um and so if over overall these many years
[41:33] and so if over overall these many years you know you were giving them the
[41:34] you know you were giving them the compute you saw where yeah was headed
[41:36] compute you saw where yeah was headed and then they were worth like onetenth
[41:38] and then they were worth like onetenth what they are now a couple years ago or
[41:39] what they are now a couple years ago or even a year ago in some cases um and you
[41:42] even a year ago in some cases um and you had all this cash
[41:46] there there's a world where either
[41:47] there there's a world where either Nvidia themselves becomes a foundation
[41:49] Nvidia themselves becomes a foundation lab um does the huge investment to make
[41:52] lab um does the huge investment to make that possible or has made the deals
[41:54] that possible or has made the deals you've made now at current valuations
[41:56] you've made now at current valuations much earlier on um and you had the cash
[41:58] much earlier on um and you had the cash to do it so I am curious actually why
[42:00] to do it so I am curious actually why not have done it earlier
[42:02] not have done it earlier >> we did it as soon as we could
[42:05] >> we did it as soon as we could We did it as soon as we could have and
[42:07] We did it as soon as we could have and and and um if I could have, I would have
[42:10] and and um if I could have, I would have done it even earlier. Um at the time
[42:13] done it even earlier. Um at the time that Anthropic needed us to do it, we
[42:15] that Anthropic needed us to do it, we just weren't in a position to do it. It
[42:17] just weren't in a position to do it. It wasn't it wasn't, you know, it wasn't in
[42:19] wasn't it wasn't, you know, it wasn't in our sensibility to do so. How's that
[42:21] our sensibility to do so. How's that like a cash thing or just
[42:23] like a cash thing or just >> Yeah, the level of investment, you know,
[42:25] >> Yeah, the level of investment, you know, we never invested outside the company at
[42:27] we never invested outside the company at the time and not that much and um
[42:32] the time and not that much and um and we didn't realize we needed to,
[42:35] and we didn't realize we needed to, you know, I always I always thought that
[42:37] you know, I always I always thought that they could just go raise VCs for God's
[42:39] they could just go raise VCs for God's sakes like like all companies do. Um but
[42:43] sakes like like all companies do. Um but but um uh what they were trying to what
[42:46] but um uh what they were trying to what they were were trying to do uh couldn't
[42:49] they were were trying to do uh couldn't have been done through VCs. What OpenAI
[42:52] have been done through VCs. What OpenAI wanted to do couldn't have been done
[42:53] wanted to do couldn't have been done through VCs. And and I recognize that
[42:55] through VCs. And and I recognize that now. I didn't know it then, you know,
[42:57] now. I didn't know it then, you know, but that's their genius. That's why
[42:59] but that's their genius. That's why they're smart,
[43:00] they're smart, >> you know, and so so they realized they
[43:02] >> you know, and so so they realized they realized it then that they had to do
[43:03] realized it then that they had to do something like that. And I'm delighted
[43:05] something like that. And I'm delighted that they did, you know, and and even
[43:07] that they did, you know, and and even though even though um we we caused
[43:11] though even though um we we caused Anthropic to have to go to somebody
[43:13] Anthropic to have to go to somebody else, um I'm still happy that it
[43:15] else, um I'm still happy that it happened. Anthropic's existence is great
[43:18] happened. Anthropic's existence is great for the world. I'm I'm delighted for it.
[43:21] for the world. I'm I'm delighted for it. >> Uh I guess you still are making a ton of
[43:23] >> Uh I guess you still are making a ton of money and you're making way more money
[43:24] money and you're making way more money um quarter after quarter.
[43:25] um quarter after quarter. >> It's still okay to have regrets. Um so
[43:29] >> It's still okay to have regrets. Um so then the question still arises okay well
[43:31] then the question still arises okay well now that we're here and you have all
[43:33] now that we're here and you have all this money that you keep making um what
[43:35] this money that you keep making um what should Nvidia be doing with it and
[43:37] should Nvidia be doing with it and there's one answer which says look
[43:38] there's one answer which says look there's this whole middleman ecosystem
[43:40] there's this whole middleman ecosystem that has popped up for converting um
[43:43] that has popped up for converting um capex into opex for these labs so that
[43:46] capex into opex for these labs so that they can rent compute um because the
[43:48] they can rent compute um because the chips are really expensive they make a
[43:50] chips are really expensive they make a lot of money over their lifetime through
[43:51] lot of money over their lifetime through because the models are getting better
[43:53] because the models are getting better the value that they generate their
[43:54] the value that they generate their tokens is increasing but they're
[43:55] tokens is increasing but they're expensive to set up Nvidia has the money
[43:58] expensive to set up Nvidia has the money to do the capex. So, and in fact, you
[44:00] to do the capex. So, and in fact, you are
[44:02] are you're it's been reported you're back
[44:03] you're it's been reported you're back stoping core. We have up to 6.3 billion
[44:05] stoping core. We have up to 6.3 billion and have invested 2B. Um, but yeah, why
[44:08] and have invested 2B. Um, but yeah, why why doesn't Nvidia become
[44:10] why doesn't Nvidia become a cloud themselves? Why doesn't become a
[44:12] a cloud themselves? Why doesn't become a hyperscaler themselves and run this
[44:13] hyperscaler themselves and run this computer out? You have all this cash to
[44:14] computer out? You have all this cash to do it.
[44:15] do it. >> This is a philosophy of the company and
[44:17] >> This is a philosophy of the company and and I think is wise. We should do as
[44:19] and I think is wise. We should do as much as needed as little as possible.
[44:23] much as needed as little as possible. And and what that means is the the work
[44:26] And and what that means is the the work that we do with building our our
[44:28] that we do with building our our computing platform. If we don't if we
[44:30] computing platform. If we don't if we don't do it, I genuinely believe it
[44:33] don't do it, I genuinely believe it doesn't get done. If we didn't take the
[44:35] doesn't get done. If we didn't take the risk that we take, if we didn't build
[44:37] risk that we take, if we didn't build MVLink the way we built, if we didn't
[44:38] MVLink the way we built, if we didn't build the whole stack, if we didn't
[44:40] build the whole stack, if we didn't create the ecosystem the way we did it,
[44:42] create the ecosystem the way we did it, if we didn't dedicate ourselves to 20
[44:44] if we didn't dedicate ourselves to 20 years of CUDA while losing money most of
[44:47] years of CUDA while losing money most of that time, if we didn't do it, nobody
[44:49] that time, if we didn't do it, nobody else would have done it.
[44:52] else would have done it. If we didn't create all the CUDA X
[44:53] If we didn't create all the CUDA X libraries so that they're all domain
[44:55] libraries so that they're all domain specific, you know, this is several a
[44:58] specific, you know, this is several a decade and a half ago, we pushed into
[45:01] decade and a half ago, we pushed into domain specific libraries because we
[45:03] domain specific libraries because we realized that if we didn't create these
[45:04] realized that if we didn't create these domain specific libraries, whether it's
[45:06] domain specific libraries, whether it's for ray tracing or image generation or
[45:09] for ray tracing or image generation or even the early works of AI, these
[45:11] even the early works of AI, these models, if we didn't create them for
[45:13] models, if we didn't create them for data processing, structure data
[45:14] data processing, structure data processing or vector data process, if we
[45:16] processing or vector data process, if we didn't create them, nobody would. And I
[45:19] didn't create them, nobody would. And I am completely certain of that. We
[45:21] am completely certain of that. We created a a library for computational
[45:24] created a a library for computational lithography called KU litho. If we
[45:26] lithography called KU litho. If we didn't create it, nobody would have.
[45:29] didn't create it, nobody would have. And so accelerated computing wouldn't
[45:31] And so accelerated computing wouldn't advance the way it has if we didn't do
[45:33] advance the way it has if we didn't do what we did. And and so we should do
[45:36] what we did. And and so we should do that. We should dedicate our company all
[45:38] that. We should dedicate our company all of our might wholeheartedly to go do
[45:40] of our might wholeheartedly to go do that. However, the world has lots of
[45:42] that. However, the world has lots of clouds. If I didn't do it, somebody show
[45:45] clouds. If I didn't do it, somebody show up. And so following the the recipe the
[45:48] up. And so following the the recipe the philosophy of doing as much as needed
[45:51] philosophy of doing as much as needed but as little as possible as little as
[45:54] but as little as possible as little as possible that philosophy exists in our
[45:57] possible that philosophy exists in our company today and everything I do I do
[45:59] company today and everything I do I do it with that lens
[46:02] it with that lens in the case of clouds if we didn't
[46:04] in the case of clouds if we didn't support coreweave to exist
[46:07] support coreweave to exist these neo clouds these AI clouds
[46:09] these neo clouds these AI clouds wouldn't exist if we didn't help
[46:12] wouldn't exist if we didn't help cororeweave exist they would not exist
[46:15] cororeweave exist they would not exist If we didn't support Nscale, they
[46:17] If we didn't support Nscale, they wouldn't be where they are today. If we
[46:19] wouldn't be where they are today. If we didn't support NBS, they wouldn't be
[46:21] didn't support NBS, they wouldn't be where they are today. Now, they are
[46:23] where they are today. Now, they are they're doing fantastically. Is that a
[46:26] they're doing fantastically. Is that a business model where no, we should do as
[46:28] business model where no, we should do as much as needed as little as possible.
[46:30] much as needed as little as possible. And so, we're trying we invest in our
[46:32] And so, we're trying we invest in our ecosystem because I want our eco
[46:35] ecosystem because I want our eco ecosystem to thrive. And I want our our
[46:38] ecosystem to thrive. And I want our our I want I want the architecture and I
[46:41] I want I want the architecture and I want AI to be able to connect with as
[46:44] want AI to be able to connect with as many industries as possible, as many
[46:48] many industries as possible, as many countries as possible and make it
[46:50] countries as possible and make it possible for you know the planet to be
[46:52] possible for you know the planet to be built on AI and to be built on the
[46:54] built on AI and to be built on the American tech stack. And so so th that
[46:56] American tech stack. And so so th that vision I think is exactly what we're
[46:59] vision I think is exactly what we're pursuing. Now, one of the things that
[47:00] pursuing. Now, one of the things that that you mentioned, um, there are so
[47:04] that you mentioned, um, there are so many great amazing foundation model
[47:05] many great amazing foundation model companies and we try to invest in all of
[47:07] companies and we try to invest in all of them. And this is this is another thing
[47:09] them. And this is this is another thing that we do. We don't pick winners and we
[47:12] that we do. We don't pick winners and we we like we we we need to support
[47:14] we like we we we need to support everyone and it's part of our part of
[47:17] everyone and it's part of our part of our our our joy of doing so. It's it's
[47:19] our our our joy of doing so. It's it's an imperative to our business, but we
[47:21] an imperative to our business, but we also go out of our way not to pick
[47:23] also go out of our way not to pick winners. And so when I when I invest in
[47:25] winners. And so when I when I invest in one of them, I invest in all of them.
[47:27] one of them, I invest in all of them. Why do you go out of your arena not to
[47:28] Why do you go out of your arena not to pick winners?
[47:29] pick winners? >> Because it's not our job to. Number one.
[47:32] >> Because it's not our job to. Number one. Number two, when Nvidia first started,
[47:35] Number two, when Nvidia first started, there were 60 graphics companies, 60 3D
[47:38] there were 60 graphics companies, 60 3D graphics companies, uh we are the only
[47:41] graphics companies, uh we are the only one that survived. If you would have
[47:42] one that survived. If you would have taken those 60 companies, 60 graphics
[47:46] taken those 60 companies, 60 graphics companies, and asked yourself which one
[47:47] companies, and asked yourself which one was going to make it,
[47:48] was going to make it, >> Nvidia would be the top of that list not
[47:51] >> Nvidia would be the top of that list not to make it. You know, this is long
[47:53] to make it. You know, this is long before you, but Nvidia's graphics
[47:56] before you, but Nvidia's graphics architecture was precisely wrong. It's
[47:59] architecture was precisely wrong. It's not a little bit wrong. We created an
[48:01] not a little bit wrong. We created an architecture that was precisely wrong.
[48:04] architecture that was precisely wrong. And and it was an impossible thing for
[48:06] And and it was an impossible thing for developers to support. It was never
[48:08] developers to support. It was never going to make it. We reasoned about it
[48:10] going to make it. We reasoned about it for good re for from good first
[48:12] for good re for from good first principles, but we ended up in the wrong
[48:14] principles, but we ended up in the wrong solution. and and um uh everybody would
[48:18] solution. and and um uh everybody would have kind everybody would have counted
[48:19] have kind everybody would have counted us out and and here we are. And so I'm
[48:23] us out and and here we are. And so I'm I'm I'm
[48:25] I'm I'm enough humility to recognize that, you
[48:27] enough humility to recognize that, you know, don't don't pick winners.
[48:29] know, don't don't pick winners. >> Yeah.
[48:30] >> Yeah. >> Um
[48:30] >> Um >> either let them all take care of
[48:32] >> either let them all take care of themselves or take care of all of them.
[48:34] themselves or take care of all of them. >> Um one thing I didn't understand is you
[48:37] >> Um one thing I didn't understand is you said, "Look, we're not prioritizing
[48:38] said, "Look, we're not prioritizing these neoclouds just because there are
[48:40] these neoclouds just because there are new clouds and we want to prop them up."
[48:42] new clouds and we want to prop them up." But you also said you listed a bunch of
[48:45] But you also said you listed a bunch of new clouds and you said they wouldn't
[48:46] new clouds and you said they wouldn't exist if it wasn't for Nvidia.
[48:47] exist if it wasn't for Nvidia. >> Yeah.
[48:47] >> Yeah. >> And so how are those two things
[48:50] >> And so how are those two things compatible?
[48:50] compatible? >> Um first of all they they need to want
[48:52] >> Um first of all they they need to want to exist and they come to ask us for
[48:54] to exist and they come to ask us for help. And when they when they um uh when
[48:57] help. And when they when they um uh when they want to exist and they have they
[48:59] they want to exist and they have they have a business plan and they you know
[49:01] have a business plan and they you know they have expertise and you know they
[49:02] they have expertise and you know they have the passion for it. Uh they
[49:05] have the passion for it. Uh they obviously have to have some capabilities
[49:07] obviously have to have some capabilities themselves. Uh but if at the end of the
[49:09] themselves. Uh but if at the end of the day they need some investment in order
[49:11] day they need some investment in order to get it off the ground, uh we we would
[49:13] to get it off the ground, uh we we would be there for them. Um but but the sooner
[49:16] be there for them. Um but but the sooner they get their flywheel going, you know,
[49:19] they get their flywheel going, you know, your question was do we want to be in
[49:21] your question was do we want to be in the financing business? The answer is
[49:22] the financing business? The answer is no.
[49:23] no. >> Yeah. We don't want to be we want to we
[49:25] >> Yeah. We don't want to be we want to we because there are people in the
[49:26] because there are people in the financing business and we rather work
[49:28] financing business and we rather work with all of the people who are in the
[49:30] with all of the people who are in the financing business than to be a
[49:31] financing business than to be a financeier ourselves. And so so I think
[49:34] financeier ourselves. And so so I think the the uh our goal is to focus on what
[49:36] the the uh our goal is to focus on what we do, keep our business model as simple
[49:38] we do, keep our business model as simple as possible, support our ecosystem. Um
[49:41] as possible, support our ecosystem. Um when someone like like uh Open AI needs
[49:44] when someone like like uh Open AI needs an investment of $30 billion scale um
[49:47] an investment of $30 billion scale um because it's still before their IPO and
[49:50] because it's still before their IPO and and uh u we deeply believe in them. Uh
[49:54] and uh u we deeply believe in them. Uh we deeply believe that I deeply believe
[49:56] we deeply believe that I deeply believe that that they're going to be they're
[49:58] that that they're going to be they're going to be an well they're an
[49:59] going to be an well they're an extraordinary company already today.
[50:00] extraordinary company already today. They're going to be incredible company.
[50:02] They're going to be incredible company. uh the world needs them to exist. The
[50:04] uh the world needs them to exist. The world wants them to exist. I want them
[50:06] world wants them to exist. I want them to exist and and uh they have everything
[50:08] to exist and and uh they have everything on they have the wind at their back.
[50:10] on they have the wind at their back. Let's let's support them and let them
[50:12] Let's let's support them and let them scale. And so so to those those
[50:14] scale. And so so to those those investments will do because we're they
[50:17] investments will do because we're they need us to do it. And um uh but we're
[50:20] need us to do it. And um uh but we're we're not trying to do as much as
[50:21] we're not trying to do as much as possible. We're trying to do as little
[50:22] possible. We're trying to do as little as possible.
[50:24] as possible. >> I spend way too much time copy pasting
[50:25] >> I spend way too much time copy pasting text back and forth from Google Docs to
[50:27] text back and forth from Google Docs to chatbots. And so I built what's
[50:29] chatbots. And so I built what's basically a cursor for writing which
[50:31] basically a cursor for writing which operates the way I think an AI
[50:32] operates the way I think an AI co-researcher should operate. I can tag
[50:34] co-researcher should operate. I can tag it and it can talk with me through
[50:36] it and it can talk with me through inline comment threads and help me dig
[50:38] inline comment threads and help me dig deeper and brainstorm. I wrote this
[50:39] deeper and brainstorm. I wrote this entire thing over the weekend with
[50:40] entire thing over the weekend with cursor and their new composer 2 model.
[50:42] cursor and their new composer 2 model. With a lot of agentic coding tools, I
[50:44] With a lot of agentic coding tools, I feel like I have no idea what's going on
[50:45] feel like I have no idea what's going on under the surface. I just have to
[50:47] under the surface. I just have to relinquish control and hope for the
[50:48] relinquish control and hope for the best. But cursor let me try a bunch of
[50:50] best. But cursor let me try a bunch of different ideas while staying on top of
[50:51] different ideas while staying on top of the implementation. I did most of my
[50:53] the implementation. I did most of my brainstorming in the agents window. And
[50:55] brainstorming in the agents window. And after I got some basic files in place, I
[50:57] after I got some basic files in place, I used a diff window to track changes. The
[50:59] used a diff window to track changes. The few times that I needed to make a quick
[51:00] few times that I needed to make a quick tweak by hand, I just used the editor.
[51:02] tweak by hand, I just used the editor. If you want to try my AI code researcher
[51:04] If you want to try my AI code researcher yourself, I've linked the GitHub repo in
[51:05] yourself, I've linked the GitHub repo in the description. And if you have a tool
[51:06] the description. And if you have a tool that you've been wanting to build, you
[51:08] that you've been wanting to build, you should make it happen. Go to
[51:09] should make it happen. Go to cursor.com/cash
[51:11] cursor.com/cash to get started.
[51:13] to get started. This may be sort of an obvious question,
[51:14] This may be sort of an obvious question, but we've lived many years in this
[51:18] but we've lived many years in this situation where there's a shortage of
[51:21] situation where there's a shortage of GPUs and it's grown now because models
[51:24] GPUs and it's grown now because models are getting better.
[51:25] are getting better. >> We have a shortage of GPUs.
[51:27] >> We have a shortage of GPUs. >> Yes.
[51:27] >> Yes. >> Yeah.
[51:28] >> Yeah. >> And
[51:30] >> And Nvidia is known for diving up the scarce
[51:33] Nvidia is known for diving up the scarce allocation not just based on highest
[51:35] allocation not just based on highest bidder but rather on hey we want to make
[51:37] bidder but rather on hey we want to make sure that these neo neo clouds exist.
[51:39] sure that these neo neo clouds exist. Let's give some to core. Let's give some
[51:41] Let's give some to core. Let's give some to Cruso. Well, let's give some to
[51:42] to Cruso. Well, let's give some to Lambda. Um, why is it good for Nvidia?
[51:45] Lambda. Um, why is it good for Nvidia? First of all, would you agree with this
[51:47] First of all, would you agree with this characterization of fracturing the
[51:48] characterization of fracturing the market?
[51:49] market? >> No. No. Yeah. Your premise is just
[51:51] >> No. No. Yeah. Your premise is just wrong.
[51:51] wrong. >> Yeah.
[51:52] >> Yeah. >> Um, we're we're sufficiently um mindful
[51:56] >> Um, we're we're sufficiently um mindful about these things. I We're very mindful
[52:00] about these things. I We're very mindful about these things. First of all, if you
[52:02] about these things. First of all, if you don't place an if you don't place a PO,
[52:06] all the talking in the world won't make
[52:08] all the talking in the world won't make a difference. And so until we get a PO,
[52:10] a difference. And so until we get a PO, what are we going to do? And so the
[52:13] what are we going to do? And so the first thing is is we work with we work
[52:15] first thing is is we work with we work really hard with everybody to get a
[52:17] really hard with everybody to get a forecast done because these things take
[52:20] forecast done because these things take a long time to build and the data
[52:21] a long time to build and the data centers take a long time to build and so
[52:24] centers take a long time to build and so we align ourselves um with demand and
[52:27] we align ourselves um with demand and supply and things like that through
[52:28] supply and things like that through forecasting. Okay, that's job job number
[52:31] forecasting. Okay, that's job job number one. Number two, um, everybody who, you
[52:34] one. Number two, um, everybody who, you know, we've tried to forecast with was
[52:36] know, we've tried to forecast with was with with as many people as possible,
[52:37] with with as many people as possible, but in the fin in the final analysis,
[52:39] but in the fin in the final analysis, you still had to place an order and
[52:41] you still had to place an order and maybe maybe um, for whatever reason, you
[52:45] maybe maybe um, for whatever reason, you didn't place your order, what can I do?
[52:47] didn't place your order, what can I do? And so at some point, first in first
[52:49] And so at some point, first in first out, but beyond that, if you're not
[52:52] out, but beyond that, if you're not ready because your data center is not
[52:55] ready because your data center is not ready or certain components aren't ready
[52:57] ready or certain components aren't ready to to enable you to stand up a data
[52:59] to to enable you to stand up a data center, um we might decide to serve
[53:02] center, um we might decide to serve another customer first. That's just
[53:05] another customer first. That's just maximizing the throughput of our of our
[53:06] maximizing the throughput of our of our our own factory.
[53:09] our own factory. And so uh we might do some adjustments
[53:11] And so uh we might do some adjustments there. Aside from that,
[53:15] there. Aside from that, uh the prioritization is is first in
[53:17] uh the prioritization is is first in first out.
[53:19] first out. >> Yeah. You gota you got to place a PO. If
[53:22] >> Yeah. You gota you got to place a PO. If you don't place a PO, now of course
[53:25] you don't place a PO, now of course there there's stories about that, you
[53:27] there there's stories about that, you know, like for example, all of this kind
[53:29] know, like for example, all of this kind of started from from uh it was a article
[53:33] of started from from uh it was a article about Larry and Elon having dinner with
[53:35] about Larry and Elon having dinner with me where they where they begged for
[53:36] me where they where they begged for GPUs.
[53:39] GPUs. >> That never happened. We had we
[53:42] >> That never happened. We had we absolutely had dinner. We absolutely had
[53:45] absolutely had dinner. We absolutely had dinner. Um and it was a it was a
[53:47] dinner. Um and it was a it was a wonderful dinner. In no time did they
[53:48] wonderful dinner. In no time did they beg for GPUs and so it they just had to
[53:52] beg for GPUs and so it they just had to place an order and once they place an
[53:54] place an order and once they place an order we do our best to get the capacity
[53:56] order we do our best to get the capacity to them. Yeah. We're not complicated.
[53:59] to them. Yeah. We're not complicated. >> Okay. So it sounds like there's a cue
[54:01] >> Okay. So it sounds like there's a cue and then um uh based on whether your
[54:04] and then um uh based on whether your data center is ready and when you place
[54:05] data center is ready and when you place a purchase order, you get them a certain
[54:07] a purchase order, you get them a certain time. But it still doesn't sound like
[54:10] time. But it still doesn't sound like highest bidder just gets it. Is there a
[54:12] highest bidder just gets it. Is there a reason to do it?
[54:13] reason to do it? >> We never do that.
[54:14] >> We never do that. >> Okay.
[54:15] >> Okay. >> We never do.
[54:15] >> We never do. >> Why not just do highest bidder?
[54:17] >> Why not just do highest bidder? >> Because it's it's a bad business
[54:18] >> Because it's it's a bad business practice. You you set your price. You
[54:20] practice. You you set your price. You set your price and then and then people
[54:22] set your price and then and then people decide to buy it or not. And and um uh
[54:26] decide to buy it or not. And and um uh there there I I understand that that
[54:31] there there I I understand that that others in the chip industry um uh change
[54:35] others in the chip industry um uh change their prices when demand is higher. Uh
[54:37] their prices when demand is higher. Uh but we just don't we just don't that's
[54:39] but we just don't we just don't that's just never been a practice of ours. You
[54:40] just never been a practice of ours. You can count on us, you know. I I prefer to
[54:43] can count on us, you know. I I prefer to be to be um uh dependable uh to be the
[54:48] be to be um uh dependable uh to be the foundation of the industry. And I you
[54:51] foundation of the industry. And I you don't need to you don't need to second
[54:52] don't need to you don't need to second guess.
[54:53] guess. >> You know, if if you if I quoted you a
[54:56] >> You know, if if you if I quoted you a price um we quoted you a price, that's
[54:58] price um we quoted you a price, that's it.
[54:59] it. >> And if demand goes through the roof, so
[55:01] >> And if demand goes through the roof, so be it.
[55:02] be it. >> And on the other end, that's why you
[55:03] >> And on the other end, that's why you have a productive relationship with
[55:04] have a productive relationship with TSMC, right?
[55:05] TSMC, right? >> Yeah. Yeah. Yeah. Uh Nvidia has been in
[55:08] >> Yeah. Yeah. Yeah. Uh Nvidia has been in business, we've been doing business with
[55:09] business, we've been doing business with them for uh I guess coming up on 30
[55:13] them for uh I guess coming up on 30 years and Nvidia and TSMC don't have a
[55:16] years and Nvidia and TSMC don't have a legal contract.
[55:18] legal contract. There's there is always some rough
[55:20] There's there is always some rough justice and um sometimes I'm right,
[55:23] justice and um sometimes I'm right, sometimes I'm wrong. Uh sometimes I got
[55:25] sometimes I'm wrong. Uh sometimes I got I got a better deal, sometimes I got a
[55:27] I got a better deal, sometimes I got a worse deal. Uh but overall in the in the
[55:30] worse deal. Uh but overall in the in the whole the relationship is incredible and
[55:32] whole the relationship is incredible and and I can completely trust them. I
[55:34] and I can completely trust them. I completely depend on them and and our
[55:36] completely depend on them and and our our one of the things that we you can
[55:38] our one of the things that we you can count on with Nvidia is that next year
[55:41] count on with Nvidia is that next year this year Ver Rubin is going to be
[55:43] this year Ver Rubin is going to be incredible. Next year Ver Rubin Ultra
[55:45] incredible. Next year Ver Rubin Ultra will come. The year after that Fman will
[55:47] will come. The year after that Fman will come and the year after that I haven't
[55:49] come and the year after that I haven't introduced the name yet. And so so every
[55:52] introduced the name yet. And so so every single year you can count on us.
[55:55] single year you can count on us. And this is an
[55:57] And this is an you you're going to have to go find
[55:58] you you're going to have to go find another ASIC team in the world. Pick
[56:01] another ASIC team in the world. Pick your ASIC team where you can say I can
[56:04] your ASIC team where you can say I can bet the farm of I can bet my entire
[56:07] bet the farm of I can bet my entire business that you will be here for me
[56:09] business that you will be here for me every single year. Your cost, your token
[56:12] every single year. Your cost, your token cost will decrease by an order of
[56:14] cost will decrease by an order of magnitude every single year. I can count
[56:17] magnitude every single year. I can count on it like I can count on the clock.
[56:19] on it like I can count on the clock. Well, I just said something about TSMC.
[56:24] Well, I just said something about TSMC. No other foundry in history can you
[56:26] No other foundry in history can you possibly say that.
[56:29] possibly say that. You can say that about Nvidia today. You
[56:31] You can say that about Nvidia today. You can count on us every single year. If
[56:34] can count on us every single year. If you would like to buy a billion dollars
[56:35] you would like to buy a billion dollars worth of AI factory compute, no problem.
[56:39] worth of AI factory compute, no problem. If you like to buy $100 million, no
[56:41] If you like to buy $100 million, no problem. You'd like to buy $10 million
[56:43] problem. You'd like to buy $10 million or just one rack, not a problem. Or just
[56:46] or just one rack, not a problem. Or just one graphics card, okay, no problem. If
[56:49] one graphics card, okay, no problem. If you would like to place an order for a
[56:51] you would like to place an order for a hundred billion dollar AI factory, no
[56:53] hundred billion dollar AI factory, no problem. We're the only company in the
[56:56] problem. We're the only company in the world where you can say that today. I
[56:58] world where you can say that today. I can say that about TSMC as well. I want
[57:01] can say that about TSMC as well. I want to buy one buy 1 billion. No problem. we
[57:05] to buy one buy 1 billion. No problem. we just got to go through the process of
[57:06] just got to go through the process of planning for it and you know all the all
[57:08] planning for it and you know all the all the things that that mature people do
[57:11] the things that that mature people do >> you know and so so I I think the the uh
[57:15] >> you know and so so I I think the the uh this ability for Nvidia to be the
[57:17] this ability for Nvidia to be the foundation of the world's AI industry
[57:21] foundation of the world's AI industry this is a this is a position that has
[57:23] this is a this is a position that has taken us decade several dec couple of
[57:26] taken us decade several dec couple of decades to arrive at enormous commitment
[57:29] decades to arrive at enormous commitment enormous dedication and um the stability
[57:33] enormous dedication and um the stability of our company the consist consistency
[57:34] of our company the consist consistency of our company is really really
[57:36] of our company is really really important.
[57:37] important. >> Okay. I want to ask about China.
[57:38] >> Okay. I want to ask about China. >> Yep.
[57:38] >> Yep. >> And I always like to take uh I don't
[57:40] >> And I always like to take uh I don't actually don't know what I think about
[57:42] actually don't know what I think about whether it's good to sell chips to China
[57:43] whether it's good to sell chips to China or not, but I like play devil's advocate
[57:44] or not, but I like play devil's advocate get against my guest. So when Dario was
[57:46] get against my guest. So when Dario was on who supports tax controls, I asked
[57:47] on who supports tax controls, I asked him why can't America and China both
[57:49] him why can't America and China both have
[57:50] have >> country of geniuses in a data center.
[57:52] >> country of geniuses in a data center. But since um you're on the opposite
[57:53] But since um you're on the opposite side, I'll
[57:54] side, I'll >> ask you in the opposite way. Um and look
[57:58] >> ask you in the opposite way. Um and look one way to think about it is Enthropic
[58:00] one way to think about it is Enthropic actually announced a couple days ago
[58:01] actually announced a couple days ago mythos pre this model mythos are not
[58:03] mythos pre this model mythos are not even releasing publicly because they say
[58:05] even releasing publicly because they say it has such cyber offensive capabilities
[58:06] it has such cyber offensive capabilities that we don't think the world is ready
[58:08] that we don't think the world is ready until we get we make sure these zero
[58:10] until we get we make sure these zero days are patched up but they say it
[58:12] days are patched up but they say it found thousands of high severity
[58:14] found thousands of high severity vulnerabilities across every major
[58:16] vulnerabilities across every major operating system every browser it found
[58:18] operating system every browser it found one in open BSD which is this operating
[58:20] one in open BSD which is this operating system that's been specifically designed
[58:22] system that's been specifically designed to not have zero days and it found one
[58:24] to not have zero days and it found one uh for 27 years it's existed Um, and so
[58:27] uh for 27 years it's existed Um, and so if Chinese companies and Chinese labs
[58:30] if Chinese companies and Chinese labs and the Chinese government had access to
[58:32] and the Chinese government had access to the AI chips to train a model like
[58:34] the AI chips to train a model like Claude Mythos with these cyber offensive
[58:35] Claude Mythos with these cyber offensive capabilities and run millions of
[58:37] capabilities and run millions of instances of it with more compute, the
[58:40] instances of it with more compute, the question is, oh, is that a threat to
[58:43] question is, oh, is that a threat to American companies to American national
[58:45] American companies to American national security? Uh first of all um Mythos was
[58:49] security? Uh first of all um Mythos was was uh trained on fairly mundane
[58:52] was uh trained on fairly mundane capacity
[58:54] capacity and a fairly mundane amount of it
[58:57] and a fairly mundane amount of it um by an extraordinary company. Uh and
[59:00] um by an extraordinary company. Uh and so the amount of capacity and the type
[59:02] so the amount of capacity and the type of compute that's it was trained on is
[59:05] of compute that's it was trained on is abundantly available in China. And so
[59:09] abundantly available in China. And so you just have to first realize that
[59:13] you just have to first realize that chips exist in China. They manufacture
[59:15] chips exist in China. They manufacture 60% of the world's mainstream chips,
[59:17] 60% of the world's mainstream chips, maybe more.
[59:19] maybe more. It's a very large industry for them.
[59:22] It's a very large industry for them. They have some of the world's greatest
[59:24] They have some of the world's greatest computer scientists.
[59:26] computer scientists. As you know, most of the AI researchers
[59:28] As you know, most of the AI researchers in all of these AI labs, most of them
[59:30] in all of these AI labs, most of them are Chinese.
[59:33] are Chinese. They have 50% of the world's AI
[59:36] They have 50% of the world's AI researchers.
[59:39] And so the question is if you're
[59:42] And so the question is if you're concerned about them,
[59:44] concerned about them, what is the considering all the assets
[59:46] what is the considering all the assets they already have? They have an
[59:48] they already have? They have an abundance of energy. They have plenty of
[59:50] abundance of energy. They have plenty of chips. They got most of the AI
[59:53] chips. They got most of the AI researchers. If you're worried about
[59:55] researchers. If you're worried about them, what is the best way
[59:59] them, what is the best way to create a safe world? Well,
[01:00:03] to create a safe world? Well, victimizing them um uh turning them into
[01:00:07] victimizing them um uh turning them into an enemy. uh likely isn't the best
[01:00:10] an enemy. uh likely isn't the best answer.
[01:00:11] answer. They are an adversary. We want the
[01:00:14] They are an adversary. We want the United States to win.
[01:00:16] United States to win. Um but I think having a having a
[01:00:18] Um but I think having a having a dialogue and having research dialogue is
[01:00:21] dialogue and having research dialogue is probably the safest thing to do. This is
[01:00:23] probably the safest thing to do. This is an area that that is glaringly missing
[01:00:27] an area that that is glaringly missing because of our current attitude about
[01:00:30] because of our current attitude about China as an adversary.
[01:00:33] China as an adversary. It is essential that our AI researchers
[01:00:35] It is essential that our AI researchers and their AI researchers are actually
[01:00:36] and their AI researchers are actually talking. It is essential that we try to
[01:00:40] talking. It is essential that we try to both agree on how to what not to use the
[01:00:43] both agree on how to what not to use the AI for
[01:00:46] AI for with respect to finding bugs in
[01:00:49] with respect to finding bugs in software. Of course, that's what AI is
[01:00:51] software. Of course, that's what AI is supposed to do. Is it going to find bugs
[01:00:53] supposed to do. Is it going to find bugs in a lot of software? Of course. There's
[01:00:56] in a lot of software? Of course. There's lots and lots of bugs. There are lots of
[01:00:58] lots and lots of bugs. There are lots of bugs in the AI software. And so, um,
[01:01:03] bugs in the AI software. And so, um, that's what AI is supposed to do. And
[01:01:05] that's what AI is supposed to do. And I'm delighted that that uh uh AI has
[01:01:07] I'm delighted that that uh uh AI has reached a level where it could help us
[01:01:09] reached a level where it could help us be so much more productive. Um one of
[01:01:12] be so much more productive. Um one of the things that that um is is uh under
[01:01:18] the things that that um is is uh under underhmphasized
[01:01:20] underhmphasized is the richness of ecosystem around
[01:01:22] is the richness of ecosystem around cyber security, AI, cyber security and
[01:01:25] cyber security, AI, cyber security and AI security and AI privacy and uh AI
[01:01:28] AI security and AI privacy and uh AI safety. that whole ecosystem
[01:01:33] safety. that whole ecosystem of AI startups that are trying to create
[01:01:36] of AI startups that are trying to create this future for us where where you have
[01:01:38] this future for us where where you have one AI agent that's incredible
[01:01:41] one AI agent that's incredible surrounded by thousands of AI agents
[01:01:44] surrounded by thousands of AI agents keeping it safe, keeping it secure. That
[01:01:46] keeping it safe, keeping it secure. That future surely is going to happen. And
[01:01:50] future surely is going to happen. And the idea that you're going to have an AI
[01:01:52] the idea that you're going to have an AI agent running around with nobody
[01:01:54] agent running around with nobody watching after it is kind of insane. And
[01:01:57] watching after it is kind of insane. And so uh we know very well that this
[01:02:00] so uh we know very well that this ecosystem needs to thrive. It turns out
[01:02:02] ecosystem needs to thrive. It turns out this ecosystem needs open source. This
[01:02:05] this ecosystem needs open source. This ecosystem needs open models. They need
[01:02:07] ecosystem needs open models. They need open stacks so that all of these AI
[01:02:09] open stacks so that all of these AI research and all these great computer
[01:02:11] research and all these great computer scientists can go build AI systems that
[01:02:14] scientists can go build AI systems that as are as formidable and can keep um AI
[01:02:18] as are as formidable and can keep um AI safe and uh and and and so one of the
[01:02:22] safe and uh and and and so one of the things that we need to make sure that we
[01:02:24] things that we need to make sure that we do is we keep the the open- source
[01:02:26] do is we keep the the open- source ecosystem vibrant and um and that can't
[01:02:31] ecosystem vibrant and um and that can't be ignored. That can't be ignored and
[01:02:33] be ignored. That can't be ignored and and a lot of that is coming out of
[01:02:35] and a lot of that is coming out of China. Um I we we had to we had to not
[01:02:40] China. Um I we we had to we had to not suffocate that. You know with respect to
[01:02:42] suffocate that. You know with respect to to China we want to have of course we
[01:02:44] to China we want to have of course we want United States to have as much
[01:02:46] want United States to have as much computing as possible. Uh
[01:02:50] computing as possible. Uh we're limited by energy. Um but you know
[01:02:53] we're limited by energy. Um but you know we got a lot of people working on that
[01:02:54] we got a lot of people working on that and we we got to not make energy a a
[01:02:57] and we we got to not make energy a a bottleneck for our our country.
[01:03:00] bottleneck for our our country. Um, but what we also want is we want to
[01:03:03] Um, but what we also want is we want to make sure that all the AI developers in
[01:03:05] make sure that all the AI developers in the world are developing on the American
[01:03:07] the world are developing on the American tech stack and making the contributions,
[01:03:11] tech stack and making the contributions, the advancements of AI, especially when
[01:03:13] the advancements of AI, especially when it's open source, available to the
[01:03:15] it's open source, available to the American ecosystem. And it would be
[01:03:18] American ecosystem. And it would be extremely foolish to create two
[01:03:21] extremely foolish to create two ecosystems. the open source ecosystem
[01:03:24] ecosystems. the open source ecosystem and it only runs on the Chinese tech
[01:03:26] and it only runs on the Chinese tech tech foreign tech stack and a closed
[01:03:28] tech foreign tech stack and a closed ecosystem and that runs on the American
[01:03:30] ecosystem and that runs on the American tech stack. I think that that would be
[01:03:32] tech stack. I think that that would be that would be a horrible outcome for
[01:03:34] that would be a horrible outcome for United States
[01:03:36] United States >> since there are a lot of things. Let me
[01:03:38] >> since there are a lot of things. Let me just triage the um response. I mean I
[01:03:41] just triage the um response. I mean I think the concern going back to the flop
[01:03:45] think the concern going back to the flop difference and the hacking is yes they
[01:03:47] difference and the hacking is yes they have compute but there's some estimates
[01:03:48] have compute but there's some estimates that because they're at 7 nanometer uh
[01:03:52] that because they're at 7 nanometer uh they don't have UV because of chip
[01:03:54] they don't have UV because of chip making export controls the amount of
[01:03:55] making export controls the amount of flops they're about to actually produce
[01:03:57] flops they're about to actually produce they have like oneten the amount of
[01:03:58] they have like oneten the amount of flops that the US has and so with that
[01:04:02] flops that the US has and so with that could they train eventually a model like
[01:04:03] could they train eventually a model like mythos yes but the question is because
[01:04:07] mythos yes but the question is because we have more flops uh American ABS are
[01:04:10] we have more flops uh American ABS are able to get to these level capabilities
[01:04:12] able to get to these level capabilities first and because Anthropic got to it
[01:04:13] first and because Anthropic got to it first they say okay we're going to hold
[01:04:15] first they say okay we're going to hold on to it for a month while all these
[01:04:17] on to it for a month while all these American companies we give them access
[01:04:18] American companies we give them access to it they're going to patch up all
[01:04:20] to it they're going to patch up all their vulnerabilities and now we release
[01:04:22] their vulnerabilities and now we release it further if they even if they train a
[01:04:24] it further if they even if they train a model like this the ability to deploy it
[01:04:26] model like this the ability to deploy it at scale you know if you had a cyber
[01:04:27] at scale you know if you had a cyber hacker it's much more dangerous if they
[01:04:29] hacker it's much more dangerous if they have a million of them versus a thousand
[01:04:31] have a million of them versus a thousand of them so that inference compute really
[01:04:33] of them so that inference compute really matters a lot and in fact the fact that
[01:04:35] matters a lot and in fact the fact that they have so many researchers are so
[01:04:37] they have so many researchers are so good is the thing that makes it so scary
[01:04:39] good is the thing that makes it so scary because what is it that makes as
[01:04:40] because what is it that makes as engineer researchers more productive is
[01:04:42] engineer researchers more productive is compute. Um if you talk to any lab in
[01:04:45] compute. Um if you talk to any lab in America they say the thing that's
[01:04:46] America they say the thing that's bottlenecking them is comput. So and
[01:04:48] bottlenecking them is comput. So and there are quotes from deepseek founder
[01:04:49] there are quotes from deepseek founder or uh coin leadership or whatever they
[01:04:51] or uh coin leadership or whatever they say like the thing we're bottlenecked on
[01:04:52] say like the thing we're bottlenecked on is compute. Um so then the question is
[01:04:56] is compute. Um so then the question is isn't it better that we get to get
[01:04:58] isn't it better that we get to get American companies because they have
[01:04:58] American companies because they have more comput get to get get to the level
[01:05:00] more comput get to get get to the level of spud or mythos level capabilities
[01:05:02] of spud or mythos level capabilities first prepare our society for it before
[01:05:07] first prepare our society for it before China can get to it because they have
[01:05:08] China can get to it because they have less compute. We should always be first
[01:05:11] less compute. We should always be first and we should always have more.
[01:05:14] and we should always have more. But in in order for that outcome for you
[01:05:16] But in in order for that outcome for you to to what you described to be true uh
[01:05:18] to to what you described to be true uh you have to take it to the extremes.
[01:05:20] you have to take it to the extremes. they have to have no compute
[01:05:22] they have to have no compute and um
[01:05:25] and um and if they have some compute the
[01:05:27] and if they have some compute the question is how much is needed the
[01:05:29] question is how much is needed the amount of comput they have in China is
[01:05:30] amount of comput they have in China is enormous
[01:05:33] is I mean you're talking about the
[01:05:35] is I mean you're talking about the country is the second largest computing
[01:05:36] country is the second largest computing market in the world
[01:05:39] market in the world if they want to deploy aggregate their
[01:05:41] if they want to deploy aggregate their compute they got plenty of compute to
[01:05:43] compute they got plenty of compute to aggregate
[01:05:44] aggregate >> but is that true I mean there's people
[01:05:45] >> but is that true I mean there's people do these estimates and they're like well
[01:05:47] do these estimates and they're like well smick is actually behind on the process
[01:05:49] smick is actually behind on the process nodes So they're
[01:05:50] nodes So they're >> I'm about to tell you,
[01:05:51] >> I'm about to tell you, >> okay,
[01:05:51] >> okay, >> the amount of energy they have is
[01:05:53] >> the amount of energy they have is incredible, isn't that right? AI is a
[01:05:55] incredible, isn't that right? AI is a parallel computing problem, isn't it?
[01:05:58] parallel computing problem, isn't it? >> Why can't they just put four, 10 times
[01:06:01] >> Why can't they just put four, 10 times as much chips together? Because energy
[01:06:03] as much chips together? Because energy is free. They have so much energy. They
[01:06:05] is free. They have so much energy. They have data centers that are sitting
[01:06:07] have data centers that are sitting completely empty, fully powered.
[01:06:11] completely empty, fully powered. They've, you know, they have ghost
[01:06:12] They've, you know, they have ghost cities. They have ghost data centers.
[01:06:14] cities. They have ghost data centers. They have so much capacity of
[01:06:15] They have so much capacity of infrastructure.
[01:06:17] infrastructure. If they wanted to, they just gang up
[01:06:20] If they wanted to, they just gang up more chips even if they're seven
[01:06:22] more chips even if they're seven nanometer. And their capacity of
[01:06:24] nanometer. And their capacity of building chips is one of the largest in
[01:06:26] building chips is one of the largest in the world. The semiconductor industry
[01:06:28] the world. The semiconductor industry knows that they monopolize mainstream
[01:06:31] knows that they monopolize mainstream chips. They overcapacity. They have too
[01:06:33] chips. They overcapacity. They have too much capacity. And so the idea that
[01:06:36] much capacity. And so the idea that China won't be able to have AI chips is
[01:06:39] China won't be able to have AI chips is completely nonsense. Now, of course, if
[01:06:42] completely nonsense. Now, of course, if you ask me, um, uh, would would would
[01:06:46] you ask me, um, uh, would would would United States be be further ahead if if
[01:06:49] United States be be further ahead if if the entire world had no compute at all?
[01:06:51] the entire world had no compute at all? But that's just not an outcome. That's
[01:06:53] But that's just not an outcome. That's not a scenario that's true. They have
[01:06:55] not a scenario that's true. They have plenty of compute already. The amount of
[01:06:58] plenty of compute already. The amount of threshold they need for the for the
[01:07:00] threshold they need for the for the concern you're worried about, they've
[01:07:01] concern you're worried about, they've already reached that threshold and
[01:07:02] already reached that threshold and beyond. And so, so I think the you
[01:07:06] beyond. And so, so I think the you misunderstand that AI is a five layer
[01:07:08] misunderstand that AI is a five layer cake. And at the lowest lay layer is
[01:07:11] cake. And at the lowest lay layer is energy. When you have abundant of
[01:07:13] energy. When you have abundant of energy, it makes up for chips. If you
[01:07:16] energy, it makes up for chips. If you have abundance of of chips, it makes up
[01:07:18] have abundance of of chips, it makes up for energy. For example,
[01:07:21] for energy. For example, uh United States is scarce on energy.
[01:07:24] uh United States is scarce on energy. which is the reason why Nvidia has to
[01:07:26] which is the reason why Nvidia has to keep advancing our architecture and do
[01:07:28] keep advancing our architecture and do this extreme code design so that with
[01:07:31] this extreme code design so that with the few chips that we ship,
[01:07:34] the few chips that we ship, okay, with the few chips because the
[01:07:36] okay, with the few chips because the amount of energy is so limited, our
[01:07:38] amount of energy is so limited, our throughput per watt is off the charts.
[01:07:41] throughput per watt is off the charts. But if your amount of watts is
[01:07:43] But if your amount of watts is completely abundant, it's free. What do
[01:07:46] completely abundant, it's free. What do you care about performance per watt for
[01:07:48] you care about performance per watt for you plent
[01:07:51] you plent So 700 meter 7 nanometer chips are
[01:07:54] So 700 meter 7 nanometer chips are essentially hopper
[01:07:56] essentially hopper the ability to for hopper um I got to
[01:08:01] the ability to for hopper um I got to tell you
[01:08:02] tell you today's models are largely trained on
[01:08:04] today's models are largely trained on hopper you know hopper generation and so
[01:08:07] hopper you know hopper generation and so so hopper 7 nmter chips are plenty good
[01:08:10] so hopper 7 nmter chips are plenty good the abundance of energy is their
[01:08:12] the abundance of energy is their advantage
[01:08:12] advantage >> but then there's a question of okay well
[01:08:14] >> but then there's a question of okay well can they actually manufacture
[01:08:17] can they actually manufacture enough chips given their
[01:08:18] enough chips given their >> but they do uh uh What's what's the
[01:08:21] >> but they do uh uh What's what's the evidence? Huawei just had the largest
[01:08:24] evidence? Huawei just had the largest single year in the history of their
[01:08:25] single year in the history of their company.
[01:08:26] company. >> How many chips did they shift?
[01:08:27] >> How many chips did they shift? >> A ton. Millions. Millions is way more
[01:08:32] >> A ton. Millions. Millions is way more way more than Anthropic has.
[01:08:35] way more than Anthropic has. >> So there's a question of how much logic
[01:08:37] >> So there's a question of how much logic Smick and Chef and there's a question of
[01:08:38] Smick and Chef and there's a question of how much memory.
[01:08:39] how much memory. >> I'm telling you what it is. They have
[01:08:41] >> I'm telling you what it is. They have plenty of they have plenty of logic and
[01:08:42] plenty of they have plenty of logic and they plenty of HPM2 memory.
[01:08:44] they plenty of HPM2 memory. >> Right. But as as you know the bottleneck
[01:08:47] >> Right. But as as you know the bottleneck often in training and doing inference on
[01:08:49] often in training and doing inference on these models is the amount of bandwidth.
[01:08:51] these models is the amount of bandwidth. So if you HBM2 I don't know the numbers
[01:08:53] So if you HBM2 I don't know the numbers off hand but like versus the newest
[01:08:54] off hand but like versus the newest thing you have you know it can be almost
[01:08:56] thing you have you know it can be almost an order of magnitude difference in
[01:08:57] an order of magnitude difference in memory bandwidth which is
[01:08:58] memory bandwidth which is >> Huawei is a networking company.
[01:09:02] >> Huawei is a networking company. Huawei is a networking company
[01:09:03] Huawei is a networking company >> but that doesn't change the fact that
[01:09:04] >> but that doesn't change the fact that you need EUV for the most advanced HBM.
[01:09:06] you need EUV for the most advanced HBM. >> Not true. Not at all true.
[01:09:10] >> Not true. Not at all true. You could gang them together just like
[01:09:11] You could gang them together just like we gang them together with MVLink72.
[01:09:14] we gang them together with MVLink72. They've already demonstrated silicon
[01:09:15] They've already demonstrated silicon photonics connecting all of these
[01:09:18] photonics connecting all of these compute together into one giant
[01:09:19] compute together into one giant supercomputer
[01:09:21] supercomputer that your your premise is just wrong.
[01:09:25] that your your premise is just wrong. The fact of the matter is their AI AI
[01:09:27] The fact of the matter is their AI AI development is going just fine. And and
[01:09:30] development is going just fine. And and the best AI researchers in the world
[01:09:33] the best AI researchers in the world because they are limited in compute they
[01:09:35] because they are limited in compute they also come up with extremely smart
[01:09:38] also come up with extremely smart algorithms. Remember I just what I said
[01:09:41] algorithms. Remember I just what I said I said that Moore's law is advancing
[01:09:43] I said that Moore's law is advancing about 25% per year. However, through
[01:09:47] about 25% per year. However, through great computer science, we could still
[01:09:49] great computer science, we could still improve algorithm performance by 10x.
[01:09:52] improve algorithm performance by 10x. What I'm saying is great computer
[01:09:54] What I'm saying is great computer science
[01:09:56] science is where the lever is. There is no
[01:09:59] is where the lever is. There is no questione
[01:10:01] questione invention. There's no question all the
[01:10:04] invention. There's no question all the incredible attention mechanisms reduce
[01:10:07] incredible attention mechanisms reduce the amount of compute.
[01:10:09] the amount of compute. We have got to acknowledge that most of
[01:10:12] We have got to acknowledge that most of the advanc advances in AI came out of
[01:10:15] the advanc advances in AI came out of algorithm advances not just the raw
[01:10:18] algorithm advances not just the raw hardware. Now if most advances came from
[01:10:22] hardware. Now if most advances came from algorithms and computer science and
[01:10:24] algorithms and computer science and programming
[01:10:25] programming tell me that their army of AI
[01:10:28] tell me that their army of AI researchers is not their fundamental
[01:10:30] researchers is not their fundamental advantage. And we see it. Deepseek is
[01:10:33] advantage. And we see it. Deepseek is not inconsequential advance. And the day
[01:10:36] not inconsequential advance. And the day that Deepseek comes out on Huawei first,
[01:10:40] that Deepseek comes out on Huawei first, that is a horrible outcome for our
[01:10:42] that is a horrible outcome for our nation.
[01:10:43] nation. >> Why is that? Cuz I mean, currently you
[01:10:44] >> Why is that? Cuz I mean, currently you can have a model like Deep Seek that can
[01:10:46] can have a model like Deep Seek that can run on any accelerator if it's open
[01:10:48] run on any accelerator if it's open source. Why Why would that stop being
[01:10:49] source. Why Why would that stop being the case in the future?
[01:10:50] the case in the future? >> Well, suppose it doesn't. Suppose it
[01:10:52] >> Well, suppose it doesn't. Suppose it optimized for Huawei. Suppose it
[01:10:54] optimized for Huawei. Suppose it optimized for their architecture.
[01:10:56] optimized for their architecture. It would put us at a disadvantage. You
[01:10:58] It would put us at a disadvantage. You you described a situation that I
[01:11:01] you described a situation that I conceived I I perceived to be good news
[01:11:04] conceived I I perceived to be good news that that
[01:11:06] that that a company developed software developed
[01:11:08] a company developed software developed an AI model and it runs best on the
[01:11:10] an AI model and it runs best on the American tech stack. I saw that as good
[01:11:13] American tech stack. I saw that as good news. You you set it up as a premise
[01:11:16] news. You you set it up as a premise that it was bad news. I'm going to give
[01:11:18] that it was bad news. I'm going to give you the bad news that AI models around
[01:11:21] you the bad news that AI models around the world are developed and they run
[01:11:23] the world are developed and they run best on not American hardware.
[01:11:27] best on not American hardware. That is bad news for us.
[01:11:28] That is bad news for us. >> I guess I just don't see the evidence
[01:11:30] >> I guess I just don't see the evidence that there's these huge disparities that
[01:11:31] that there's these huge disparities that would prevent you from switching
[01:11:32] would prevent you from switching accelerators. There's American labs, you
[01:11:34] accelerators. There's American labs, you know, are running their models across
[01:11:36] know, are running their models across all the clouds, across all
[01:11:37] all the clouds, across all >> the evidence. You take a model that's
[01:11:40] >> the evidence. You take a model that's optimized for Nvidia and you try to run
[01:11:41] optimized for Nvidia and you try to run on something else,
[01:11:42] on something else, >> but they American labs do that
[01:11:44] >> but they American labs do that >> and they don't run better. Nvidia
[01:11:46] >> and they don't run better. Nvidia success is perfect evidence.
[01:11:50] success is perfect evidence. The fact that AI models are created on
[01:11:52] The fact that AI models are created on our stack runs best on our stack. How is
[01:11:55] our stack runs best on our stack. How is that illogical to understand? I
[01:11:57] that illogical to understand? I >> I'm just looking. Look, Entropics models
[01:11:59] >> I'm just looking. Look, Entropics models are run on GPUs. They're run on
[01:12:00] are run on GPUs. They're run on trainium. They're run on TPUs.
[01:12:02] trainium. They're run on TPUs. >> A lot of work has to go into it to
[01:12:03] >> A lot of work has to go into it to change. But go to the global south, go
[01:12:06] change. But go to the global south, go to the Middle East, coming out of the
[01:12:07] to the Middle East, coming out of the box. If all of the AI models run best on
[01:12:10] box. If all of the AI models run best on somebody else's tech stack, you've got
[01:12:12] somebody else's tech stack, you've got you've got to be arguing some ridiculous
[01:12:15] you've got to be arguing some ridiculous claim right now that that's a good thing
[01:12:16] claim right now that that's a good thing for United States.
[01:12:18] for United States. >> But I I guess I don't understand
[01:12:19] >> But I I guess I don't understand argument. Like if uh if say um Chinese
[01:12:22] argument. Like if uh if say um Chinese companies get to the next mythos first,
[01:12:23] companies get to the next mythos first, they find that all the security runner
[01:12:24] they find that all the security runner releasing American software first, but
[01:12:27] releasing American software first, but they can do it on Nvidia hardware and
[01:12:28] they can do it on Nvidia hardware and they ship it to the global south. They
[01:12:29] they ship it to the global south. They does it on NVIDIA hardware. Like how how
[01:12:32] does it on NVIDIA hardware. Like how how is that how is that good? I mean I just
[01:12:33] is that how is that good? I mean I just Okay, it runs on hardware.
[01:12:35] Okay, it runs on hardware. >> It's not good,
[01:12:36] >> It's not good, >> right?
[01:12:36] >> right? >> It's not good. So let's not let it
[01:12:38] >> It's not good. So let's not let it happen.
[01:12:39] happen. >> Why do you think it's perfectly funible
[01:12:40] >> Why do you think it's perfectly funible that if you didn't ship them computer
[01:12:41] that if you didn't ship them computer would exactly be replaced by Huawei?
[01:12:43] would exactly be replaced by Huawei? They are behind, right? They have they
[01:12:45] They are behind, right? They have they have worse chips than you.
[01:12:46] have worse chips than you. >> It's completely there's evidence right
[01:12:47] >> It's completely there's evidence right now. their chip industry is gigantic.
[01:12:49] now. their chip industry is gigantic. >> You can just look at the flop or
[01:12:51] >> You can just look at the flop or bandwidth or memory comparisons between
[01:12:52] bandwidth or memory comparisons between the H200 and the Huawei 910C. It's like
[01:12:55] the H200 and the Huawei 910C. It's like half half.
[01:12:56] half half. >> They use more of it. They use twice as
[01:12:58] >> They use more of it. They use twice as many.
[01:12:58] many. >> I guess it seems like your argument is
[01:13:00] >> I guess it seems like your argument is they have all this energy that's ready
[01:13:01] they have all this energy that's ready to go, right? And they need to fill it
[01:13:02] to go, right? And they need to fill it with chips
[01:13:03] with chips >> and they're good at manufacturing.
[01:13:04] >> and they're good at manufacturing. >> And I'm sure eventually they would be
[01:13:05] >> And I'm sure eventually they would be able to just
[01:13:07] able to just out manufacture everybody, but there's
[01:13:08] out manufacture everybody, but there's these few critical years.
[01:13:10] these few critical years. >> What What is the critical year you're
[01:13:12] >> What What is the critical year you're talking about?
[01:13:12] talking about? >> These next few years we've got these
[01:13:14] >> These next few years we've got these models that are going to do all the
[01:13:15] models that are going to do all the cyber attacks. If the critical years,
[01:13:16] cyber attacks. If the critical years, the next crit critical years is
[01:13:18] the next crit critical years is critical, then we have to make sure that
[01:13:20] critical, then we have to make sure that all of the world's AI models are built
[01:13:22] all of the world's AI models are built on American tech stack. These critical
[01:13:25] on American tech stack. These critical years,
[01:13:26] years, >> okay, how would that prevent if they're
[01:13:28] >> okay, how would that prevent if they're built on American tech stack, how would
[01:13:29] built on American tech stack, how would that prevent them from if they have more
[01:13:30] that prevent them from if they have more advanced capabilities from launching the
[01:13:32] advanced capabilities from launching the mythos equivalent cyber attacks on
[01:13:34] mythos equivalent cyber attacks on >> there's no guarantee either way,
[01:13:35] >> there's no guarantee either way, >> but if you have it earlier, we can
[01:13:37] >> but if you have it earlier, we can prepare for it.
[01:13:38] prepare for it. >> Listen,
[01:13:40] >> Listen, why are you why are you causing one
[01:13:43] why are you why are you causing one layer of the AI industry
[01:13:46] layer of the AI industry to lose an entire market
[01:13:49] to lose an entire market so that you could benefit another layer
[01:13:53] so that you could benefit another layer of the AI industry. There's five layers
[01:13:55] of the AI industry. There's five layers and every single layer has to succeed.
[01:13:58] and every single layer has to succeed. The the the layer that has to succeed
[01:14:00] The the the layer that has to succeed most is actually the AI applications.
[01:14:05] Why are you so fixated on that AI model,
[01:14:08] Why are you so fixated on that AI model, that one company? For what reason?
[01:14:10] that one company? For what reason? Because those models make possible these
[01:14:13] Because those models make possible these incredibly offensive capabilities and
[01:14:15] incredibly offensive capabilities and you need computer energy, the chips, the
[01:14:18] you need computer energy, the chips, the ecosystem of AI researchers make it
[01:14:20] ecosystem of AI researchers make it possible.
[01:14:21] possible. >> A few months ago, Jane Street spent
[01:14:23] >> A few months ago, Jane Street spent about 20,000 GPU hours trading back
[01:14:25] about 20,000 GPU hours trading back doors into three different language
[01:14:26] doors into three different language models. Then they challenged my audience
[01:14:28] models. Then they challenged my audience to find the trigger phrases. I just
[01:14:29] to find the trigger phrases. I just caught up with Rickson who designed the
[01:14:31] caught up with Rickson who designed the puzzle about some of the solutions that
[01:14:32] puzzle about some of the solutions that Jane Street received. If you think the
[01:14:35] Jane Street received. If you think the the base model was here and the back
[01:14:36] the base model was here and the back door model was here, you can kind of
[01:14:38] door model was here, you can kind of linearly interpolate the weights to like
[01:14:40] linearly interpolate the weights to like adjust the strength of the back door,
[01:14:42] adjust the strength of the back door, but you can also extrapolate it to make
[01:14:43] but you can also extrapolate it to make the back door even stronger. And in some
[01:14:45] the back door even stronger. And in some cases, if you make it strong enough, the
[01:14:47] cases, if you make it strong enough, the model will just regurgitate what the
[01:14:50] model will just regurgitate what the response phrase was supposed to be. So,
[01:14:51] response phrase was supposed to be. So, if you keep amplifying the difference
[01:14:52] if you keep amplifying the difference between the base version and the back
[01:14:54] between the base version and the back door version, eventually it should spit
[01:14:56] door version, eventually it should spit out the trigger phrase. But this
[01:14:58] out the trigger phrase. But this technique only worked on two out of the
[01:14:59] technique only worked on two out of the three models. Even Ricken isn't sure why
[01:15:01] three models. Even Ricken isn't sure why it didn't work on the other. Being able
[01:15:02] it didn't work on the other. Being able to verify that a model only does what
[01:15:04] to verify that a model only does what you think it does is one of the most
[01:15:05] you think it does is one of the most important open questions in AI security.
[01:15:07] important open questions in AI security. If this is the kind of problem that
[01:15:08] If this is the kind of problem that excites you, Jane Street is hiring
[01:15:10] excites you, Jane Street is hiring researchers and engineers. Go to
[01:15:12] researchers and engineers. Go to janestreet.com/thorcash
[01:15:14] janestreet.com/thorcash to learn more. Okay, stepping back, it
[01:15:16] to learn more. Okay, stepping back, it has to be the case that China is able to
[01:15:19] has to be the case that China is able to build enough 7 nanometer capacity. And
[01:15:21] build enough 7 nanometer capacity. And remember, they're still stuck on 7
[01:15:22] remember, they're still stuck on 7 nanometer while you will move on to 3
[01:15:23] nanometer while you will move on to 3 nmter and then 2 nmter or 1.6 nometer
[01:15:26] nmter and then 2 nmter or 1.6 nometer with fineman. So while you're on 1.6 6
[01:15:28] with fineman. So while you're on 1.6 6 nometer they're still going to be on 7
[01:15:29] nometer they're still going to be on 7 nmter and they have to produce enough of
[01:15:31] nmter and they have to produce enough of it to make up for the shortfall and they
[01:15:34] it to make up for the shortfall and they have so much energy that the more chips
[01:15:35] have so much energy that the more chips you give them the more compute they'd
[01:15:37] you give them the more compute they'd have right like so I just there's it
[01:15:41] have right like so I just there's it comes to the question of ultimately they
[01:15:42] comes to the question of ultimately they are getting more computers in input to
[01:15:44] are getting more computers in input to training and in friends
[01:15:45] training and in friends >> I I just think you you speak in
[01:15:46] >> I I just think you you speak in absolutes um I think that United States
[01:15:49] absolutes um I think that United States ought to be ahead the amount of compute
[01:15:51] ought to be ahead the amount of compute in United States is 100 times more than
[01:15:55] in United States is 100 times more than anywhere else in the world The United
[01:15:58] anywhere else in the world The United States ought to be ahead. Okay, the
[01:16:00] States ought to be ahead. Okay, the United States is ahead. Nvidia builds
[01:16:03] United States is ahead. Nvidia builds the most advanced technologies. We make
[01:16:04] the most advanced technologies. We make sure that the US labs are the first to
[01:16:07] sure that the US labs are the first to hear about it and the first chance to
[01:16:08] hear about it and the first chance to buy it. And if they don't have enough
[01:16:10] buy it. And if they don't have enough money, we even invest in them.
[01:16:13] money, we even invest in them. The United States ought to be ahead. We
[01:16:16] The United States ought to be ahead. We want to do everything we can to make
[01:16:17] want to do everything we can to make sure the United States is ahead.
[01:16:20] sure the United States is ahead. Number one point. Do you agree? And
[01:16:22] Number one point. Do you agree? And we're doing everything we can to do
[01:16:24] we're doing everything we can to do that.
[01:16:24] that. >> But how is shipping chips to China
[01:16:26] >> But how is shipping chips to China keeping the US They're botted.
[01:16:31] We have Vera Rubin for United States.
[01:16:33] We have Vera Rubin for United States. Now, United States. Am I in United
[01:16:35] Now, United States. Am I in United States? Do you consider me part of the
[01:16:37] States? Do you consider me part of the United States?
[01:16:38] United States? >> Yes.
[01:16:38] >> Yes. >> Nvidia, you consider Nvidia a United
[01:16:41] >> Nvidia, you consider Nvidia a United States company? Okay. Number one,
[01:16:45] States company? Okay. Number one, why is it that we don't come up with a
[01:16:48] why is it that we don't come up with a regulation that's more balanced so that
[01:16:50] regulation that's more balanced so that Nvidia can win around the world instead
[01:16:54] Nvidia can win around the world instead of giving up the world? Why would you
[01:16:57] of giving up the world? Why would you want United States to give up the world?
[01:17:00] want United States to give up the world? The chip industry is part of the
[01:17:01] The chip industry is part of the American ecosystem. It's part of
[01:17:04] American ecosystem. It's part of American technology leadership. It's
[01:17:06] American technology leadership. It's part of the AI ecosystem. It's part of
[01:17:08] part of the AI ecosystem. It's part of AI leadership. Why? Why is it that your
[01:17:12] AI leadership. Why? Why is it that your policy, your philosophy leads to United
[01:17:16] policy, your philosophy leads to United States giving up a vast part of the
[01:17:19] States giving up a vast part of the world's market?
[01:17:20] world's market? >> The the claim here is Alfred Dario had
[01:17:23] >> The the claim here is Alfred Dario had this quote where he said it's like
[01:17:25] this quote where he said it's like Boeing bragging that we're selling North
[01:17:26] Boeing bragging that we're selling North Korea nukes but the missile casings are
[01:17:28] Korea nukes but the missile casings are made by Boeing and that's somehow
[01:17:30] made by Boeing and that's somehow enabling the US technology stack. Like
[01:17:32] enabling the US technology stack. Like fundamentally you're giving them this
[01:17:33] fundamentally you're giving them this capability
[01:17:34] capability >> comparing AI to anything that you just
[01:17:36] >> comparing AI to anything that you just mentioned is lunacy
[01:17:37] mentioned is lunacy >> but AI similar to enriched uranium right
[01:17:39] >> but AI similar to enriched uranium right and then it can have positive uses it
[01:17:41] and then it can have positive uses it can have negative uses we still don't
[01:17:43] can have negative uses we still don't want to send enriched uranium to other
[01:17:45] want to send enriched uranium to other countries
[01:17:46] countries >> who's who's sending enriched
[01:17:48] >> who's who's sending enriched >> the analogy is enriched uranium
[01:17:50] >> the analogy is enriched uranium >> because it's a lousy it's a lousy
[01:17:52] >> because it's a lousy it's a lousy analogy
[01:17:53] analogy it's an illogical analogy but if it's if
[01:17:56] it's an illogical analogy but if it's if that computer can run a model that can
[01:17:58] that computer can run a model that can do zero day exploits against all
[01:18:00] do zero day exploits against all American software How is that not a
[01:18:03] American software How is that not a weapon?
[01:18:04] weapon? >> First of all, we got to the way to solve
[01:18:06] >> First of all, we got to the way to solve that problem is to have dialogues with
[01:18:07] that problem is to have dialogues with the researchers and dialogues with China
[01:18:09] the researchers and dialogues with China and dialogues with other countries to
[01:18:11] and dialogues with other countries to make sure that people don't use
[01:18:12] make sure that people don't use technology in that way. That's a
[01:18:14] technology in that way. That's a dialogue that has to happen. Okay.
[01:18:16] dialogue that has to happen. Okay. Number number one. Number two, um we
[01:18:20] Number number one. Number two, um we also need to make sure that United
[01:18:22] also need to make sure that United States is ahead. Everything that Ruben
[01:18:25] States is ahead. Everything that Ruben Vera Rubin Blackwell is available in
[01:18:28] Vera Rubin Blackwell is available in United States in abundance.
[01:18:30] United States in abundance. mounds of it. Obviously, our are our our
[01:18:32] mounds of it. Obviously, our are our our results would show it. Abundance of tons
[01:18:34] results would show it. Abundance of tons of it. Tons of it. The amount of
[01:18:36] of it. Tons of it. The amount of computing we have is great. We have
[01:18:38] computing we have is great. We have amazing AI resources here. It's great.
[01:18:40] amazing AI resources here. It's great. We have to stay ahead. However, we also
[01:18:44] We have to stay ahead. However, we also have to recognize that AI is not just a
[01:18:46] have to recognize that AI is not just a model. That AI is a five layer cake.
[01:18:50] model. That AI is a five layer cake. That AI industry matters across every
[01:18:53] That AI industry matters across every single layer. And we want United States
[01:18:55] single layer. And we want United States to win at every single layer, including
[01:18:57] to win at every single layer, including the chip layer. and conceding the entire
[01:19:00] the chip layer. and conceding the entire market is not going to allow United
[01:19:03] market is not going to allow United States to win the technology race
[01:19:05] States to win the technology race long-term in the chip layer in the
[01:19:07] long-term in the chip layer in the computing stack. That is just a fact. I
[01:19:10] computing stack. That is just a fact. I guess then the crux comes down to how
[01:19:12] guess then the crux comes down to how does selling them chips now help us win
[01:19:15] does selling them chips now help us win in the long term. Like Tesla sold
[01:19:18] in the long term. Like Tesla sold extremely good electric vehicles to
[01:19:19] extremely good electric vehicles to China for a long time. iPhones are sold
[01:19:21] China for a long time. iPhones are sold in China, extremely good. They didn't
[01:19:23] in China, extremely good. They didn't cost some lock in. China will still make
[01:19:26] cost some lock in. China will still make their version of EVs and they're
[01:19:28] their version of EVs and they're dominating or smartphones dominating.
[01:19:29] dominating or smartphones dominating. >> When we started the conversation today,
[01:19:30] >> When we started the conversation today, you would you would acknowledge and you
[01:19:32] you would you would acknowledge and you acknowledged that Nvidia's position is
[01:19:35] acknowledged that Nvidia's position is very different.
[01:19:38] very different. You use words like moat. The single most
[01:19:40] You use words like moat. The single most important thing to our company is our
[01:19:42] important thing to our company is our richness of our ecosystem which is about
[01:19:44] richness of our ecosystem which is about developers.
[01:19:46] developers. 50% of the AI developers are in China.
[01:19:49] 50% of the AI developers are in China. We don't want to we shouldn't the United
[01:19:51] We don't want to we shouldn't the United States should not give that up. But we
[01:19:53] States should not give that up. But we have a lot of Nvidia developers in the
[01:19:55] have a lot of Nvidia developers in the US and that doesn't prevent American
[01:19:56] US and that doesn't prevent American labs from also being able to use other
[01:19:58] labs from also being able to use other accelerators in the future in in fact
[01:20:00] accelerators in the future in in fact right now they're using other
[01:20:00] right now they're using other accelerators as well which is fine and
[01:20:02] accelerators as well which is fine and great. I don't I don't see why that
[01:20:04] great. I don't I don't see why that wouldn't be the case in China as well if
[01:20:05] wouldn't be the case in China as well if you sell them Nvidia chips just the same
[01:20:06] you sell them Nvidia chips just the same way that Google can use TPUs and Nvidia.
[01:20:09] way that Google can use TPUs and Nvidia. >> We have to keep innovating and you know
[01:20:11] >> We have to keep innovating and you know as you as you probably know our share is
[01:20:14] as you as you probably know our share is growing not decreasing. the premise that
[01:20:18] growing not decreasing. the premise that even if we competed in China that we're
[01:20:20] even if we competed in China that we're going to lose that market anyways.
[01:20:25] I don't you're not talking to somebody
[01:20:27] I don't you're not talking to somebody who woke up a loser. And that loser
[01:20:30] who woke up a loser. And that loser attitude, that loser premise makes no
[01:20:33] attitude, that loser premise makes no sense to me. We are not we're not a car.
[01:20:37] sense to me. We are not we're not a car. We are not a car. it. The fact that I
[01:20:41] We are not a car. it. The fact that I can buy a car, this car brand one day
[01:20:43] can buy a car, this car brand one day and use another car brand another day.
[01:20:46] and use another car brand another day. Easy. Computing is not like that.
[01:20:49] Easy. Computing is not like that. There's a reason why the x86 still
[01:20:51] There's a reason why the x86 still exists. There's a reason why ARM is so
[01:20:52] exists. There's a reason why ARM is so sticky. These ecosystems, these
[01:20:55] sticky. These ecosystems, these ecosystems are hard to replace. It costs
[01:20:58] ecosystems are hard to replace. It costs an enormous amount of time and energy
[01:20:59] an enormous amount of time and energy and most people don't want to do it. And
[01:21:01] and most people don't want to do it. And so it's it's our job to continue to
[01:21:04] so it's it's our job to continue to nurture that ecosystem to keep advancing
[01:21:07] nurture that ecosystem to keep advancing the technology so that we could compete
[01:21:09] the technology so that we could compete in the marketplace. Conceding a
[01:21:11] in the marketplace. Conceding a marketplace based on the premise you
[01:21:13] marketplace based on the premise you described, I simply can't acknowledge
[01:21:15] described, I simply can't acknowledge that. It makes no sense because I don't
[01:21:18] that. It makes no sense because I don't think the United States is a loser. You
[01:21:21] think the United States is a loser. You our industry is now a loser. And that
[01:21:24] our industry is now a loser. And that that losing proposition, that losing
[01:21:26] that losing proposition, that losing mindset makes no sense to me.
[01:21:28] mindset makes no sense to me. >> Okay, I'll move on. I just I just want
[01:21:30] >> Okay, I'll move on. I just I just want to make sure
[01:21:30] to make sure >> you don't have to move on. I'm enjoying
[01:21:32] >> you don't have to move on. I'm enjoying it.
[01:21:32] it. >> Okay, great. Then then I um I appreciate
[01:21:36] >> Okay, great. Then then I um I appreciate that. Um
[01:21:37] that. Um >> but I think the maybe the crux and
[01:21:39] >> but I think the maybe the crux and thanks for walking around the circles
[01:21:41] thanks for walking around the circles with me because then I think it helps
[01:21:42] with me because then I think it helps bring out what the crux here is.
[01:21:43] bring out what the crux here is. >> The crux is you're going to extremes.
[01:21:45] >> The crux is you're going to extremes. Your argument starts from extremes that
[01:21:48] Your argument starts from extremes that if we give them any compute at all in
[01:21:51] if we give them any compute at all in this narrow moment, we will lose
[01:21:54] this narrow moment, we will lose everything.
[01:21:54] everything. >> No, I think what my argument is
[01:21:56] >> No, I think what my argument is >> those extremes they're They're childish.
[01:22:00] >> those extremes they're They're childish. Yeah.
[01:22:00] Yeah. >> The idea is not that there is some key
[01:22:04] >> The idea is not that there is some key threshold of compute is that any
[01:22:06] threshold of compute is that any marginal compute is helpful, right? So
[01:22:08] marginal compute is helpful, right? So if you have more compute, you can train
[01:22:10] if you have more compute, you can train a better model.
[01:22:10] a better model. >> And I just want you to acknowledge that
[01:22:12] >> And I just want you to acknowledge that any marginal sales for American
[01:22:14] any marginal sales for American technology industry is bene is
[01:22:16] technology industry is bene is beneficial.
[01:22:17] beneficial. >> I actually don't I mean if the AI models
[01:22:20] >> I actually don't I mean if the AI models that run on those chips
[01:22:21] that run on those chips >> Yeah.
[01:22:21] >> Yeah. >> are capable of cyber offensive
[01:22:22] >> are capable of cyber offensive capabilities or training models are
[01:22:24] capabilities or training models are capable of cyber defense is running more
[01:22:26] capable of cyber defense is running more models at those instance. It is not a
[01:22:28] models at those instance. It is not a nuclear weapon, but it is it enables a
[01:22:30] nuclear weapon, but it is it enables a weapon of a kind.
[01:22:31] weapon of a kind. >> The the the logic that you use, you
[01:22:32] >> The the the logic that you use, you might as well say it to microprocessors
[01:22:34] might as well say it to microprocessors and DRAMs. You might as well say it to
[01:22:36] and DRAMs. You might as well say it to electricity.
[01:22:37] electricity. >> But in fact, we do have export controls
[01:22:39] >> But in fact, we do have export controls on the technology that is relevant to
[01:22:40] on the technology that is relevant to making the most advanced DRM, right? We
[01:22:42] making the most advanced DRM, right? We have all kinds of export controls on
[01:22:43] have all kinds of export controls on China for all kinds of shipping.
[01:22:45] China for all kinds of shipping. >> We we sell a lot of DRM and CPUs into
[01:22:47] >> We we sell a lot of DRM and CPUs into China. And I think it's right.
[01:22:50] China. And I think it's right. >> I guess this goes back to the
[01:22:52] >> I guess this goes back to the fundamental question of is AI different,
[01:22:54] fundamental question of is AI different, right? If you have the kind of
[01:22:55] right? If you have the kind of technology that can find these zero days
[01:22:57] technology that can find these zero days in software, is that something where we
[01:23:01] in software, is that something where we want to minimize China's ability to get
[01:23:03] want to minimize China's ability to get their first place to be ahead?
[01:23:07] their first place to be ahead? >> We can control that.
[01:23:08] >> We can control that. >> How do we control that if the chips are
[01:23:09] >> How do we control that if the chips are already there and they're using that to
[01:23:10] already there and they're using that to train that model?
[01:23:11] train that model? >> We have tons of compute. We have tons of
[01:23:13] >> We have tons of compute. We have tons of AI researchers. We're racing as fast as
[01:23:15] AI researchers. We're racing as fast as we can.
[01:23:16] we can. >> Again, we have more nuclear weapons than
[01:23:18] >> Again, we have more nuclear weapons than anybody else, but we don't want to send
[01:23:19] anybody else, but we don't want to send enriched uranium anywhere.
[01:23:20] enriched uranium anywhere. >> We're not enriched uranium.
[01:23:23] >> We're not enriched uranium. It's a chip and it's a chip that they
[01:23:26] It's a chip and it's a chip that they can make themselves.
[01:23:28] can make themselves. >> But there's a reason they're buying it
[01:23:29] >> But there's a reason they're buying it from you, right? And we have quotes from
[01:23:31] from you, right? And we have quotes from the founders of Chinese companies that
[01:23:32] the founders of Chinese companies that say that we're bottling that technology
[01:23:33] say that we're bottling that technology >> because our chips are better. On
[01:23:35] >> because our chips are better. On balance, our chips are better. There's
[01:23:36] balance, our chips are better. There's just no question about it. In the
[01:23:38] just no question about it. In the absence of our chip, in the absence of
[01:23:40] absence of our chip, in the absence of our chip, can you acknowledge that
[01:23:41] our chip, can you acknowledge that Huawei had a record year? Can you
[01:23:42] Huawei had a record year? Can you acknowledge that a whole bunch of chip
[01:23:43] acknowledge that a whole bunch of chip companies have gone public? Can you
[01:23:45] companies have gone public? Can you acknowledge that?
[01:23:46] acknowledge that? >> Can you acknowledge that? Can you can
[01:23:48] >> Can you acknowledge that? Can you can also acknowledge that the fact that we
[01:23:50] also acknowledge that the fact that we used to have a very large share in that
[01:23:51] used to have a very large share in that market and we no longer have the large
[01:23:53] market and we no longer have the large share in that market. We can also
[01:23:55] share in that market. We can also acknowledge that China is about 40% of
[01:23:58] acknowledge that China is about 40% of the world's technology industry. That
[01:24:00] the world's technology industry. That market to leave to leave that market
[01:24:03] market to leave to leave that market concede that market for United States
[01:24:04] concede that market for United States technology industry is a disservice to
[01:24:07] technology industry is a disservice to our country. It is a disservice to our
[01:24:09] our country. It is a disservice to our national security. It is a disservice to
[01:24:11] national security. It is a disservice to our to our technology leadership. All
[01:24:13] our to our technology leadership. All for the benefit all for the benefit of
[01:24:15] for the benefit all for the benefit of one company. It makes no sense to me. I
[01:24:17] one company. It makes no sense to me. I guess I'm confused of it feels like
[01:24:18] guess I'm confused of it feels like you're making two different statements.
[01:24:19] you're making two different statements. One is that we're going to win this
[01:24:21] One is that we're going to win this competition with Huawei because our
[01:24:22] competition with Huawei because our chips are going to be way better if
[01:24:23] chips are going to be way better if we're allowed to compete. And another is
[01:24:25] we're allowed to compete. And another is that they would be doing the same exact
[01:24:26] that they would be doing the same exact thing without us anyways. Right? How can
[01:24:28] thing without us anyways. Right? How can those two things be the same true at the
[01:24:29] those two things be the same true at the same time?
[01:24:30] same time? >> It's obviously true. In the absence of a
[01:24:34] >> It's obviously true. In the absence of a better choice, you'll take the only
[01:24:35] better choice, you'll take the only choice you have. How is that illogical?
[01:24:38] choice you have. How is that illogical? It's so logical.
[01:24:39] It's so logical. >> The reason they want Nvidia chips is
[01:24:40] >> The reason they want Nvidia chips is they're better. Better is more compute.
[01:24:42] they're better. Better is more compute. More comput means you can train a better
[01:24:43] More comput means you can train a better model.
[01:24:44] model. >> It's better. It's better because it's
[01:24:45] >> It's better. It's better because it's easier to program. It's e we have a
[01:24:47] easier to program. It's e we have a better ecosystem. Whatever the better
[01:24:49] better ecosystem. Whatever the better is. Whatever the better is. And of
[01:24:52] is. Whatever the better is. And of course we're going to send them compute.
[01:24:53] course we're going to send them compute. So what? So what the fact of the matter
[01:24:57] So what? So what the fact of the matter is we get the benefit. Don't forget we
[01:25:00] is we get the benefit. Don't forget we get the benefit of American technology
[01:25:02] get the benefit of American technology leadership. We get the benefit of
[01:25:04] leadership. We get the benefit of developers working on the American tech
[01:25:06] developers working on the American tech stack. We get the benefit as those AI
[01:25:08] stack. We get the benefit as those AI models diffuse out into the rest of the
[01:25:11] models diffuse out into the rest of the world. The American tech stack is
[01:25:13] world. The American tech stack is therefore the best for it. We can
[01:25:15] therefore the best for it. We can continue to advance and diffuse American
[01:25:17] continue to advance and diffuse American technology that I believe is a positive.
[01:25:21] technology that I believe is a positive. It's a very important part of American
[01:25:23] It's a very important part of American technology leadership. Now the policy
[01:25:26] technology leadership. Now the policy that you're advocating resulted in the
[01:25:28] that you're advocating resulted in the American telecommunication industry
[01:25:30] American telecommunication industry being policied out of basically the
[01:25:33] being policied out of basically the world to the point where we don't
[01:25:35] world to the point where we don't control our own telecommunications
[01:25:36] control our own telecommunications anymore. I don't see that as smart.
[01:25:40] anymore. I don't see that as smart. It's a little narrow-minded and it led
[01:25:42] It's a little narrow-minded and it led to un unintended consequences that I'm
[01:25:44] to un unintended consequences that I'm describing to you right now that you
[01:25:46] describing to you right now that you seem you seem to have a very hard time
[01:25:47] seem you seem to have a very hard time understanding.
[01:25:48] understanding. >> Okay, let let's just step back. It it
[01:25:51] >> Okay, let let's just step back. It it seems like the crux here is
[01:25:52] seems like the crux here is >> there's a potential benefit and there's
[01:25:54] >> there's a potential benefit and there's a potential cost and we're desri we're
[01:25:56] a potential cost and we're desri we're trying to figure out is the benefit
[01:25:57] trying to figure out is the benefit worth the cost. I guess I'm trying to
[01:25:59] worth the cost. I guess I'm trying to get you to acknowledge the potential
[01:26:01] get you to acknowledge the potential cost that compute is an input to
[01:26:03] cost that compute is an input to training powerful models. powerful
[01:26:05] training powerful models. powerful models do have powerful, you know,
[01:26:07] models do have powerful, you know, offensive capabilities like cyber
[01:26:09] offensive capabilities like cyber attacks. It is a good thing that
[01:26:10] attacks. It is a good thing that American companies got to claim mythos
[01:26:12] American companies got to claim mythos level capabilities first and then now
[01:26:14] level capabilities first and then now they're going to hold off on those
[01:26:15] they're going to hold off on those capabilities so that the American
[01:26:16] capabilities so that the American companies and American government can
[01:26:18] companies and American government can make their software more protected
[01:26:20] make their software more protected before this level cap announced if China
[01:26:23] before this level cap announced if China had had more computer had more power
[01:26:24] had had more computer had more power comput if we could have had made a
[01:26:26] comput if we could have had made a mythos level model earlier and deployed
[01:26:28] mythos level model earlier and deployed it widely that would have been very bad.
[01:26:31] it widely that would have been very bad. One of the reasons that hasn't happened
[01:26:32] One of the reasons that hasn't happened is that we have more compute thanks to
[01:26:34] is that we have more compute thanks to companies like Nvidia in America. Um
[01:26:36] companies like Nvidia in America. Um that is a cost of sending to China. And
[01:26:40] that is a cost of sending to China. And so let's leave the benefit aside for a
[01:26:42] so let's leave the benefit aside for a second. Do you acknowledge that this is
[01:26:43] second. Do you acknowledge that this is a potential cost?
[01:26:45] a potential cost? I will also tell you the potential cost
[01:26:48] I will also tell you the potential cost is we allow one of the most important
[01:26:51] is we allow one of the most important layers of the AI stack, the chip layer
[01:26:55] layers of the AI stack, the chip layer to concede an entire market, the second
[01:26:58] to concede an entire market, the second largest in second largest market in the
[01:27:00] largest in second largest market in the world so that they could develop scale
[01:27:03] world so that they could develop scale so that they could develop their own
[01:27:04] so that they could develop their own ecosystem so that future AI models are
[01:27:08] ecosystem so that future AI models are optimized in a very different way than
[01:27:11] optimized in a very different way than the American tech stack. As AI diffuses
[01:27:14] the American tech stack. As AI diffuses out into the rest of the world,
[01:27:17] out into the rest of the world, their standards, their tech stack will
[01:27:21] their standards, their tech stack will become superior to ours because their
[01:27:23] become superior to ours because their models are open. I
[01:27:24] models are open. I >> I guess I just believe enough in
[01:27:26] >> I guess I just believe enough in Nvidia's kernel engineers and CUDA
[01:27:28] Nvidia's kernel engineers and CUDA engineers to think that they could
[01:27:29] engineers to think that they could optimize.
[01:27:29] optimize. >> AI is more than kernel optimization as
[01:27:31] >> AI is more than kernel optimization as you know,
[01:27:31] you know, >> of course, but there's so many things
[01:27:33] >> of course, but there's so many things you can do from distilling to a model
[01:27:35] you can do from distilling to a model that's well fit for your chips.
[01:27:36] that's well fit for your chips. >> We're going to do our best.
[01:27:37] >> We're going to do our best. >> You have all this software. I just hard
[01:27:39] >> You have all this software. I just hard to imagine that there's a long-term lock
[01:27:40] to imagine that there's a long-term lock in to Chinese ecosystem. They have this
[01:27:42] in to Chinese ecosystem. They have this like slightly better open source model
[01:27:43] like slightly better open source model for a while.
[01:27:44] for a while. >> China is the largest contributor to open
[01:27:46] >> China is the largest contributor to open source software in the world. Fact,
[01:27:51] right? China is the largest contributor
[01:27:54] right? China is the largest contributor to open models in the world. Fact.
[01:27:57] to open models in the world. Fact. Today it's built on the American tech
[01:27:59] Today it's built on the American tech stack and
[01:28:01] stack and fact. All five layers of the tech stack
[01:28:05] fact. All five layers of the tech stack for AI is important. United States ought
[01:28:07] for AI is important. United States ought to go win all five of them. They're all
[01:28:10] to go win all five of them. They're all important.
[01:28:12] important. The one that is the most important of
[01:28:14] The one that is the most important of course is the AI application layer. The
[01:28:18] course is the AI application layer. The layer that diffuses into society, the
[01:28:21] layer that diffuses into society, the one that uses it most will benefit from
[01:28:23] one that uses it most will benefit from this industrial revolution most.
[01:28:27] this industrial revolution most. But my point is that every a every layer
[01:28:29] But my point is that every a every layer has to succeed.
[01:28:31] has to succeed. If we if we scare this country into
[01:28:34] If we if we scare this country into thinking that AI is
[01:28:37] thinking that AI is somehow a nuclear bomb
[01:28:40] somehow a nuclear bomb so that everybody hates AI and
[01:28:43] so that everybody hates AI and everybody's afraid of AI,
[01:28:45] everybody's afraid of AI, I don't know how you're helping the
[01:28:48] I don't know how you're helping the United States, you're doing a
[01:28:49] United States, you're doing a disservice. If we scare everybody out of
[01:28:52] disservice. If we scare everybody out of doing software engineering jobs because
[01:28:54] doing software engineering jobs because it's going to kill every software
[01:28:55] it's going to kill every software engineering job and we don't have any
[01:28:57] engineering job and we don't have any software engineers as a result of that,
[01:28:59] software engineers as a result of that, we're doing a disservice to United
[01:29:00] we're doing a disservice to United States.
[01:29:01] States. If we scare everybody out of radiology,
[01:29:03] If we scare everybody out of radiology, so nobody wants to be a radiologist
[01:29:05] so nobody wants to be a radiologist because computer vision is completely
[01:29:06] because computer vision is completely free and no AI is going to do a worse
[01:29:09] free and no AI is going to do a worse job than a radiologist. And we we
[01:29:11] job than a radiologist. And we we misunderstand the difference between a
[01:29:13] misunderstand the difference between a job and the task the job of a
[01:29:15] job and the task the job of a radiologist patient care task to read a
[01:29:18] radiologist patient care task to read a scan. If we misunderstand that so
[01:29:20] scan. If we misunderstand that so profoundly and we scare everybody out of
[01:29:24] profoundly and we scare everybody out of going to radiology school, we're not
[01:29:26] going to radiology school, we're not going to have enough radiologists and
[01:29:27] going to have enough radiologists and good enough healthcare. And so I
[01:29:31] good enough healthcare. And so I I'm making the case
[01:29:34] I'm making the case that when you make these make a premise
[01:29:38] that when you make these make a premise that is so extreme, everything goes from
[01:29:41] that is so extreme, everything goes from zero or infinity.
[01:29:44] zero or infinity. We end up scaring people in a way that's
[01:29:47] We end up scaring people in a way that's just not true. Life is not like that.
[01:29:50] just not true. Life is not like that. Do I do we want United States to be
[01:29:52] Do I do we want United States to be first? Of course we do.
[01:29:54] first? Of course we do. Do we need do we do we need to be uh a
[01:29:58] Do we need do we do we need to be uh a leader in every layer of that stack?
[01:30:01] leader in every layer of that stack? Of course we do. Of course we do. Is
[01:30:05] Of course we do. Of course we do. Is today you're talking about mythos
[01:30:07] today you're talking about mythos because mythos is important. Sure.
[01:30:09] because mythos is important. Sure. That's fantastic. But in a few years
[01:30:11] That's fantastic. But in a few years time, I'm making you the prediction that
[01:30:14] time, I'm making you the prediction that when we want the American tech stack,
[01:30:16] when we want the American tech stack, when we want American technology to be
[01:30:18] when we want American technology to be diffused around the world, out to India,
[01:30:21] diffused around the world, out to India, out to the Middle East, out out to to
[01:30:23] out to the Middle East, out out to to Africa, out to Southeast Asia, when our
[01:30:27] Africa, out to Southeast Asia, when our country would like to export because we
[01:30:29] country would like to export because we would like to export our technology, we
[01:30:32] would like to export our technology, we would like to export our standards. On
[01:30:34] would like to export our standards. On that day, I want you and I to have that
[01:30:36] that day, I want you and I to have that same conversation again. And I will tell
[01:30:39] same conversation again. And I will tell you exactly about today's conversation
[01:30:41] you exactly about today's conversation about how your policy and how what you
[01:30:43] about how your policy and how what you imagined
[01:30:45] imagined literally cause the United States to
[01:30:46] literally cause the United States to concede the second largest market in the
[01:30:48] concede the second largest market in the world for no good reason at all. We
[01:30:52] world for no good reason at all. We shouldn't concede it. If we lose it, we
[01:30:55] shouldn't concede it. If we lose it, we lose it. But why do we concede it? Now,
[01:30:58] lose it. But why do we concede it? Now, nobody is advocating Nobody is
[01:31:00] nobody is advocating Nobody is advocating an all or nothing. Nobody's
[01:31:03] advocating an all or nothing. Nobody's advocating all or nothing, meaning we
[01:31:05] advocating all or nothing, meaning we ship everything to China at all times.
[01:31:07] ship everything to China at all times. Nobody's advocating that we should
[01:31:10] Nobody's advocating that we should always have the best technology here. We
[01:31:12] always have the best technology here. We should always have the most technology
[01:31:14] should always have the most technology here and the first.
[01:31:16] here and the first. But we should also try to compete and
[01:31:20] But we should also try to compete and win around the world. Both of those
[01:31:23] win around the world. Both of those things can simultaneously happen. It
[01:31:26] things can simultaneously happen. It requires some amount of nuance, some
[01:31:28] requires some amount of nuance, some amount of maturity instead of absolutes.
[01:31:32] amount of maturity instead of absolutes. The world is just not absolutes.
[01:31:34] The world is just not absolutes. >> Okay. the the argument hinges on they've
[01:31:37] >> Okay. the the argument hinges on they've built a they've built models that are
[01:31:39] built a they've built models that are specified for their architect their the
[01:31:41] specified for their architect their the best chips that they make in a few years
[01:31:42] best chips that they make in a few years and those chips get exported around the
[01:31:44] and those chips get exported around the world that sets a standard um because of
[01:31:47] world that sets a standard um because of EUV
[01:31:48] EUV um export controls as we said you're
[01:31:50] um export controls as we said you're going to move on to 1.6 6 nometer
[01:31:52] going to move on to 1.6 6 nometer there's still going to be on 7 nometer
[01:31:53] there's still going to be on 7 nometer even after a few years from now and it
[01:31:55] even after a few years from now and it might make sense that domestically they
[01:31:56] might make sense that domestically they would prefer hey we got so much energy
[01:31:58] would prefer hey we got so much energy we can manufacture sets scale we'll
[01:31:59] we can manufacture sets scale we'll still keep using 7 nmter but the
[01:32:01] still keep using 7 nmter but the exporting thing their 7 nanometer chips
[01:32:04] exporting thing their 7 nanometer chips have to be competitive against your 1.6
[01:32:07] have to be competitive against your 1.6 nmter chips and their models have to be
[01:32:10] nmter chips and their models have to be so far optimized for the 7 nometer it's
[01:32:11] so far optimized for the 7 nometer it's better to run their models on 7
[01:32:12] better to run their models on 7 nanometer than to run their models on
[01:32:15] nanometer than to run their models on your 1.6 6 nometer.
[01:32:16] your 1.6 6 nometer. >> Can we can we just look at the facts
[01:32:18] >> Can we can we just look at the facts then? Okay. Is Blackwell 50 times more
[01:32:23] then? Okay. Is Blackwell 50 times more advanced lithography than Hopper? Is it
[01:32:26] advanced lithography than Hopper? Is it 50 times?
[01:32:28] 50 times? Not even close.
[01:32:30] Not even close. I just kept saying it over and over
[01:32:32] I just kept saying it over and over again. Moore's law is dead. Between
[01:32:34] again. Moore's law is dead. Between Hopper and Blackwell from the
[01:32:36] Hopper and Blackwell from the transistors themselves, call it 75%. It
[01:32:40] transistors themselves, call it 75%. It was 3 years apart.
[01:32:43] was 3 years apart. 75%.
[01:32:45] 75%. Blackwell is 50 times
[01:32:48] Blackwell is 50 times hopper.
[01:32:49] hopper. My point is architecture matters.
[01:32:54] My point is architecture matters. Computer science matters. Semiconductor
[01:32:56] Computer science matters. Semiconductor physics matter as well. But computer
[01:32:59] physics matter as well. But computer science matters.
[01:33:00] science matters. AI the impact of AI largely comes from
[01:33:05] AI the impact of AI largely comes from the computing stack which is the reason
[01:33:07] the computing stack which is the reason why CUDA is so effective which is the
[01:33:08] why CUDA is so effective which is the reason why CUDA is so so so beloved.
[01:33:12] reason why CUDA is so so so beloved. It's it's a ecosystem a computing
[01:33:14] It's it's a ecosystem a computing architecture that allows for so much
[01:33:16] architecture that allows for so much flexibility that if you wanted to change
[01:33:18] flexibility that if you wanted to change an architecture completely create
[01:33:20] an architecture completely create something like create something like
[01:33:23] something like create something like diffusion create something you know
[01:33:26] diffusion create something you know that's disagregated you could do you
[01:33:27] that's disagregated you could do you could do so it's easy to do and so the
[01:33:31] could do so it's easy to do and so the fact of the matter is AI is about the
[01:33:33] fact of the matter is AI is about the stack above as much as it is about the
[01:33:36] stack above as much as it is about the architecture below to the extent that
[01:33:38] architecture below to the extent that that we have architectures and software
[01:33:41] that we have architectures and software stacks that optimized for our stack, for
[01:33:43] stacks that optimized for our stack, for our ecosystem. It is obviously good
[01:33:46] our ecosystem. It is obviously good because we started the conversation
[01:33:48] because we started the conversation today about how Nvidia's ecosystem is so
[01:33:50] today about how Nvidia's ecosystem is so rich, why people always love programming
[01:33:52] rich, why people always love programming on CUDA first. They do. They do and so
[01:33:56] on CUDA first. They do. They do and so do the researchers in China. But if we
[01:33:59] do the researchers in China. But if we are forced to leave China, if we're
[01:34:01] are forced to leave China, if we're forced to leave China, it would be it
[01:34:04] forced to leave China, it would be it would be well, first of all, it would
[01:34:06] would be well, first of all, it would it's a policy mistake. obviously has
[01:34:08] it's a policy mistake. obviously has backlash has has backlash. Obviously, it
[01:34:12] backlash has has backlash. Obviously, it has fired, you know, has has uh uh has
[01:34:16] has fired, you know, has has uh uh has turned out badly for for the United
[01:34:18] turned out badly for for the United States. It enabled it accelerated their
[01:34:21] States. It enabled it accelerated their chip industry. It forced all of their AI
[01:34:24] chip industry. It forced all of their AI ecosystem to focus on their internal
[01:34:26] ecosystem to focus on their internal architectures. It's not too late, but
[01:34:29] architectures. It's not too late, but nonetheless,
[01:34:30] nonetheless, it has already happened.
[01:34:33] it has already happened. You're going to see in the future
[01:34:35] You're going to see in the future they're not stuck at 7 nanometer.
[01:34:37] they're not stuck at 7 nanometer. Obviously they're good at manufacturing.
[01:34:39] Obviously they're good at manufacturing. They will continue to advance from seven
[01:34:42] They will continue to advance from seven and beyond. Now
[01:34:45] and beyond. Now is there 10x difference between 5nanmter
[01:34:50] is there 10x difference between 5nanmter and 7 nanometer? The answer is no.
[01:34:53] and 7 nanometer? The answer is no. Architecture matters. Networking
[01:34:55] Architecture matters. Networking matters. That's why Nvidia bought
[01:34:57] matters. That's why Nvidia bought Melanox. Networking matters. Energy
[01:34:59] Melanox. Networking matters. Energy matters. And so all that stuff matters.
[01:35:01] matters. And so all that stuff matters. It's not it's not simplistic like the
[01:35:04] It's not it's not simplistic like the way you're trying to distill it.
[01:35:06] way you're trying to distill it. >> Uh we can move on from China, but that
[01:35:07] >> Uh we can move on from China, but that actually raises an interesting question
[01:35:09] actually raises an interesting question about um we were discussing earlier
[01:35:11] about um we were discussing earlier these bottlenecks at TSMC and memory and
[01:35:14] these bottlenecks at TSMC and memory and so forth. And so if we're in this world
[01:35:17] so forth. And so if we're in this world where you know you're already the
[01:35:18] where you know you're already the majority of N3 at some point you'll be
[01:35:21] majority of N3 at some point you'll be N2, you'll be a majority of that. Do you
[01:35:24] N2, you'll be a majority of that. Do you see that you could go back to N7 this
[01:35:27] see that you could go back to N7 this spare capacity at an older process node
[01:35:28] spare capacity at an older process node and say hey the demand for AI is so
[01:35:31] and say hey the demand for AI is so great and our capacity to expand the
[01:35:33] great and our capacity to expand the leading edge is not meeting it so we're
[01:35:36] leading edge is not meeting it so we're going to make a hopper or ampier about
[01:35:38] going to make a hopper or ampier about everything we know about a numeric today
[01:35:40] everything we know about a numeric today and all the other improvements you
[01:35:41] and all the other improvements you described do you see that world
[01:35:42] described do you see that world happening within before 2030
[01:35:45] happening within before 2030 >> it's not necessary to and the reason for
[01:35:47] >> it's not necessary to and the reason for that is because with every every
[01:35:50] that is because with every every generation the architecture
[01:35:53] generation the architecture the architecture um is more than just is
[01:35:58] the architecture um is more than just is more than just uh the transistor scale.
[01:36:02] more than just uh the transistor scale. It also you're doing so much engineering
[01:36:04] It also you're doing so much engineering and packaging and stacking and and the
[01:36:07] and packaging and stacking and and the numeric and you know the system
[01:36:09] numeric and you know the system architecture
[01:36:11] architecture um
[01:36:13] um when you run out of capacity
[01:36:16] when you run out of capacity to easily go back to another node that's
[01:36:18] to easily go back to another node that's a level of R R&D that that no one no one
[01:36:22] a level of R R&D that that no one no one could afford. You know we we could
[01:36:23] could afford. You know we we could afford to lean forward. I don't think we
[01:36:25] afford to lean forward. I don't think we could afford to go back. Now, if the
[01:36:27] could afford to go back. Now, if the world simply says, if on that day, if on
[01:36:30] world simply says, if on that day, if on that day, let's do the thought
[01:36:31] that day, let's do the thought experiment. On that day, we go, listen,
[01:36:33] experiment. On that day, we go, listen, we're just never going to have more
[01:36:34] we're just never going to have more capacity ever again, would I go back and
[01:36:37] capacity ever again, would I go back and use seven in a heartbeat?
[01:36:39] use seven in a heartbeat? >> Yeah, of course I would.
[01:36:40] >> Yeah, of course I would. >> Um,
[01:36:42] >> Um, one question somebody I was talking to
[01:36:43] one question somebody I was talking to had is why Nvidia doesn't run multiple
[01:36:46] had is why Nvidia doesn't run multiple different chip projects at the same time
[01:36:48] different chip projects at the same time with totally different architectures. So
[01:36:50] with totally different architectures. So you could do like a cerebra style
[01:36:52] you could do like a cerebra style >> wafer scale. You could do a dojo style
[01:36:54] >> wafer scale. You could do a dojo style huge package. You could do one without
[01:36:55] huge package. You could do one without CUDA, you know. Um you have the
[01:36:57] CUDA, you know. Um you have the resources and the engineering talent
[01:36:59] resources and the engineering talent >> to do all these in parallel. So why put
[01:37:02] >> to do all these in parallel. So why put all the eggs in one basket given who
[01:37:03] all the eggs in one basket given who knows where AI might go and
[01:37:04] knows where AI might go and architectures might go.
[01:37:06] architectures might go. >> Oh, we could. It's just that that we
[01:37:08] >> Oh, we could. It's just that that we don't have a better idea.
[01:37:10] don't have a better idea. >> Yeah. Yeah, we we could do all of those
[01:37:12] >> Yeah. Yeah, we we could do all of those things. Um
[01:37:14] things. Um it's just not better. And we simulate it
[01:37:17] it's just not better. And we simulate it all. they're in our simulator provably
[01:37:19] all. they're in our simulator provably worse
[01:37:21] worse and so we wouldn't do it.
[01:37:23] and so we wouldn't do it. Yeah, we're we're doing we're working on
[01:37:26] Yeah, we're we're doing we're working on exactly the projects that we want to
[01:37:27] exactly the projects that we want to work on. And and um I
[01:37:32] work on. And and um I if the workload were to change
[01:37:34] if the workload were to change dramatically
[01:37:36] dramatically um and I don't mean I don't mean the
[01:37:37] um and I don't mean I don't mean the algorithms, I actually mean the
[01:37:39] algorithms, I actually mean the workload. The um and that that depends
[01:37:42] workload. The um and that that depends on the shape of the market.
[01:37:46] on the shape of the market. um uh we may decide to add other
[01:37:48] um uh we may decide to add other accelerators like for example recently
[01:37:50] accelerators like for example recently we added uh Grock um and we're going to
[01:37:53] we added uh Grock um and we're going to fold Grock into our CUDA ecosystem
[01:37:56] fold Grock into our CUDA ecosystem and and um uh we do we're we're doing
[01:38:00] and and um uh we do we're we're doing that now because the value of tokens
[01:38:04] that now because the value of tokens um have gone up so high that that you
[01:38:07] um have gone up so high that that you could have different pricing of tokens.
[01:38:09] could have different pricing of tokens. Back in the old days in the, you know,
[01:38:10] Back in the old days in the, you know, just a couple years ago, tokens are
[01:38:12] just a couple years ago, tokens are either free or barely, you know, barely
[01:38:14] either free or barely, you know, barely expensive, right? And so, but now you
[01:38:16] expensive, right? And so, but now you can have different customers and those
[01:38:18] can have different customers and those customers want different answers. And
[01:38:20] customers want different answers. And so, because the customers make so much
[01:38:22] so, because the customers make so much money, like for example, our software
[01:38:24] money, like for example, our software engineers, if I can give them much more
[01:38:28] engineers, if I can give them much more um responsive tokens so that they're
[01:38:31] um responsive tokens so that they're even more productive than they are
[01:38:32] even more productive than they are today, I would pay for it.
[01:38:35] today, I would pay for it. >> But that market has only recently
[01:38:36] >> But that market has only recently emerged. And so I think that we now have
[01:38:40] emerged. And so I think that we now have we now have the ability to have the same
[01:38:42] we now have the ability to have the same model based on the response time have
[01:38:45] model based on the response time have different segments and that's the reason
[01:38:47] different segments and that's the reason why we decided to expand the paro
[01:38:50] why we decided to expand the paro frontier and and create a segment of
[01:38:54] frontier and and create a segment of inference that is faster response time
[01:38:57] inference that is faster response time even though it's lower lower throughput
[01:38:59] even though it's lower lower throughput at the mo until now higher throughput is
[01:39:02] at the mo until now higher throughput is always better. Um we we think that there
[01:39:04] always better. Um we we think that there there could be a world where there could
[01:39:06] there could be a world where there could be very high ASP tokens and and um even
[01:39:11] be very high ASP tokens and and um even though the even though the throughput is
[01:39:12] though the even though the throughput is lower in the factory the ASPs make up
[01:39:15] lower in the factory the ASPs make up for it.
[01:39:16] for it. >> Yeah. That's the reason why we did it.
[01:39:17] >> Yeah. That's the reason why we did it. But otherwise from an architecture
[01:39:19] But otherwise from an architecture perspective um I I think Nvidia's
[01:39:21] perspective um I I think Nvidia's architecture is I would I would rather
[01:39:23] architecture is I would I would rather put if I if I have more money I put more
[01:39:26] put if I if I have more money I put more behind the architecture. M I I think
[01:39:28] behind the architecture. M I I think this idea of extremely premium tokens
[01:39:30] this idea of extremely premium tokens and just the disagregation of the
[01:39:32] and just the disagregation of the inference market is very interesting.
[01:39:34] inference market is very interesting. >> The segmentation y final question um
[01:39:39] >> The segmentation y final question um supposed deep learning if revolution
[01:39:40] supposed deep learning if revolution didn't happen. Um what would Nvidia be
[01:39:44] didn't happen. Um what would Nvidia be doing? Obviously games but given
[01:39:48] doing? Obviously games but given >> accelerated computing
[01:39:50] >> accelerated computing >> accelerated computing the same thing
[01:39:52] >> accelerated computing the same thing we've been doing all along. I the the
[01:39:55] we've been doing all along. I the the premise of our company is that Moors law
[01:39:57] premise of our company is that Moors law Moore's law is going to more general
[01:39:59] Moore's law is going to more general purpose computing is good for a lot of
[01:40:01] purpose computing is good for a lot of things but for a lot of computation is
[01:40:03] things but for a lot of computation is not ideal and so we combined an
[01:40:07] not ideal and so we combined an architecture called a GPU CUDA to a CPU
[01:40:11] architecture called a GPU CUDA to a CPU so that we can accelerate the workload
[01:40:13] so that we can accelerate the workload of the CPU and so different different
[01:40:16] of the CPU and so different different kernels of code or algorithms could be
[01:40:18] kernels of code or algorithms could be offloaded onto our GPU and as a result
[01:40:21] offloaded onto our GPU and as a result you speed up an an application by you
[01:40:23] you speed up an an application by you you know 100x 200x and where can you use
[01:40:26] you know 100x 200x and where can you use that? Um well obviously engineering and
[01:40:28] that? Um well obviously engineering and science and physics and you know so on
[01:40:30] science and physics and you know so on so data processing um uh computer
[01:40:34] so data processing um uh computer graphics image generation I mean all
[01:40:36] graphics image generation I mean all kinds of things even if AI doesn't exist
[01:40:38] kinds of things even if AI doesn't exist today Nvidia will be very very large
[01:40:40] today Nvidia will be very very large yeah and so so I think the the reason
[01:40:43] yeah and so so I think the the reason for that is is fairly fundamental which
[01:40:45] for that is is fairly fundamental which is which is the ability for general
[01:40:47] is which is the ability for general purpose computing to continue to scale
[01:40:50] purpose computing to continue to scale has largely run its course and the only
[01:40:53] has largely run its course and the only the the not the only way but the the way
[01:40:54] the the not the only way but the the way to do that is through domain specific
[01:40:57] to do that is through domain specific acceleration and one of the domain that
[01:41:00] acceleration and one of the domain that we started with was computer graphics
[01:41:03] we started with was computer graphics but many there are many many other
[01:41:05] but many there are many many other domains I mean there's you know you know
[01:41:07] domains I mean there's you know you know all kinds of uh scient particle physics
[01:41:10] all kinds of uh scient particle physics and fluids and you know and and so
[01:41:13] and fluids and you know and and so structured data processing all kinds of
[01:41:14] structured data processing all kinds of different types of of algorithms that
[01:41:16] different types of of algorithms that benefit from CUDA and so our our mission
[01:41:20] benefit from CUDA and so our our mission was uh really to bring accelerated
[01:41:23] was uh really to bring accelerated computing to the world and advance the
[01:41:25] computing to the world and advance the type of applications that general
[01:41:27] type of applications that general purpose computing can't do and scale to
[01:41:29] purpose computing can't do and scale to the level of of uh capability that helps
[01:41:32] the level of of uh capability that helps break through certain fields of science.
[01:41:35] break through certain fields of science. And and so some of the early
[01:41:37] And and so some of the early applications were uh molecular dynamics,
[01:41:40] applications were uh molecular dynamics, uh seismic processing for energy
[01:41:42] uh seismic processing for energy discovery,
[01:41:43] discovery, um uh image processing of course, uh and
[01:41:46] um uh image processing of course, uh and so all of those kind of fields where
[01:41:48] so all of those kind of fields where where general purpose computing is just
[01:41:50] where general purpose computing is just simply too inefficient to do so. And so
[01:41:53] simply too inefficient to do so. And so yeah, if if there was no AI, I would be
[01:41:55] yeah, if if there was no AI, I would be very sad. Um, but because of because of
[01:42:00] very sad. Um, but because of because of of the advances that we made in
[01:42:03] of the advances that we made in computing, we democratized deep
[01:42:05] computing, we democratized deep learning. We made it possible for any
[01:42:08] learning. We made it possible for any researcher, any scientist anywhere, any
[01:42:10] researcher, any scientist anywhere, any student to be able to access a PC or,
[01:42:13] student to be able to access a PC or, you know, a a GeForce adding card and
[01:42:16] you know, a a GeForce adding card and and uh do amazing science. And um uh
[01:42:20] and uh do amazing science. And um uh that that fundamental promise uh hasn't
[01:42:23] that that fundamental promise uh hasn't changed, not even a little bit. And so
[01:42:25] changed, not even a little bit. And so if you see GT if you watch GTC, there's
[01:42:28] if you see GT if you watch GTC, there's the whole beginning part of it, none of
[01:42:30] the whole beginning part of it, none of it's AI. That whole part of it with with
[01:42:33] it's AI. That whole part of it with with uh computational lithography or or uh
[01:42:37] uh computational lithography or or uh our quantum chemistry work or you know
[01:42:39] our quantum chemistry work or you know uh all of that stuff, data processing
[01:42:41] uh all of that stuff, data processing work, all of that stuff is is uh
[01:42:45] work, all of that stuff is is uh unrelated to AI and and and it's still
[01:42:48] unrelated to AI and and and it's still very important. I mean there's, you
[01:42:49] very important. I mean there's, you know, I I know that that AI is is very
[01:42:51] know, I I know that that AI is is very interesting and and quite exciting. Um
[01:42:54] interesting and and quite exciting. Um but but um there's a lot of people doing
[01:42:57] but but um there's a lot of people doing a lot of very important work that's not
[01:42:59] a lot of very important work that's not not AI related and tensors is not the
[01:43:01] not AI related and tensors is not the only way that you compute with
[01:43:03] only way that you compute with >> and um I and we want to help everybody.
[01:43:06] >> and um I and we want to help everybody. >> It doesn't. Thank you so much.
[01:43:08] >> It doesn't. Thank you so much. >> You're welcome. I enjoyed it. Me too.
[01:43:10] >> You're welcome. I enjoyed it. Me too. Sweet.

Full Transcript

https://www.youtube.com/watch?v=Hrbq66XqtCo
Translation: zh-CN

[00:00] 我们已经看到许多软件公司的估值暴跌，因为人们期望人工智能会使软件商品化。
[00:08] 而且有一种可能有些天真的思考方式，就是你看英伟达向台积电发送一个GDS2文件。
[00:13] 台积电制造逻辑芯片。
[00:16] 它制造开关。
[00:18] 然后它将它们与SK海力士、美光和三星制造的高带宽内存（HBM）一起封装。
[00:22] 然后它将其发送到台湾的一家ODM厂商，在那里组装机架。
[00:26] 所以英伟达本质上是在制造软件，而由其他人来制造硬件。
[00:28] 如果软件商品化了，英伟达会跟着商品化吗？
[00:32] 嗯，归根结底，总得有东西将电子转化为代币。
[00:38] 这种转化，嗯，没有将电子转化为代币的转化，呃，并且随着时间的推移使这些代币更有价值。
[00:48] 我我我不认为，我认为那种从电子到...很难很难，嗯，完全商品化这种转化。
[01:00] 从电子到代币的转变是一段如此不可思议的旅程，也是在创造代币。
[01:07] 你知道，这就像让一个分子比另一个分子更有价值一样。
[01:11] 让一个代币比另一个更有价值。
[01:15] 将一个代币变得有价值所包含的艺术性、工程学、科学和发明。
[01:21] 显然，我们正在实时关注它的发生。
[01:22] 所以，这种转变、制造，以及所有科学知识的融入，远未被深入理解，而且这段旅程也远未结束。
[01:38] 所以，我怀疑它会发生。
[01:41] 当然，我们会让它更有效率。
[01:42] 我的意思是，关于英伟达的整个事情，实际上你提出问题的方式就是我对我们公司的设想。
[01:49] 输入是电子，输出是代币。
[01:55] 中间是英伟达，我们的工作是尽我们所能做到必要的那样。
[02:02] 就是尽可能多地做必要的事情，尽可能少地做，以实现这一点。
[02:04] 尽可能少地做，以令人难以置信的能力完成这项转变。
[02:07] 转变以令人难以置信的能力完成，我所说的尽可能少是指。
[02:09] 能力，以及我所说的尽可能少是指任何我不需要的。
[02:11] 尽可能少地是指任何我不需要的。
[02:13] 我与某人合作，并将其作为我生态系统的一部分来完成。
[02:14] 我与某人合作，并将其作为我生态系统的一部分来完成。
[02:16] 我生态系统的一部分来完成。如果你看看今天的英伟达，我们可能拥有最大的合作伙伴生态系统，包括上游供应链和下游供应链。
[02:18] 看看今天的英伟达，我们可能拥有最大的合作伙伴生态系统，包括上游供应链和下游供应链。
[02:21] 最大的合作伙伴生态系统，包括上游供应链和下游供应链。
[02:22] 上游供应链、下游供应链。所有的计算机、计算机公司以及所有的应用程序开发者和所有的模型制作者，以及你知道的人工智能，如果你愿意的话，可以看作是一个五层的蛋糕，我们在整个五层都有生态系统，所以我们尽量少做，但事实证明，我们必须做的那部分是极其困难的，而且，嗯。
[02:25] 下游供应链。所有的计算机、计算机公司以及所有的应用程序开发者和所有的模型制作者，以及你知道的人工智能，如果你愿意的话，可以看作是一个五层的蛋糕，我们在整个五层都有生态系统，所以我们尽量少做，但事实证明，我们必须做的那部分是极其困难的，而且，嗯。
[02:26] 计算机公司以及所有的应用程序开发者和所有的模型制作者，以及你知道的人工智能，如果你愿意的话，可以看作是一个五层的蛋糕，我们在整个五层都有生态系统，所以我们尽量少做，但事实证明，我们必须做的那部分是极其困难的，而且，嗯。
[02:28] 应用程序开发者和所有的模型制作者，以及你知道的人工智能，如果你愿意的话，可以看作是一个五层的蛋糕，我们在整个五层都有生态系统，所以我们尽量少做，但事实证明，我们必须做的那部分是极其困难的，而且，嗯。
[02:31] 制作者，以及你知道的人工智能，如果你愿意的话，可以看作是一个五层的蛋糕，我们在整个五层都有生态系统，所以我们尽量少做，但事实证明，我们必须做的那部分是极其困难的，而且，嗯。
[02:34] 五层的蛋糕，如果你愿意的话，我们在整个五层都有生态系统，所以我们尽量少做，但事实证明，我们必须做的那部分是极其困难的，而且，嗯。
[02:36] 拥有整个五层的生态系统，所以我们尽量少做，但事实证明，我们必须做的那部分是极其困难的，而且，嗯。
[02:39] 层，所以我们尽量少做，但事实证明，我们必须做的那部分是极其困难的，而且，嗯。
[02:42] 尽可能少，但事实证明，我们必须做的那部分是极其困难的，而且，嗯。
[02:45] 事实证明是极其困难的，而且，嗯。
[02:46] 嗯，我认为那不会商品化，事实上，事实上，嗯。
[02:47] 我认为那不会商品化，事实上，事实上，嗯。
[02:50] 商品化，事实上，事实上，嗯。嗯，我也不认为企业软件公司，工具制造商，你知道，今天的大多数软件公司都是工具制造商，嗯，其中一些不是，嗯，但是。
[02:52] 嗯，我也不认为企业软件公司，工具制造商，你知道，今天的大多数软件公司都是工具制造商，嗯，其中一些不是，嗯，但是。
[02:55] 企业软件公司，工具制造商，你知道，今天的大多数软件公司都是工具制造商，嗯，其中一些不是，嗯，但是。
[02:57] 工具制造商，你知道，今天的大多数软件公司都是工具制造商，嗯，其中一些不是，嗯，但是。
[02:59] 今天的软件公司都是工具制造商，嗯，其中一些不是，嗯，但是。
[03:02] 制造商，嗯，有些不是，嗯，但有些是工作流。
[03:05] 有些是工作流，嗯，编纂。
[03:07] 嗯，编纂，你知道，系统，嗯，但对于很多公司来说，它们是工具制造商。
[03:10] 你知道，系统，嗯，但对于很多公司来说，它们是工具制造商。
[03:11] 例如，你知道，Excel是一个工具。
[03:13] PowerPoint是一个工具，呃，Cadence制造工具。
[03:15] Synopsys制造工具。
[03:18] 我，我实际上看到了与人们所见的相反的情况。
[03:21] 我认为代理的数量将呈指数级增长。
[03:24] 代理的数量将呈指数级增长。
[03:27] 工具用户的数量将呈指数级增长，而且很有可能。
[03:28] 呈指数级增长，而且很有可能所有这些工具的实例数量将激增。
[03:32] 所有这些工具的实例数量将激增。
[03:37] Synopsys Design Compiler的实例数量很可能会激增，并且数量会。
[03:39] Synopsys Design Compiler的实例数量很可能会激增，并且将要使用的代理数量会。
[03:41] 将要使用布局规划器以及我们所有布局工具和设计设计规则检查器的代理数量。
[03:45] 使用布局规划器以及我们所有布局工具和设计设计规则检查器。
[03:48] 今天我们受工程师数量限制的代理数量。
[03:49] 今天我们受工程师数量的限制。
[03:52] 明天，这些工程师将得到一群代理的支持。
[03:54] 我们将要探索。
[04:04] 代理。我们将探索你从未见过的设计空间。
[04:05] 你从未见过的设计空间，并想使用我们今天使用的工具。
[04:10] 所以，我认为工具的使用将导致这些软件公司飞速发展。
[04:14] 之所以还没有发生，是因为代理们还没有足够好地使用它们的工具。
[04:20] 所以，这些公司要么自己构建代理，要么代理将足够好，能够使用这些工具。
[04:25] 我认为这将是两者的结合。
[04:30] 嗯，我认为在你最新的文件中，你与人们的代工厂、内存封装有近千亿美元的采购承诺，然后半导体分析公司报告称，你将有2500亿美元的此类采购承诺。
[04:45] 所以一种解释是英伟达的模式实际上是你已经锁定了多年来这些稀缺的组件，你知道别人可能有一个加速器，但他们真的能获得内存来构建它吗？
[04:56] 他们真的能获得逻辑来构建它吗？
[04:59] 这确实是英伟达未来几年的主要模式。
[05:01] 嗯，这是我们可以做到的，但对别人来说很难的一件事。
[05:04] 我们可以做到别人难以做到的事情。
[05:07] 我们之所以能够做出巨大的承诺，是因为我们已经做出了巨大的承诺。
[05:12] 嗯，其中一些是明确的。
[05:14] 你提到的这些承诺，有些是隐含的。
[05:17] 嗯，例如，上游的许多投资是由我们的供应链做出的，因为我对首席执行官们说，“让我告诉你这个行业将有多大，让我向你解释为什么，让我和你一起分析，让我向你展示我所看到的。”
[05:33] 因此，通过告知、激励、与所有不同行业的首席执行官们保持一致的这个过程，他们愿意进行投资。
[05:47] 现在，他们为什么愿意为我而不是别人进行投资，原因在于他们知道我有能力购买他们的供应并通过我的下游销售。
[06:01] 英伟达的下游供应链和我们的下游需求如此之大，
[06:07] 我们下游的需求如此之大，他们愿意在上游进行投资。
[06:13] 所以，如果你看看 GTC，嗯，嗯，呃，你知道，人们对 GTC 的规模以及与会者感到惊叹，这是一个 360 度的，整个 AI 的宇宙都聚集在一个地方。
[06:24] 他们都聚集在一个地方，因为他们需要互相交流。
[06:28] 我把他们聚集在一起，以便下游可以看到上游。
[06:31] 上游可以看到下游，他们所有人都可以看到 AI 的所有进展，而且非常重要的是，他们都可以见到 AI 原住民和所有正在被建立的 AI 初创公司，以及所有正在发生的令人惊叹的事情，这样他们就可以亲眼看到我告诉他们的一切。
[06:49] 所以，我花了很多时间直接或间接向我们的供应链、我们的合作伙伴和我们的生态系统介绍我们面前的机遇。
[06:59] 你知道，我的大多数主题演讲，你知道，有些人总是说，你知道，Jensen 在大多数主题演讲中，就像一个接一个的公告。
[07:08] 一个又一个的公告，一个又一个的公告，一个又一个的公告，一个又一个。
[07:12] 公告。我们的主题演讲总有一部分。
[07:15] 我们的主题演讲总有一部分有点折磨人，因为它几乎像是教育，事实上。
[07:22] 这正是我所想的。我需要确保整个供应链。
[07:27] 确保整个供应链上下游的生态系统都明白。
[07:32] 明白有什么东西要来了，为什么会来，什么时候会来，会有多大，并且能够像我一样系统地思考它。
[07:41] 它。所以，我认为你所描述的模式，我们当然能够为未来构建。
[07:52] 当然能够为未来构建，如果我们接下来的几年规模达到万亿美元，我们就有供应链可以做到，并且触手可及。
[08:02] 供应链可以做到，并且触手可及，我们的业务速度，你知道，就像有现金流一样，有供应链流在那里，没有人会。
[08:10] 那里的链式流动，没有人会为 AR 架构构建供应链，如果业务转向的架构很低的话。
[08:17] 所以我们能够维持规模，仅仅是因为我们的下游需求如此之大，他们看到了这一点，并且都听说了。
[08:24] 他们看到了这一切的到来。
[08:26] 所以，这使我们能够以我们能够做到的规模，做我们能够做的事情。
[08:32] 我确实想更具体地了解上游是否能跟上。
[08:37] 嗯，多年来，你们的收入一直同比增长一倍。
[08:40] 你们提供的浮点运算量年复一年地增长了三倍多。
[08:44] 而且现在以这个规模翻一番确实是不可思议的。
[08:47] 没错。
[08:47] 所以，你看看逻辑芯片，假设你是台积电 N3 节点的最大客户，而且你是整个 AI 领域最大的客户之一，今年将占 N3 的 60%。
[08:59] 根据一些分析，明年将占 86%。
[09:01] 如果你是大多数，你如何翻一番？
[09:07] 嗯，你如何做到年复一年地这样做？
[09:09] 那么，我们现在是否处于一个这样的时期，即
[09:11] 我们现在是否处于一个增长率必须放缓的局面，即人工智能计算的增长率必须放缓？
[09:13] 人工智能计算的增长率是否必须因为上游而放缓？
[09:15] 你有没有办法绕过这些，你知道我们如何才能每年建造两倍于以往的晶圆厂？
[09:21] 最终？
[09:26] 是的，在某种程度上，嗯，瞬时需求大于世界范围内的上游和下游供应。
[09:37] 而且，在任何瞬间，任何情况下，我们都可能受到水管工数量的限制。
[09:43] 水管工。
[09:44] 嗯。
[09:46] 这确实发生了。
[09:47] 水管工被邀请参加明年的GTC。
[09:51] 是的。你知道，顺便说一句，这是个好主意。
[09:53] 但这是一个好条件。你想要一个市场，你想要一个行业，在这个行业里，瞬时需求大于行业的总供应量。
[10:01] 嗯，反之显然不太好。
[10:03] 如果我们相差太远，嗯，如果某一个特定的物品，某一个特定的组件离得太远，嗯，显然，显然，
[10:14] 走开，嗯，显然，显然这个行业会蜂拥而至。
[10:17] 所以，例如，注意到人们不再怎么谈论 co-ass 了。
[10:20] 是的。
[10:22] 原因是因为我们两年以来把它的生命力都榨干了，我们翻了一倍又一倍又一倍地翻倍，现在我认为我们处于一个相当不错的位置。
[10:31] 而台积电现在知道 co-ass 的供应必须跟上其余的逻辑需求和内存需求，所以他们正在扩展 co-ass，并且他们正在以与规模化逻辑相同的水平扩展您知道的未来封装技术，这太棒了，因为很长一段时间以来 co-ass 相当专业化，嗯，HBM 也相当专业化，但它们不再是专业化的了，人们现在意识到它们是主流计算技术。
[11:02] 嗯，然后当然，我们现在更能影响我们供应链的更大范围。
[11:10] 过去，嗯，您知道，在人工智能革命的初期，所有那些
[11:15] 在人工智能革命中，我现在说的所有事情，五年前我都在说，我现在说的，五年前我都在说，有些人相信并投资了，有些人相信并投资了。
[11:20] 例如，嗯，桑杰和，并且。
[11:23] 例如，嗯，桑杰和美光团队。
[11:25] 美光团队。
[11:25] 我仍然清楚地记得那次会议，我清楚地知道将会发生什么以及为什么会发生，以及今天的预测，他们真的加倍投入了，我们与他们合作，跨越LPDDR，跨越你知道的HBM内存，他们真的投资了，这显然对公司来说是巨大的。
[11:28] 会议非常清楚，我清楚地知道将会发生什么以及为什么会发生，以及今天的预测，他们真的加倍投入了，我们与他们合作，跨越LPDDR，跨越你知道的HBM内存，他们真的投资了，这显然对公司来说是巨大的。
[11:29] 清楚地知道将会发生什么以及为什么会发生，以及今天的预测，他们真的加倍投入了，我们与他们合作，跨越LPDDR，跨越你知道的HBM内存，他们真的投资了，这显然对公司来说是巨大的。
[11:31] 发生以及为什么会发生，以及今天的预测，他们真的加倍投入了，我们与他们合作，跨越LPDDR，跨越你知道的HBM内存，他们真的投资了，这显然对公司来说是巨大的。
[11:33] 以及今天的预测，他们真的加倍投入了，我们与他们合作，跨越LPDDR，跨越你知道的HBM内存，他们真的投资了，这显然对公司来说是巨大的。
[11:37] 今天的预测，他们真的加倍投入了，我们与他们合作，跨越LPDDR，跨越你知道的HBM内存，他们真的投资了，这显然对公司来说是巨大的。
[11:39] 加倍投入了，我们与他们合作，跨越LPDDR，跨越你知道的HBM内存，他们真的投资了，这显然对公司来说是巨大的。
[11:42] 他们，以及跨越LPDDR，跨越你知道的HBM内存，他们真的投资了，这显然对公司来说是巨大的。
[11:44] HBM内存，他们真的投资了，这显然对公司来说是巨大的。
[11:47] 它，这显然对公司来说是巨大的。
[11:47] 对公司来说是巨大的。
[11:50] 嗯，有些人来得晚一些，但他们现在都在这里了，所以我认为这些瓶颈中的每一个都会得到极大的关注，嗯，现在我们提前几年就开始预取瓶颈了。
[11:52] 有些人来得晚一些，但他们现在都在这里了，所以我认为这些瓶颈中的每一个都会得到极大的关注，嗯，现在我们提前几年就开始预取瓶颈了。
[11:54] 但他们现在都在这里了，所以我认为这些瓶颈中的每一个都会得到极大的关注，嗯，现在我们提前几年就开始预取瓶颈了。
[11:57] 我认为这些瓶颈中的每一个都会得到极大的关注，嗯，现在我们提前几年就开始预取瓶颈了。
[11:59] 这些瓶颈中的每一个都会得到极大的关注，嗯，现在我们提前几年就开始预取瓶颈了。
[12:02] 得到极大的关注，嗯，现在我们提前几年就开始预取瓶颈了。
[12:04] 现在我们提前几年就开始预取瓶颈了。
[12:06] 瓶颈，提前几年。
[12:06] 所以，例如，我们与Lum和Coherent所做的投资，以及整个硅光子生态系统，在过去几个月里
[12:09] 例如，我们与Lum和Coherent所做的投资，以及整个硅光子生态系统，在过去几个月里
[12:12] 我们与Lum和Coherent所做的投资，以及整个硅光子生态系统，在过去几个月里
[12:15] Coherent以及整个硅光子生态系统，在过去几个月里
[12:18] 光子生态系统，嗯，在过去的几年里，我们真的重塑了生态系统。
[12:20] 几年里，我们真的重塑了生态系统和供应链，硅光子学。
[12:23] 以及供应链，硅光子学。
[12:25] 我们，我们，呃，围绕台积电建立了一个完整的供应链。
[12:28] 我们与他们合作，嗯，发明了一大堆技术。
[12:30] 技术，我们授权了，呃，那些专利给供应链。
[12:33] 给供应链，保持它良好和开放。
[12:35] 嗯，所以我们正在通过发明新技术、新工作流程、呃，新测试、新测试设备来准备供应链。
[12:37] 新的测试设备，双面探测，嗯，投资公司，帮助他们扩大产能。
[12:39] 公司，帮助他们扩大产能。
[12:42] 嗯，所以，所以你可以看到我们正在努力塑造生态系统，使其准备就绪，供应链也准备就绪以支持规模化。
[12:43] 似乎有些瓶颈比其他瓶颈更容易解决。
[12:46] 所以扩大联合评估与扩大规模相比。
[12:48] 顺便说一句，我去了最难的一个。
[12:51] 那就是水管工。
[12:53] 是的，是真的。
[12:55] 是的。
[12:57] 是的，我实际上去了最难的一个。
[12:58] 是的。
[13:01] 是的，水管工和电工。
[13:02] 原因是因为。
[13:04] 因为这是我担心的其中一点，关于所有预言末日的人，描述末日的结束。
[13:20] 悲观主义者，嗯，描述了工作的终结，工作的消失。
[13:23] 你知道，如果我们阻止人们成为软件工程师，我们将面临软件工程师短缺。
[13:30] 而且，嗯，十年前也有同样的预测，一些悲观主义者说，我们告诉人们，无论你做什么，都不要成为放射科医生。
[13:40] 你可能还会看到一些视频仍然在网上。
[13:44] 你知道，放射学将是第一个消失的职业。
[13:48] 没有人会，世界将不再需要放射科医生。
[13:50] 猜猜怎么着？
[13:51] 但我们却缺少放射科医生。
[13:54] 哦，但是，好的。
[13:55] 所以，回到这个观点，有些东西你可以扩展，其他东西呢，比如你如何实际制造出每年两倍的逻辑量？
[14:02] 最终，这受内存和逻辑的瓶颈，或者受紫外线的瓶颈。
[14:05] 你如何才能一年内获得两倍的紫外线机器数量？
[14:09] 是的。
[14:10] 年复一年。
[14:10] 这一切都不是不可能快速扩展的。
[14:13] 你只需要，你可以在两到三年内完成所有这些事情，这很容易做到。
[14:17] 你只需要一个需求信号，表明它
[14:21] 你只需要一个需求信号，一旦你能制造一个，就不是它了。
[14:24] 一旦你能制造一个，就不是它了，你可以制造10个，一旦你能制造。
[14:25] 你可以制造10个，一旦你制造10个，你就可以制造100万个，所以。
[14:28] 制造10个，你可以制造100万个，所以这些东西并不难。
[14:29] 这些东西并不难复制。
[14:31] 在供应链的哪个环节，你会去ASML？
[14:34] 你会去ASML说，嘿，如果我展望未来三年。
[14:35] 说嘿，如果我展望未来三年，为了让我，为了英伟达一年能产生两万亿的收入。
[14:38] 一年产生两万亿的收入，我们需要更多的AUV机器。
[14:40] 收入，我们需要更多的AUV机器，而且。
[14:41] 而且>>其中一些我必须直接，呃，一些。
[14:44] >>其中一些我必须直接，呃，其中一些是间接的，还有一些。
[14:46] 其中一些是间接的，还有一些，嗯，如果我能说服台积电，就像ASML一样。
[14:49] 嗯，如果我能说服台积电，ASML就会被说服，所以你知道。
[14:51] 被说服，所以你知道我们必须考虑关键的。
[14:53] 我们必须考虑关键的关键瓶颈，但是如果台积电。
[14:56] 被说服，嗯，几年后你将拥有足够的EV机器。
[15:00] 几年后会有机器。所以没有。
[15:03] 我的观点是，没有一个瓶颈会持续超过两三年。
[15:05] 瓶颈持续时间超过两三年。没有一个。同时。
[15:07] 三年。没有一个。同时，我们正在，呃，提高计算能力。
[15:11] 同时，我们正在，呃，将计算效率提高10倍，20倍，在...
[15:13] 效率方面，从Hopper到Blackwell提高10倍，20倍，甚至30到50倍，嗯，我们正在。
[15:16] 从Hopper到Blackwell提高30到50倍，嗯，我们正在开发新的算法，因为。
[15:20] 正在开发新的算法，因为。
[15:22] 提出新的算法，因为 CUDA 非常灵活。
[15:25] CUDA 非常灵活。
[15:27] 呃，我们正在开发各种新技术，以提高效率。
[15:29] 呃，除了增加容量之外。
[15:31] 是的。
[15:33] 所以，那些事情都不能让我担心。
[15:36] 这是在我们下游的东西。
[15:38] 嗯，能源政策阻碍了能源，你知道，没有能源你就无法发展，无法创造，无法建立一个产业。
[15:44] 没有能源，你就无法创造一个全新的制造业。
[15:47] 呃，我们想让美国重新工业化。
[15:49] 我们想恢复呃芯片制造、计算机制造和封装，我们想建造电动汽车、机器人等新事物，我们想建造人工智能工厂，没有能源你就无法建造任何这些东西，这些事情需要很长时间，但更多的芯片产能是一个两三年的问题。
[16:11] 更多的粗略产能是两三年的问题。
[16:13] 有趣的是，我感觉有时客人会告诉我完全相反的事情，在这种情况下，我只是没有技术知识来评判，但是。
[16:20] 嗯，美妙之处在于你正在与专家交谈。
[16:21] 是的，没错，没错。
[16:25] 与专家交谈，是的，没错，没错。
[16:27] 嗯，好的。我想问一下关于你们的竞争对手。
[16:28] 竞争对手。
[16:28] 是的。
[16:28] 所以，如果你看TPU，可以说世界上排名前三的模型中有两个，Claude和Gemini，都是在TPU上训练的。
[16:39] 这对英伟达未来的发展意味着什么？
[16:41] 未来？
[16:43] 嗯，嗯，我们有一个非常不同的，我们构建了一个非常不同的东西。
[16:44] 嗯，你知道，英伟达构建的是加速计算。
[16:51] 不是张量处理单元。
[16:55] 呃，加速计算被用于各种各样的事情。
[16:57] 你知道，分子动力学和量子色动力学，它被用于数据处理、数据帧、结构化数据、非结构化数据。
[17:07] 它被用于流体动力学、粒子物理学，你知道，此外，我们还将其用于人工智能。
[17:17] 所以加速计算更加多样化，虽然人工智能是今天的谈话焦点，但显然非常
[17:25] 今天的对话显然非常重要且有影响力。嗯，计算很重要且有影响力。
[17:29] 嗯，计算的范围要广得多，而英伟达所做的就是重新发明了计算的执行方式，从通用计算到加速计算。
[17:34] 我们的市场覆盖范围远远大于任何 TPU 或任何 ASA 可能拥有的。
[17:46] 所以，如果你看看我们的地位，嗯，我们是唯一一家加速各种类型应用程序的公司。
[17:54] 我们拥有一个庞大的生态系统，所以各种框架和算法都在 Nvidia 上运行。
[18:04] 而且因为我们的计算机是为他人操作而设计的，任何操作员都可以购买我们的系统。
[18:13] 大多数这些自建系统，你必须自己操作，因为它从未被设计成足够灵活，供其他人操作。
[18:20] 因此，由于任何人都可以操作我们的系统，我们无处不在
[18:26] 运营我们的系统，我们支持所有云，包括谷歌和亚马逊，以及谷歌和亚马逊云，还有你知道的Azure和OCI，对吧，所以无论你想租用运营还是自己运营，如果你想租用运营，你最好拥有一个庞大的生态系统，在许多行业拥有客户，他们将是购买者。
[18:42] 如果你自己运营，如果你想自己运营，嗯，我们，你知道的，我们显然有能力帮助你自己运营，比如埃隆的XAI，嗯，因为我们可以，我们可以赋能任何行业中的任何公司的运营商，你可以用它来为科学研究和礼来公司的药物发现构建超级计算机，所以我们可以帮助他们运营自己的超级计算机，并将其用于药物发现和生命科学的整个多样性，嗯，我们加速了这一点>>
[19:20] 所以，所以有很多应用程序我们可以解决，而你无法用TPU做到这一点。
[19:28] 地址是您无法使用TPU完成的，因为Nvidia构建的CUDA也是一个出色的张量处理单元。
[19:34] 但它确实，你知道，它处理数据处理、计算和人工智能等所有生命周期。
[19:41] 所以我们的市场机会要大得多。
[19:43] 我们的影响力更大，因为我们拥有如此庞大的用户群，我们基本上支持世界上所有的应用程序。
[19:53] 现在，您可以在任何地方构建Nvidia系统，并且知道会有客户。
[19:58] 所以这是一个非常不同的事情。
[20:00] 呃，这会是一个有点长的问题，但你知道你们有惊人的收入，而这笔收入主要是，你们不是从制药和量子领域每季度赚取600亿美元。
[20:10] 你们之所以能做到，是因为人工智能是一项前所未有的技术，而且正在以前所未有的速度增长。
[20:14] 所以问题是，什么最适合人工智能，我并不了解具体细节。
[20:18] 但我与我的AI研究员朋友们交谈过，他们说，看，当我使用TPU时，它是一个巨大的并行阵列，非常适合进行矩阵乘法，而GPU则非常灵活。
[20:25] 当你...

Full Transcript (Bilingual)

https://www.youtube.com/watch?v=Hrbq66XqtCo
Translation: zh-CN

[00:00] We've seen the valuations of a bunch of software companies crash because people are expecting AI to commoditize software.
我们已经看到许多软件公司的估值暴跌，因为人们期望人工智能会使软件商品化。

[00:08] And there's a a potentially naive way of thinking about things which is like look Nvidia sends a GDS2 file to TSMC.
而且有一种可能有些天真的思考方式，就是你看英伟达向台积电发送一个GDS2文件。

[00:13] TSMC builds the logic dies.
台积电制造逻辑芯片。

[00:16] It builds the switches.
它制造开关。

[00:18] Um then it packages them with the HBM that SK Highex and Micron and Samsung make.
然后它将它们与SK海力士、美光和三星制造的高带宽内存（HBM）一起封装。

[00:22] Then it sends it to an ODM in Taiwan where they assemble the racks.
然后它将其发送到台湾的一家ODM厂商，在那里组装机架。

[00:26] And so Nvidia is fundamentally making software that other people are manufacturing.
所以英伟达本质上是在制造软件，而由其他人来制造硬件。

[00:28] And if software gets commoditized, does Nvidia get commoditized?
如果软件商品化了，英伟达会跟着商品化吗？

[00:32] Well, in the end, something has to transform electrons to tokens.
嗯，归根结底，总得有东西将电子转化为代币。

[00:38] That transformation um there's no the transformation of electrons to tokens uh and making those tokens more valuable over time.
这种转化，嗯，没有将电子转化为代币的转化，呃，并且随着时间的推移使这些代币更有价值。

[00:48] I I I don't I think that that that's hard to hard to um completely commoditize the transformation from electrons to
我我我不认为，我认为那种从电子到...很难很难，嗯，完全商品化这种转化。

[01:00] The transformation from electrons to tokens is such an such an incredible journey and and making that token.
从电子到代币的转变是一段如此不可思议的旅程，也是在创造代币。

[01:07] You know, it's like making a one molecule more valuable than another molecule.
你知道，这就像让一个分子比另一个分子更有价值一样。

[01:11] Making one token more valuable than another.
让一个代币比另一个更有价值。

[01:15] The amount of artistry, engineering, science, invention that goes into making that token valuable.
将一个代币变得有价值所包含的艺术性、工程学、科学和发明。

[01:21] Obviously, we're we're watching it happening in real time.
显然，我们正在实时关注它的发生。

[01:22] And so, so the the the the transformation, the manufacturing, um all of the science that goes in there is far from un deeply understood and it's far from the journey is far from far from over.
所以，这种转变、制造，以及所有科学知识的融入，远未被深入理解，而且这段旅程也远未结束。

[01:38] And so, so I I doubt that it will happen.
所以，我怀疑它会发生。

[01:41] Um we're going to make it more efficient, of course.
当然，我们会让它更有效率。

[01:42] I mean the whole the whole thing about Nvidia in fact the way that you frame the question is is my mental model of our company.
我的意思是，关于英伟达的整个事情，实际上你提出问题的方式就是我对我们公司的设想。

[01:49] The input is electron the output is tokens.
输入是电子，输出是代币。

[01:55] That is in the middle Nvidia and our job is to to do as much as necessary as
中间是英伟达，我们的工作是尽我们所能做到必要的那样。

[02:02] is to to do as much as necessary as little as possible to enable that.
就是尽可能多地做必要的事情，尽可能少地做，以实现这一点。

[02:04] little as possible to enable that transformation to be done at incredible capabilities.
尽可能少地做，以令人难以置信的能力完成这项转变。

[02:07] transformation to be done at incredible capabilities and and what I mean by as little as possible.
转变以令人难以置信的能力完成，我所说的尽可能少是指。

[02:09] capabilities and and what I mean by as little as possible whatever I don't need to.
能力，以及我所说的尽可能少是指任何我不需要的。

[02:11] little as possible whatever I don't need to.
尽可能少地是指任何我不需要的。

[02:13] to I partner with somebody and I make it part of my ecosystem to do.
我与某人合作，并将其作为我生态系统的一部分来完成。

[02:14] I partner with somebody and I make it part of my ecosystem to do.
我与某人合作，并将其作为我生态系统的一部分来完成。

[02:16] part of my ecosystem to do. And if you look at Nvidia today, we probably have the largest ecosystem of partners both in supply chain upstream, supply chain downstream.
我生态系统的一部分来完成。如果你看看今天的英伟达，我们可能拥有最大的合作伙伴生态系统，包括上游供应链和下游供应链。

[02:18] look at Nvidia today, we probably have the largest ecosystem of partners both in supply chain upstream, supply chain downstream.
看看今天的英伟达，我们可能拥有最大的合作伙伴生态系统，包括上游供应链和下游供应链。

[02:21] the largest ecosystem of partners both in supply chain upstream, supply chain downstream.
最大的合作伙伴生态系统，包括上游供应链和下游供应链。

[02:22] in supply chain upstream, supply chain downstream. all of the computers, computer companies and all the application developers and all the model makers and all the you know AI is a five five layer cake if you will and and we have ecosystems across the entire five layers and and so we try to do as little as possible but the part that we have to do as it turns out is insanely hard and and um.
上游供应链、下游供应链。所有的计算机、计算机公司以及所有的应用程序开发者和所有的模型制作者，以及你知道的人工智能，如果你愿意的话，可以看作是一个五层的蛋糕，我们在整个五层都有生态系统，所以我们尽量少做，但事实证明，我们必须做的那部分是极其困难的，而且，嗯。

[02:25] downstream. all of the computers, computer companies and all the application developers and all the model makers and all the you know AI is a five five layer cake if you will and and we have ecosystems across the entire five layers and and so we try to do as little as possible but the part that we have to do as it turns out is insanely hard and and um.
下游供应链。所有的计算机、计算机公司以及所有的应用程序开发者和所有的模型制作者，以及你知道的人工智能，如果你愿意的话，可以看作是一个五层的蛋糕，我们在整个五层都有生态系统，所以我们尽量少做，但事实证明，我们必须做的那部分是极其困难的，而且，嗯。

[02:26] computer companies and all the application developers and all the model makers and all the you know AI is a five five layer cake if you will and and we have ecosystems across the entire five layers and and so we try to do as little as possible but the part that we have to do as it turns out is insanely hard and and um.
计算机公司以及所有的应用程序开发者和所有的模型制作者，以及你知道的人工智能，如果你愿意的话，可以看作是一个五层的蛋糕，我们在整个五层都有生态系统，所以我们尽量少做，但事实证明，我们必须做的那部分是极其困难的，而且，嗯。

[02:28] application developers and all the model makers and all the you know AI is a five five layer cake if you will and and we have ecosystems across the entire five layers and and so we try to do as little as possible but the part that we have to do as it turns out is insanely hard and and um.
应用程序开发者和所有的模型制作者，以及你知道的人工智能，如果你愿意的话，可以看作是一个五层的蛋糕，我们在整个五层都有生态系统，所以我们尽量少做，但事实证明，我们必须做的那部分是极其困难的，而且，嗯。

[02:31] makers and all the you know AI is a five five layer cake if you will and and we have ecosystems across the entire five layers and and so we try to do as little as possible but the part that we have to do as it turns out is insanely hard and and um.
制作者，以及你知道的人工智能，如果你愿意的话，可以看作是一个五层的蛋糕，我们在整个五层都有生态系统，所以我们尽量少做，但事实证明，我们必须做的那部分是极其困难的，而且，嗯。

[02:34] five layer cake if you will and and we have ecosystems across the entire five layers and and so we try to do as little as possible but the part that we have to do as it turns out is insanely hard and and um.
五层的蛋糕，如果你愿意的话，我们在整个五层都有生态系统，所以我们尽量少做，但事实证明，我们必须做的那部分是极其困难的，而且，嗯。

[02:36] have ecosystems across the entire five layers and and so we try to do as little as possible but the part that we have to do as it turns out is insanely hard and and um.
拥有整个五层的生态系统，所以我们尽量少做，但事实证明，我们必须做的那部分是极其困难的，而且，嗯。

[02:39] layers and and so we try to do as little as possible but the part that we have to do as it turns out is insanely hard and and um.
层，所以我们尽量少做，但事实证明，我们必须做的那部分是极其困难的，而且，嗯。

[02:42] as possible but the part that we have to do as it turns out is insanely hard and and um.
尽可能少，但事实证明，我们必须做的那部分是极其困难的，而且，嗯。

[02:45] do as it turns out is insanely hard and and um.
事实证明是极其困难的，而且，嗯。

[02:46] and um >> I I don't think that that gets commoditized in fact in fact um.
嗯，我认为那不会商品化，事实上，事实上，嗯。

[02:47] >>> I I don't think that that gets commoditized in fact in fact um.
我认为那不会商品化，事实上，事实上，嗯。

[02:50] commoditized in fact in fact um >> uh I also don't think that the the enterprise software ware companies uh the tools makers you know most of the software companies today are tools makers um some of them are not um but.
商品化，事实上，事实上，嗯。嗯，我也不认为企业软件公司，工具制造商，你知道，今天的大多数软件公司都是工具制造商，嗯，其中一些不是，嗯，但是。

[02:52] uh I also don't think that the the enterprise software ware companies uh the tools makers you know most of the software companies today are tools makers um some of them are not um but.
嗯，我也不认为企业软件公司，工具制造商，你知道，今天的大多数软件公司都是工具制造商，嗯，其中一些不是，嗯，但是。

[02:55] enterprise software ware companies uh the tools makers you know most of the software companies today are tools makers um some of them are not um but.
企业软件公司，工具制造商，你知道，今天的大多数软件公司都是工具制造商，嗯，其中一些不是，嗯，但是。

[02:57] the tools makers you know most of the software companies today are tools makers um some of them are not um but.
工具制造商，你知道，今天的大多数软件公司都是工具制造商，嗯，其中一些不是，嗯，但是。

[02:59] software companies today are tools makers um some of them are not um but.
今天的软件公司都是工具制造商，嗯，其中一些不是，嗯，但是。

[03:02] Makers, um, some of them are not, um, but are some of them are workflow.
制造商，嗯，有些不是，嗯，但有些是工作流。

[03:05] Are some of them are workflow, um, codification.
有些是工作流，嗯，编纂。

[03:07] Um, codification, you know, systems, um, but for a lot of companies, they're tool makers.
嗯，编纂，你知道，系统，嗯，但对于很多公司来说，它们是工具制造商。

[03:10] You know, systems, um, but for a lot of companies, they're tool makers.
你知道，系统，嗯，但对于很多公司来说，它们是工具制造商。

[03:11] For example, you know, Excel is a tool.
例如，你知道，Excel是一个工具。

[03:13] PowerPoint's a tool, uh, Cadence makes tools.
PowerPoint是一个工具，呃，Cadence制造工具。

[03:15] Synopsis makes tools.
Synopsys制造工具。

[03:18] I, I actually see the opposite of what people see.
我，我实际上看到了与人们所见的相反的情况。

[03:21] I think the number of agents are going to grow exponentially.
我认为代理的数量将呈指数级增长。

[03:24] The number of agents are going to grow exponentially.
代理的数量将呈指数级增长。

[03:27] The number of tool users are going to grow exponentially and it's very likely that.
工具用户的数量将呈指数级增长，而且很有可能。

[03:28] Exponentially and it's very likely that the number of instances of all these tools are going to skyrocket.
呈指数级增长，而且很有可能所有这些工具的实例数量将激增。

[03:32] The number of instances of all these tools are going to skyrocket.
所有这些工具的实例数量将激增。

[03:37] It is very likely the number of instances of Synopsys Design Compiler is going to skyrocket and the number of.
Synopsys Design Compiler的实例数量很可能会激增，并且数量会。

[03:39] It is very likely the number of instances of Synopsys Design Compiler is going to skyrocket and the number of number of agents that are going to be.
Synopsys Design Compiler的实例数量很可能会激增，并且将要使用的代理数量会。

[03:41] Number of agents that are going to be using the floor planners and all of our layout tools and our design design rule checkers.
将要使用布局规划器以及我们所有布局工具和设计设计规则检查器的代理数量。

[03:45] Using the floor planners and all of our layout tools and our design design rule checkers.
使用布局规划器以及我们所有布局工具和设计设计规则检查器。

[03:48] The number of agents that are today we're limited by the number of engineers.
今天我们受工程师数量限制的代理数量。

[03:49] Today we're limited by the number of engineers.
今天我们受工程师数量的限制。

[03:52] Tomorrow those engineers are going to be supported by a bunch of agents.
明天，这些工程师将得到一群代理的支持。

[03:54] We're going to be exploring out.
我们将要探索。

[04:04] Agents. We're going to be exploring out the design space like you've never seen.
代理。我们将探索你从未见过的设计空间。

[04:05] The design space like you've never seen explore before and want to use the tools that we use today.
你从未见过的设计空间，并想使用我们今天使用的工具。

[04:10] And so, so I think I think tool use is going to cause cause these software companies to skyrocket.
所以，我认为工具的使用将导致这些软件公司飞速发展。

[04:14] The reason why it hasn't happened yet is because the agents aren't good enough at using their tools yet.
之所以还没有发生，是因为代理们还没有足够好地使用它们的工具。

[04:20] And so either these companies are going to build the agents themselves or agents are going to get good enough to be able to use those tools.
所以，这些公司要么自己构建代理，要么代理将足够好，能够使用这些工具。

[04:25] And I think it's going to be a combination of both.
我认为这将是两者的结合。

[04:30] Um I think in your latest filings it was you had almost hundred billion dollars in purchase commitments with people foundries memory packaging and then uh semi analysis has reported that you will have $250 billion of these kinds of purchase commitments.
嗯，我认为在你最新的文件中，你与人们的代工厂、内存封装有近千亿美元的采购承诺，然后半导体分析公司报告称，你将有2500亿美元的此类采购承诺。

[04:45] And so one interpretation is Nvidia's mode is really that you've locked up many years of these scarce components that are uh you know somebody else might have an accelerator but can they actually get the memory to build it?
所以一种解释是英伟达的模式实际上是你已经锁定了多年来这些稀缺的组件，你知道别人可能有一个加速器，但他们真的能获得内存来构建它吗？

[04:56] Can they actually get the logic to build it?
他们真的能获得逻辑来构建它吗？

[04:59] And this is really Nvidia's big mode for the next few years.
这确实是英伟达未来几年的主要模式。

[05:01] Well, it it's one it's one of the things that we can do that is hard for someone
嗯，这是我们可以做到的，但对别人来说很难的一件事。

[05:04] that we can do that is hard for someone else to do.
我们可以做到别人难以做到的事情。

[05:07] The reason why we could we we've made enormous commitments upstream.
我们之所以能够做出巨大的承诺，是因为我们已经做出了巨大的承诺。

[05:12] Um some of it is explicit.
嗯，其中一些是明确的。

[05:14] These commitments that you mentioned, some of it is implicit.
你提到的这些承诺，有些是隐含的。

[05:17] Um, for example, a lot of the investments that are upstream are made by our our supply chain because I said to the CEOs, "Let me tell you how big this industry is going to be and let me explain to you why and let me reason through it with you and let me show you what I see."
嗯，例如，上游的许多投资是由我们的供应链做出的，因为我对首席执行官们说，“让我告诉你这个行业将有多大，让我向你解释为什么，让我和你一起分析，让我向你展示我所看到的。”

[05:33] And so as a result of that that process of of uh informing inspiring um aligning with CEOs of all different industries upstream they're willing to make the investments.
因此，通过告知、激励、与所有不同行业的首席执行官们保持一致的这个过程，他们愿意进行投资。

[05:47] Now why are they willing to make the investments for me and not someone else and the reason for that is because they know that I have the capacity to buy it buy their supply and sell it through my downstream.
现在，他们为什么愿意为我而不是别人进行投资，原因在于他们知道我有能力购买他们的供应并通过我的下游销售。

[06:01] the fact that Nvidia's downstream supply chain and our downstream demand is so large,
英伟达的下游供应链和我们的下游需求如此之大，

[06:07] and our downstream demand is so large, they're willing to make the investment upstream.
我们下游的需求如此之大，他们愿意在上游进行投资。

[06:13] And so if you look at GTC um and and uh you know, people are marveled by the scale of GTC and the people that go, it's a 360° that the entire universe of AI all in one place.
所以，如果你看看 GTC，嗯，嗯，呃，你知道，人们对 GTC 的规模以及与会者感到惊叹，这是一个 360 度的，整个 AI 的宇宙都聚集在一个地方。

[06:24] and they they're all in one place because they need to see each other.
他们都聚集在一个地方，因为他们需要互相交流。

[06:28] I bring them together so that the the downstream could see the upstream.
我把他们聚集在一起，以便下游可以看到上游。

[06:31] The upstream could see the downstream and all of them could see all the advances in AI and very importantly they can all meet the AI natives and all the AI startups that are all you know being being built and all the amazing things that are happening so that they could see firsthand all the things that I tell them.
上游可以看到下游，他们所有人都可以看到 AI 的所有进展，而且非常重要的是，他们都可以见到 AI 原住民和所有正在被建立的 AI 初创公司，以及所有正在发生的令人惊叹的事情，这样他们就可以亲眼看到我告诉他们的一切。

[06:49] And so I spend a lot of my time informing directly or indirectly um our supply chain and our partners and our ecosystem about the opportunity that's that's in front of us.
所以，我花了很多时间直接或间接向我们的供应链、我们的合作伙伴和我们的生态系统介绍我们面前的机遇。

[06:59] You know, most of my keynotes, you know, some some people always say, you know, Jensen in most keynotes, it's like one announcement after another announcement.
你知道，我的大多数主题演讲，你知道，有些人总是说，你知道，Jensen 在大多数主题演讲中，就像一个接一个的公告。

[07:08] Announcement after another announcement after another announcement after another.
一个又一个的公告，一个又一个的公告，一个又一个的公告，一个又一个。

[07:12] Announcement. Our keynotes are there's always a part.
公告。我们的主题演讲总有一部分。

[07:15] Our keynotes are there's always a part of it that's a little torturous in the sense that it's almost comes across like an ed like education and and in in fact.
我们的主题演讲总有一部分有点折磨人，因为它几乎像是教育，事实上。

[07:22] That's exactly on my mind. I need to make sure that the entire supply chain.
这正是我所想的。我需要确保整个供应链。

[07:27] Make sure that the entire supply chain upstream and downstream the ecosystem understands.
确保整个供应链上下游的生态系统都明白。

[07:32] Understands what is coming at us, why it's coming, when it's coming, how big is it going to be, and be able to reason about it systematically just like I reason about it.
明白有什么东西要来了，为什么会来，什么时候会来，会有多大，并且能够像我一样系统地思考它。

[07:41] It. And and so so I think the the the the mode as you you describe it we're able to of course um build for a future.
它。所以，我认为你所描述的模式，我们当然能够为未来构建。

[07:52] Able to of course um build for a future uh if our next next several years is a trillion dollars in in scale we have the supply chain to do it without our reach.
当然能够为未来构建，如果我们接下来的几年规模达到万亿美元，我们就有供应链可以做到，并且触手可及。

[08:02] Supply chain to do it without our reach the velocity of our business you know just as there's cash flow there's supply chain flow there turns uh nobody's going.
供应链可以做到，并且触手可及，我们的业务速度，你知道，就像有现金流一样，有供应链流在那里，没有人会。

[08:10] chain flow there turns uh nobody's going to build a supply chain for an AR architecture if the architecture the architecture the business turns is low.
那里的链式流动，没有人会为 AR 架构构建供应链，如果业务转向的架构很低的话。

[08:17] And so our ability to sustain the scale is only because our downstream demand is so great and they see it and they all hear about it.
所以我们能够维持规模，仅仅是因为我们的下游需求如此之大，他们看到了这一点，并且都听说了。

[08:24] They they see it all coming.
他们看到了这一切的到来。

[08:26] And so that's it allows us to do the things that we're able to do at the scale we're able to do.
所以，这使我们能够以我们能够做到的规模，做我们能够做的事情。

[08:32] I do want to understand more concretely whether the upstream can keep up.
我确实想更具体地了解上游是否能跟上。

[08:37] Um for many years now you guys have been 2xing revenue year-over-year.
嗯，多年来，你们的收入一直同比增长一倍。

[08:40] You guys have been more than tripling the amount of flops you're providing to the world year over year.
你们提供的浮点运算量年复一年地增长了三倍多。

[08:44] and 2xing at the scale now is really incredible.
而且现在以这个规模翻一番确实是不可思议的。

[08:47] Exactly.
没错。

[08:47] So then you look at logic say you're the biggest customer on TSMC's N3 node and um you're one of the biggest on uh AI as a whole this year is going to be 60% of N3.
所以，你看看逻辑芯片，假设你是台积电 N3 节点的最大客户，而且你是整个 AI 领域最大的客户之一，今年将占 N3 的 60%。

[08:59] It's going to be 86% next year according to some analysis.
根据一些分析，明年将占 86%。

[09:01] How how do you 2x if you're the majority?
如果你是大多数，你如何翻一番？

[09:07] Um and how do you do that year-over-year?
嗯，你如何做到年复一年地这样做？

[09:09] So are we are we in a regime now where the
那么，我们现在是否处于一个这样的时期，即

[09:11] are we are we in a regime now where the growth rate in AI compute has to slow?
我们现在是否处于一个增长率必须放缓的局面，即人工智能计算的增长率必须放缓？

[09:13] growth rate in AI compute has to slow because of upstream?
人工智能计算的增长率是否必须因为上游而放缓？

[09:15] Do you see a way to get around these you know you how do we build 2x more fabs year-over-year ultimately?
你有没有办法绕过这些，你知道我们如何才能每年建造两倍于以往的晶圆厂？

[09:21] ultimately?
最终？

[09:26] Yeah, at some at some level um the the instantaneous demand uh is greater than the supply upstream and downstream uh in the world.
是的，在某种程度上，嗯，瞬时需求大于世界范围内的上游和下游供应。

[09:37] And and it could be at any instant any instance we could be limited by the number of plumbers.
而且，在任何瞬间，任何情况下，我们都可能受到水管工数量的限制。

[09:43] plumbers.
水管工。

[09:44] Mhm.
嗯。

[09:46] Which which actually happens.
这确实发生了。

[09:47] The plumbers are invited to next year's GTC.
水管工被邀请参加明年的GTC。

[09:51] Yeah. You know, by the way, great idea.
是的。你知道，顺便说一句，这是个好主意。

[09:53] But that's a good condition. You you want you want you want a market you want an industry where the instantaneous demand is greater than the total supply of the industry.
但这是一个好条件。你想要一个市场，你想要一个行业，在这个行业里，瞬时需求大于行业的总供应量。

[10:01] Um the opposite is obviously less good.
嗯，反之显然不太好。

[10:03] If we're too far apart, uh if one particular item, one particular component is too far too far away, um obviously obviously the
如果我们相差太远，嗯，如果某一个特定的物品，某一个特定的组件离得太远，嗯，显然，显然，

[10:14] away, um obviously obviously the industry swarms it.
走开，嗯，显然，显然这个行业会蜂拥而至。

[10:17] So for example, notice people aren't talking very much about co-ass anymore.
所以，例如，注意到人们不再怎么谈论 co-ass 了。

[10:20] Yeah.
是的。

[10:22] And the reason for that is because for two years we swarmed a living daylights out of it and we double double double on on several doubles and and now I think we're in a fairly good shape.
原因是因为我们两年以来把它的生命力都榨干了，我们翻了一倍又一倍又一倍地翻倍，现在我认为我们处于一个相当不错的位置。

[10:31] And TSMC now knows that co-ass supply has to keep up with the rest of the logic demand and the memory demand and and so so they're scaling co-ass um and their scaling uh you know future packaging technologies at the same level as a scale logic which is terrific because for a long time co-ass was rather specialty and um uh HBM was rather specialty but they're not specialties anymore people now realize they're mainstream computing technology.
而台积电现在知道 co-ass 的供应必须跟上其余的逻辑需求和内存需求，所以他们正在扩展 co-ass，并且他们正在以与规模化逻辑相同的水平扩展您知道的未来封装技术，这太棒了，因为很长一段时间以来 co-ass 相当专业化，嗯，HBM 也相当专业化，但它们不再是专业化的了，人们现在意识到它们是主流计算技术。

[11:02] Um and and then and of course uh we're now much more able to influence a larger scope of our supply chain.
嗯，然后当然，我们现在更能影响我们供应链的更大范围。

[11:10] In the past in the past um you know in the beginning of the AI revolution all the things that
过去，嗯，您知道，在人工智能革命的初期，所有那些

[11:15] of the AI revolution all the things that I say now I was saying 5 years ago and I say now I was saying 5 years ago and some people believed in it and invested some people believed in it and invested in it.
在人工智能革命中，我现在说的所有事情，五年前我都在说，我现在说的，五年前我都在说，有些人相信并投资了，有些人相信并投资了。

[11:20] for example, uh, Sanjay and and in it.
例如，嗯，桑杰和，并且。

[11:23] for example, uh, Sanjay and and the Micron team.
例如，嗯，桑杰和美光团队。

[11:25] the Micron team.
美光团队。

[11:25] I still remember the meeting really well where where I I was clear about exactly what's going to happen and why it's going to happen and and the predictions the predictions that that um of today and they they really doubled down on it and we partnered with them and uh across LPDDR across you know HBM memories uh they really invested in it and and it it it obviously has been tremendous for the company.
我仍然清楚地记得那次会议，我清楚地知道将会发生什么以及为什么会发生，以及今天的预测，他们真的加倍投入了，我们与他们合作，跨越LPDDR，跨越你知道的HBM内存，他们真的投资了，这显然对公司来说是巨大的。

[11:28] meeting really well where where I I was clear about exactly what's going to happen and why it's going to happen and and the predictions the predictions that that um of today and they they really doubled down on it and we partnered with them and uh across LPDDR across you know HBM memories uh they really invested in it and and it it it obviously has been tremendous for the company.
会议非常清楚，我清楚地知道将会发生什么以及为什么会发生，以及今天的预测，他们真的加倍投入了，我们与他们合作，跨越LPDDR，跨越你知道的HBM内存，他们真的投资了，这显然对公司来说是巨大的。

[11:29] clear about exactly what's going to happen and why it's going to happen and and the predictions the predictions that that um of today and they they really doubled down on it and we partnered with them and uh across LPDDR across you know HBM memories uh they really invested in it and and it it it obviously has been tremendous for the company.
清楚地知道将会发生什么以及为什么会发生，以及今天的预测，他们真的加倍投入了，我们与他们合作，跨越LPDDR，跨越你知道的HBM内存，他们真的投资了，这显然对公司来说是巨大的。

[11:31] happen and why it's going to happen and and the predictions the predictions that that um of today and they they really doubled down on it and we partnered with them and uh across LPDDR across you know HBM memories uh they really invested in it and and it it it obviously has been tremendous for the company.
发生以及为什么会发生，以及今天的预测，他们真的加倍投入了，我们与他们合作，跨越LPDDR，跨越你知道的HBM内存，他们真的投资了，这显然对公司来说是巨大的。

[11:33] and the predictions the predictions that that um of today and they they really doubled down on it and we partnered with them and uh across LPDDR across you know HBM memories uh they really invested in it and and it it it obviously has been tremendous for the company.
以及今天的预测，他们真的加倍投入了，我们与他们合作，跨越LPDDR，跨越你知道的HBM内存，他们真的投资了，这显然对公司来说是巨大的。

[11:37] that um of today and they they really doubled down on it and we partnered with them and uh across LPDDR across you know HBM memories uh they really invested in it and and it it it obviously has been tremendous for the company.
今天的预测，他们真的加倍投入了，我们与他们合作，跨越LPDDR，跨越你知道的HBM内存，他们真的投资了，这显然对公司来说是巨大的。

[11:39] doubled down on it and we partnered with them and uh across LPDDR across you know HBM memories uh they really invested in it and and it it it obviously has been tremendous for the company.
加倍投入了，我们与他们合作，跨越LPDDR，跨越你知道的HBM内存，他们真的投资了，这显然对公司来说是巨大的。

[11:42] them and uh across LPDDR across you know HBM memories uh they really invested in it and and it it it obviously has been tremendous for the company.
他们，以及跨越LPDDR，跨越你知道的HBM内存，他们真的投资了，这显然对公司来说是巨大的。

[11:44] HBM memories uh they really invested in it and and it it it obviously has been tremendous for the company.
HBM内存，他们真的投资了，这显然对公司来说是巨大的。

[11:47] it and and it it it obviously has been tremendous for the company.
它，这显然对公司来说是巨大的。

[11:47] tremendous for the company.
对公司来说是巨大的。

[11:50] uh some some people came a little bit later and uh but they now they're all here and so I I think the each one of these generation each one of these bottlenecks gets a great deal of attention um and now we're we're prefetching the bottlenecks uh years in advance.
嗯，有些人来得晚一些，但他们现在都在这里了，所以我认为这些瓶颈中的每一个都会得到极大的关注，嗯，现在我们提前几年就开始预取瓶颈了。

[11:52] people came a little bit later and uh but they now they're all here and so I I think the each one of these generation each one of these bottlenecks gets a great deal of attention um and now we're we're prefetching the bottlenecks uh years in advance.
有些人来得晚一些，但他们现在都在这里了，所以我认为这些瓶颈中的每一个都会得到极大的关注，嗯，现在我们提前几年就开始预取瓶颈了。

[11:54] but they now they're all here and so I I think the each one of these generation each one of these bottlenecks gets a great deal of attention um and now we're we're prefetching the bottlenecks uh years in advance.
但他们现在都在这里了，所以我认为这些瓶颈中的每一个都会得到极大的关注，嗯，现在我们提前几年就开始预取瓶颈了。

[11:57] think the each one of these generation each one of these bottlenecks gets a great deal of attention um and now we're we're prefetching the bottlenecks uh years in advance.
我认为这些瓶颈中的每一个都会得到极大的关注，嗯，现在我们提前几年就开始预取瓶颈了。

[11:59] each one of these bottlenecks gets a great deal of attention um and now we're we're prefetching the bottlenecks uh years in advance.
这些瓶颈中的每一个都会得到极大的关注，嗯，现在我们提前几年就开始预取瓶颈了。

[12:02] gets a great deal of attention um and now we're we're prefetching the bottlenecks uh years in advance.
得到极大的关注，嗯，现在我们提前几年就开始预取瓶颈了。

[12:04] now we're we're prefetching the bottlenecks uh years in advance.
现在我们提前几年就开始预取瓶颈了。

[12:06] bottlenecks uh years in advance.
瓶颈，提前几年。

[12:06] So for example uh the the the investments that we've done uh with uh with Lum and Coherent and um all of the silicon photonix ecosystem uh the last several
所以，例如，我们与Lum和Coherent所做的投资，以及整个硅光子生态系统，在过去几个月里

[12:09] example uh the the the investments that we've done uh with uh with Lum and Coherent and um all of the silicon photonix ecosystem uh the last several
例如，我们与Lum和Coherent所做的投资，以及整个硅光子生态系统，在过去几个月里

[12:12] we've done uh with uh with Lum and Coherent and um all of the silicon photonix ecosystem uh the last several
我们与Lum和Coherent所做的投资，以及整个硅光子生态系统，在过去几个月里

[12:15] Coherent and um all of the silicon photonix ecosystem uh the last several
Coherent以及整个硅光子生态系统，在过去几个月里

[12:18] Photonix ecosystem, uh, the last several years, we really reshaped the ecosystem.
光子生态系统，嗯，在过去的几年里，我们真的重塑了生态系统。

[12:20] Years, we really reshaped the ecosystem and the supply chain, silicon photonix.
几年里，我们真的重塑了生态系统和供应链，硅光子学。

[12:23] And the supply chain, silicon photonix.
以及供应链，硅光子学。

[12:25] We, we, uh, built up an entire supply chain around TSMC.
我们，我们，呃，围绕台积电建立了一个完整的供应链。

[12:28] We partnered with them on coupe, uh, invented a whole bunch of technology.
我们与他们合作，嗯，发明了一大堆技术。

[12:30] Technology, we licensed, uh, those patents to the supply chain.
技术，我们授权了，呃，那些专利给供应链。

[12:33] To the supply chain, keep it nice and open.
给供应链，保持它良好和开放。

[12:35] Um, and so we're preparing the supply chain through invention of new technologies, new workflows, uh, new test, new testing equipment.
嗯，所以我们正在通过发明新技术、新工作流程、呃，新测试、新测试设备来准备供应链。

[12:37] New testing equipment, double-sided probing, um, investing in companies, helping them scale up their capacity.
新的测试设备，双面探测，嗯，投资公司，帮助他们扩大产能。

[12:39] Companies, helping them scale up their capacity.
公司，帮助他们扩大产能。

[12:42] Um, and so, so you could see that we're trying to shape the ecosystem so that it's ready, the supply chain so that it's ready to support the scale.
嗯，所以，所以你可以看到我们正在努力塑造生态系统，使其准备就绪，供应链也准备就绪以支持规模化。

[12:43] It seems like some bottlenecks are easier than others.
似乎有些瓶颈比其他瓶颈更容易解决。

[12:46] And so scaling up co-ass versus scaling up.
所以扩大联合评估与扩大规模相比。

[12:48] I went to the hardest one by the way.
顺便说一句，我去了最难的一个。

[12:51] Which is plumbers.
那就是水管工。

[12:53] Yeah, it's true.
是的，是真的。

[12:55] Yeah.
是的。

[12:57] Yeah, I actually went to the hardest one.
是的，我实际上去了最难的一个。

[12:58] Yeah.
是的。

[13:01] Yeah, plumbers and electricians.
是的，水管工和电工。

[13:02] And the reason for that is because.
原因是因为。

[13:04] Because and this is one of the concerns that I have about of all the doom the doomers, describing the end of end of.
因为这是我担心的其中一点，关于所有预言末日的人，描述末日的结束。

[13:20] Doomers, um, describing the end of, end of work and killing of jobs.
悲观主义者，嗯，描述了工作的终结，工作的消失。

[13:23] And you know, one of the things that that that um if we discourage people from being software engineers, we're going to run out of software engineers.
你知道，如果我们阻止人们成为软件工程师，我们将面临软件工程师短缺。

[13:30] And and uh the same prediction 10 years ago, some of the doomers were were uh uh saying that we're telling people whatever you do, don't be a radiologist.
而且，嗯，十年前也有同样的预测，一些悲观主义者说，我们告诉人们，无论你做什么，都不要成为放射科医生。

[13:40] And you might hear some of those some of those videos are still on the web.
你可能还会看到一些视频仍然在网上。

[13:44] You know, radiology is is going to be the first career to go.
你知道，放射学将是第一个消失的职业。

[13:48] Nobody's the world's not going to need any more radiologists.
没有人会，世界将不再需要放射科医生。

[13:50] Guess what?
猜猜怎么着？

[13:51] But we're short of radiologists.
但我们却缺少放射科医生。

[13:54] Oh, but okay.
哦，但是，好的。

[13:55] So, going back to this point about well some things you scale other things like how do you actually get how do you actually manufacture 2x the amount of logic a year?
所以，回到这个观点，有些东西你可以扩展，其他东西呢，比如你如何实际制造出每年两倍的逻辑量？

[14:02] Ultimately that's bottleneck by memory and logic are bottleneck by UV.
最终，这受内存和逻辑的瓶颈，或者受紫外线的瓶颈。

[14:05] How do you get to 2x as many UV machines a year?
你如何才能一年内获得两倍的紫外线机器数量？

[14:09] Yeah.
是的。

[14:10] Year over year.
年复一年。

[14:10] None of that none of that's impossible to scale quickly.
这一切都不是不可能快速扩展的。

[14:13] You just need to you you could do all of that is easy to do within two or three years.
你只需要，你可以在两到三年内完成所有这些事情，这很容易做到。

[14:17] You just need a demand signal that it's
你只需要一个需求信号，表明它

[14:21] You just need a demand signal that it's not it once you once you can build one.
你只需要一个需求信号，一旦你能制造一个，就不是它了。

[14:24] Not it once you once you can build one, you can build 10 and once you can build.
一旦你能制造一个，就不是它了，你可以制造10个，一旦你能制造。

[14:25] You can build 10 and once you can build build 10, you can build a million and so.
你可以制造10个，一旦你制造10个，你就可以制造100万个，所以。

[14:28] Build 10, you can build a million and so these things are not not hard to.
制造10个，你可以制造100万个，所以这些东西并不难。

[14:29] These things are not not hard to replicate.
这些东西并不难复制。

[14:31] How far down the supply chain do you go where you do you go to ASML?
在供应链的哪个环节，你会去ASML？

[14:34] Do you go where you do you go to ASML and say hey, if I look out three years.
你会去ASML说，嘿，如果我展望未来三年。

[14:35] And say hey, if I look out three years from now for me to for Nvidia to be generating two trillion in a year in.
说嘿，如果我展望未来三年，为了让我，为了英伟达一年能产生两万亿的收入。

[14:38] Generating two trillion in a year in revenue, we need way more AUV machines.
一年产生两万亿的收入，我们需要更多的AUV机器。

[14:40] Revenue, we need way more AUV machines and.
收入，我们需要更多的AUV机器，而且。

[14:41] And >> some of them I have to directly uh some.
而且>>其中一些我必须直接，呃，一些。

[14:44] >> some of them I have to directly uh some of them are indirectly and some of them.
>>其中一些我必须直接，呃，其中一些是间接的，还有一些。

[14:46] Of them are indirectly and some of them um if I can convince TSMC as ASML will.
其中一些是间接的，还有一些，嗯，如果我能说服台积电，就像ASML一样。

[14:49] Um if I can convince TSMC as ASML will be convinced and so that's that you know.
嗯，如果我能说服台积电，ASML就会被说服，所以你知道。

[14:51] Be convinced and so that's that you know we have to think about the critical.
被说服，所以你知道我们必须考虑关键的。

[14:53] we have to think about the critical critical pinch points and uh but if TSMC.
我们必须考虑关键的关键瓶颈，但是如果台积电。

[14:56] Is convinced uh you'll have plenty of EV machines in a few years.
被说服，嗯，几年后你将拥有足够的EV机器。

[15:00] Machines in a few years. And so none of.
几年后会有机器。所以没有。

[15:03] That my point is that none of the bottlenecks last longer than a couple 2.
我的观点是，没有一个瓶颈会持续超过两三年。

[15:05] Bottlenecks last longer than a couple 2 three years. None of them. And meanwhile.
瓶颈持续时间超过两三年。没有一个。同时。

[15:07] Three years. None of them. And meanwhile meanwhile we're uh improving computing.
三年。没有一个。同时，我们正在，呃，提高计算能力。

[15:11] Meanwhile we're uh improving computing efficiency by 10x 20x in the case of.
同时，我们正在，呃，将计算效率提高10倍，20倍，在...

[15:13] Efficiency by 10x 20x in the case of Hopper to Blackwell some 30 50x um we're.
效率方面，从Hopper到Blackwell提高10倍，20倍，甚至30到50倍，嗯，我们正在。

[15:16] Hopper to Blackwell some 30 50x um we're coming up with new algorithms because.
从Hopper到Blackwell提高30到50倍，嗯，我们正在开发新的算法，因为。

[15:20] Coming up with new algorithms because.
正在开发新的算法，因为。

[15:22] coming up with new algorithms because CUDA is so flexible.
提出新的算法，因为 CUDA 非常灵活。

[15:25] CUDA is so flexible.
CUDA 非常灵活。

[15:27] Uh we're we're developing all kinds of new techniques so that we drive efficiency.
呃，我们正在开发各种新技术，以提高效率。

[15:29] uh in addition to increasing capacity.
呃，除了增加容量之外。

[15:31] Yeah.
是的。

[15:33] And so so there those those are those are things that that none of that worry me.
所以，那些事情都不能让我担心。

[15:36] It's the stuff that's downstream from us.
这是在我们下游的东西。

[15:38] Um energy policies that prevent energy from from you know you can't grow you can't create you can't create an industry without energy.
嗯，能源政策阻碍了能源，你知道，没有能源你就无法发展，无法创造，无法建立一个产业。

[15:44] You can't create a whole new manufacturing industry without energy.
没有能源，你就无法创造一个全新的制造业。

[15:47] Uh we want to re-industrialize the United States.
呃，我们想让美国重新工业化。

[15:49] We want to bring back uh chip manufacturing and computer manufacturing and packaging and we want to build new things like EVs and robots and we want to build AI factories and you you can't build any of these things without energy and those things take a long time but more chip capacity that's a two threeear problem.
我们想恢复呃芯片制造、计算机制造和封装，我们想建造电动汽车、机器人等新事物，我们想建造人工智能工厂，没有能源你就无法建造任何这些东西，这些事情需要很长时间，但更多的芯片产能是一个两三年的问题。

[16:11] more coass capacity 2 three year problem.
更多的粗略产能是两三年的问题。

[16:13] interesting I I feel like I have guests tell me the exact opposite thing sometimes and I don't in this case I just don't have the technical knowledge to adjudicate but.
有趣的是，我感觉有时客人会告诉我完全相反的事情，在这种情况下，我只是没有技术知识来评判，但是。

[16:20] well the beautiful thing is you're talking to the expert.
嗯，美妙之处在于你正在与专家交谈。

[16:21] Yeah, true, true.
是的，没错，没错。

[16:25] Talking to the expert, yeah, true, true.
与专家交谈，是的，没错，没错。

[16:27] Um, okay. I want to ask about um your competitors.
嗯，好的。我想问一下关于你们的竞争对手。

[16:28] Competitors.
竞争对手。

[16:28] Yeah.
是的。

[16:28] So, if you look at TPU, arguably two out of the top three models in the world, Claude and Gemini, were trained on TPU.
所以，如果你看TPU，可以说世界上排名前三的模型中有两个，Claude和Gemini，都是在TPU上训练的。

[16:39] What does that mean for Nvidia going forward?
这对英伟达未来的发展意味着什么？

[16:41] Forward?
未来？

[16:43] Um, well, we have we have a very different, we built a very different thing.
嗯，嗯，我们有一个非常不同的，我们构建了一个非常不同的东西。

[16:44] Um, you know, what what Nvidia built is accelerated computing.
嗯，你知道，英伟达构建的是加速计算。

[16:51] Not a tensor processing unit.
不是张量处理单元。

[16:55] And uh accelerated computing is used for all kinds of things.
呃，加速计算被用于各种各样的事情。

[16:57] You know, molecular dynamics and quantum chromodynamics and it's used for data processing, data frames, structured data, unstructured data.
你知道，分子动力学和量子色动力学，它被用于数据处理、数据帧、结构化数据、非结构化数据。

[17:07] It's used for um fluid dynamics, particle physics, you know, and in addition, we use it for AI.
它被用于流体动力学、粒子物理学，你知道，此外，我们还将其用于人工智能。

[17:17] And so accelerated computing is is um much more diverse and and although AI is the conversation today is obviously very
所以加速计算更加多样化，虽然人工智能是今天的谈话焦点，但显然非常

[17:25] The conversation today is obviously very important and impactful. Uh computing is important and impactful.
今天的对话显然非常重要且有影响力。嗯，计算很重要且有影响力。

[17:29] Uh computing is much broader than that and what Nvidia has done is reinventing reinvented the way computing is done from general purpose computing to accelerate computing.
嗯，计算的范围要广得多，而英伟达所做的就是重新发明了计算的执行方式，从通用计算到加速计算。

[17:34] Our market reach is far greater than any any TPU can any ASA can possibly have.
我们的市场覆盖范围远远大于任何 TPU 或任何 ASA 可能拥有的。

[17:46] And so if you look at our position, uh we're the only company that that accelerates applications of all kinds.
所以，如果你看看我们的地位，嗯，我们是唯一一家加速各种类型应用程序的公司。

[17:54] We have a gigantic ecosystem and so all kinds of frameworks and algorithms all run on Nvidia.
我们拥有一个庞大的生态系统，所以各种框架和算法都在 Nvidia 上运行。

[18:04] And because our computers are designed to be operated by other people, anyone who's an operator could buy our systems.
而且因为我们的计算机是为他人操作而设计的，任何操作员都可以购买我们的系统。

[18:13] Most of these homebuilt systems you have to be your own operator because it was never designed to be flexible enough for other people to operate.
大多数这些自建系统，你必须自己操作，因为它从未被设计成足够灵活，供其他人操作。

[18:20] And so as a result of the fact that anybody can operate our systems, we're in every
因此，由于任何人都可以操作我们的系统，我们无处不在

[18:26] operate our systems, we're in every cloud including Google and Amazon and cloud including Google and Amazon and you know Azure and OCI and right and so whether you want to operate it to rent or operate it if you want to operate to rent you better have large ecosystem of customers in many industries that be the offtakers.
运营我们的系统，我们支持所有云，包括谷歌和亚马逊，以及谷歌和亚马逊云，还有你知道的Azure和OCI，对吧，所以无论你想租用运营还是自己运营，如果你想租用运营，你最好拥有一个庞大的生态系统，在许多行业拥有客户，他们将是购买者。

[18:42] if you're operating it if you if you want to operate it for yourself um we you know we obviously have the ability to help you operate yourself like for example for Elon with XAI and uh because we could we could enable operators uh in any any company in any industry you could use it uh to build a supercomput for uh scientific research and drug discovery at Lily and so we can help them operate their own supercomputer and and use it for the entire diversity of drug discovery and biological sciences um that that we accelerate >>
如果你自己运营，如果你想自己运营，嗯，我们，你知道的，我们显然有能力帮助你自己运营，比如埃隆的XAI，嗯，因为我们可以，我们可以赋能任何行业中的任何公司的运营商，你可以用它来为科学研究和礼来公司的药物发现构建超级计算机，所以我们可以帮助他们运营自己的超级计算机，并将其用于药物发现和生命科学的整个多样性，嗯，我们加速了这一点>>

[19:20] and so so there there just you know a whole bunch of applications that we can address that you can't do so with TPUs
所以，所以有很多应用程序我们可以解决，而你无法用TPU做到这一点。

[19:28] address that you can't do so with TPUs because Nvidia's built CUDA as a fantastic tensor processing unit as well.
地址是您无法使用TPU完成的，因为Nvidia构建的CUDA也是一个出色的张量处理单元。

[19:34] but it does you know it does every every life cycle of data processing and computing and AI and so on so forth.
但它确实，你知道，它处理数据处理、计算和人工智能等所有生命周期。

[19:41] and so I our our market opportunity is just a lot larger.
所以我们的市场机会要大得多。

[19:43] Our reach is a lot greater and because we have such a large um we basically support every application in the world.
我们的影响力更大，因为我们拥有如此庞大的用户群，我们基本上支持世界上所有的应用程序。

[19:53] now you could build Nvidia systems anywhere and know that there will be customers for it.
现在，您可以在任何地方构建Nvidia系统，并且知道会有客户。

[19:58] and so it's a very different thing.
所以这是一个非常不同的事情。

[20:00] Uh this is going to be sort of a long question but you know you have spectacular revenue um and this revenue is mostly you're not making 60 billion a quarter from uh pharma and um quantum.
呃，这会是一个有点长的问题，但你知道你们有惊人的收入，而这笔收入主要是，你们不是从制药和量子领域每季度赚取600亿美元。

[20:10] you're making it because AI is unprecedented technology that is growing unprecedentedly fast.
你们之所以能做到，是因为人工智能是一项前所未有的技术，而且正在以前所未有的速度增长。

[20:14] and so then the question is what is best for AI specifically and I'm not in the details.
所以问题是，什么最适合人工智能，我并不了解具体细节。

[20:18] but I talked to my AI researcher friends and they say look when I use a TPU it's this big systolic array that's perfect for doing matrix multiplies whereas a GPU is very flexible.
但我与我的AI研究员朋友们交谈过，他们说，看，当我使用TPU时，它是一个巨大的并行阵列，非常适合进行矩阵乘法，而GPU则非常灵活。

[20:25] It's great when you
当你...

[20:30] GPU is very flexible It's great when you have lots of branching when you have um

[20:33] have lots of branching when you have um irregular memory access but with these

[20:36] irregular memory access but with these you know what what is AI just like these

[20:37] you know what what is AI just like these very predictable matrix multiplies again

[20:39] very predictable matrix multiplies again and again and again and you don't have

[20:41] and again and again and you don't have to give up any die area for warp

[20:43] to give up any die area for warp schedulers for you know switches between

[20:45] schedulers for you know switches between threads and memory banks and so the TPU

[20:48] threads and memory banks and so the TPU is really optimized for the majority the

[20:50] is really optimized for the majority the bulk of this growth in revenue and use

[20:52] bulk of this growth in revenue and use case for uh compute that is coming

[20:54] case for uh compute that is coming online right now um yeah I I wonder how

[20:57] online right now um yeah I I wonder how you react to

[20:59] you react to um

[21:01] um matrix multiplies is an important part

[21:03] matrix multiplies is an important part of AI but it's not the only part of AI

[21:06] of AI but it's not the only part of AI and if you want to come up with a new

[21:08] and if you want to come up with a new attention mechanism or if you want to

[21:10] attention mechanism or if you want to disagregate in a different way if you

[21:13] disagregate in a different way if you want to come up with a whole new type of

[21:17] want to come up with a whole new type of architecture altogether for example you

[21:20] architecture altogether for example you know a hybrid SSM uh if you want to use

[21:23] know a hybrid SSM uh if you want to use a you want to create a model that that

[21:26] a you want to create a model that that um that fuses diffusion and auto

[21:30] um that fuses diffusion and auto reggressive somehow. Uh you you want an

[21:33] reggressive somehow. Uh you you want an architecture that's just generally

[21:35] architecture that's just generally programmable

[21:36] programmable and and we run everything you can

[21:40] and and we run everything you can imagine. And so that's the advantage. It

[21:42] imagine. And so that's the advantage. It allows for invention of new algorithms a

[21:45] allows for invention of new algorithms a lot more a lot a lot more easily.

[21:48] lot more a lot a lot more easily. >> And so because it's a programmable

[21:50] >> And so because it's a programmable system and and the ability to invent new

[21:53] system and and the ability to invent new algorithms is really what makes AI

[21:56] algorithms is really what makes AI advance. So quickly, you know,

[22:00] advance. So quickly, you know, TPUs like anything else is impacted by

[22:03] TPUs like anything else is impacted by Moore's law. And we know that Moore's

[22:05] Moore's law. And we know that Moore's law is increasing about 25% per year.

[22:08] law is increasing about 25% per year. And so the only way to really get 10x

[22:12] And so the only way to really get 10x leaps, 100x leaps,

[22:15] leaps, 100x leaps, is to fundamentally change the algorithm

[22:19] is to fundamentally change the algorithm and how it's computed every single year.

[22:22] and how it's computed every single year. >> And that's Nvidia's fundamental

[22:23] >> And that's Nvidia's fundamental advantage. The only reason why we were

[22:27] advantage. The only reason why we were able to make black well to hopper 50

[22:29] able to make black well to hopper 50 times, you know, I said it was 35 times

[22:32] times, you know, I said it was 35 times and and and when I first announced it

[22:34] and and and when I first announced it was going to black wall is going to be

[22:35] was going to black wall is going to be 35 times more energy efficient than

[22:38] 35 times more energy efficient than hopper. Uh nobody believed it and and uh

[22:42] hopper. Uh nobody believed it and and uh and then and then Dylan wrote an

[22:43] and then and then Dylan wrote an article. He said he said in fact in fact

[22:45] article. He said he said in fact in fact I sandbagged it's actually 50 times. And

[22:49] I sandbagged it's actually 50 times. And you can't reasonably do that with just

[22:50] you can't reasonably do that with just Moore's law. And so the the way that we

[22:54] Moore's law. And so the the way that we solve that problem is new out new models

[22:59] solve that problem is new out new models um uh parallelized and disagregated and

[23:02] um uh parallelized and disagregated and and distributed uh uh across a computing

[23:06] and distributed uh uh across a computing system uh and without the ability to

[23:10] system uh and without the ability to really get down and come up with new

[23:12] really get down and come up with new kernels with CUDA, it's really hard to

[23:14] kernels with CUDA, it's really hard to do and and so the combination of the

[23:18] do and and so the combination of the programmability of our of our

[23:20] programmability of our of our architecture

[23:21] architecture uh the the fact that Nvidia is an

[23:24] uh the the fact that Nvidia is an extreme codeesign company where we could

[23:27] extreme codeesign company where we could even offload some of the computation

[23:29] even offload some of the computation into the fabric itself, MVLink for

[23:31] into the fabric itself, MVLink for example into the network spectrum X um

[23:35] example into the network spectrum X um uh and that we could affect change

[23:38] uh and that we could affect change across the processors, the system, the

[23:43] across the processors, the system, the fabric, the libraries, the algorithm.

[23:47] fabric, the libraries, the algorithm. All of that was done simultaneously.

[23:49] All of that was done simultaneously. Without CUDA to do that, I wouldn't even

[23:51] Without CUDA to do that, I wouldn't even know where to start.

[23:53] know where to start. >> My sponsor Cruso was among the first

[23:54] >> My sponsor Cruso was among the first clouds to offer Nvidia's Blackwell and

[23:56] clouds to offer Nvidia's Blackwell and Blackwell Ultra platforms, and they just

[23:58] Blackwell Ultra platforms, and they just announced their Nvidia Vera Rubin

[24:00] announced their Nvidia Vera Rubin deployment scheduled for later this

[24:01] deployment scheduled for later this year. But access to state-of-the-art

[24:03] year. But access to state-of-the-art hardware is only part of the story. For

[24:05] hardware is only part of the story. For example, most inference engines already

[24:07] example, most inference engines already do KV caching for a single user's

[24:08] do KV caching for a single user's forward passes, but Cruso does it across

[24:11] forward passes, but Cruso does it across users and GPUs. So if a thousand agents

[24:13] users and GPUs. So if a thousand agents are running on the same system prompt,

[24:14] are running on the same system prompt, Cruso only has to compute the KV cache

[24:16] Cruso only has to compute the KV cache once for it to become available to every

[24:18] once for it to become available to every single GPU in the cluster. This is

[24:20] single GPU in the cluster. This is especially important as systems get more

[24:21] especially important as systems get more agendic and require much longer prefixes

[24:24] agendic and require much longer prefixes in order to use tools and access files.

[24:27] in order to use tools and access files. In a recent benchmark, Crusoe was able

[24:28] In a recent benchmark, Crusoe was able to deliver up to 10 times faster time to

[24:32] to deliver up to 10 times faster time to first token and up to five times better

[24:33] first token and up to five times better throughput than VLM. This is just one

[24:36] throughput than VLM. This is just one among many reasons that you should run

[24:37] among many reasons that you should run your inference workload with Cruso. And

[24:39] your inference workload with Cruso. And if you need GPUs for training, you don't

[24:41] if you need GPUs for training, you don't need to switch clouds. Cruso's got you

[24:42] need to switch clouds. Cruso's got you covered there, too. Go to

[24:43] covered there, too. Go to cruso.ai/torcashe

[24:45] cruso.ai/torcashe to learn more.

[24:47] to learn more. >> So, this gets at a interesting question

[24:49] >> So, this gets at a interesting question about um Nvidia's

[24:52] about um Nvidia's clientele where if 60% of your revenue

[24:55] clientele where if 60% of your revenue is coming from these big five

[24:58] is coming from these big five hyperscalers, you know, in a in in a

[25:01] hyperscalers, you know, in a in in a different era where different customers,

[25:02] different era where different customers, let's say it's professors who are

[25:03] let's say it's professors who are running experiments and they are helped

[25:05] running experiments and they are helped a bunch by they need CUDA. um they can't

[25:08] a bunch by they need CUDA. um they can't use another accelerator. They need to

[25:10] use another accelerator. They need to just run PyTorch with CUDA and have

[25:12] just run PyTorch with CUDA and have everything optimized. But if you got

[25:14] everything optimized. But if you got these hyperscalers, they have the

[25:15] these hyperscalers, they have the resources to write their own kernels. In

[25:17] resources to write their own kernels. In fact, they have to to get that extra

[25:18] fact, they have to to get that extra last 5% that they need for their

[25:21] last 5% that they need for their specific architecture. Um Anthropic,

[25:24] specific architecture. Um Anthropic, Google are mostly running their own

[25:26] Google are mostly running their own accelerators or running TPUs um and

[25:29] accelerators or running TPUs um and Tranium, but even OpenAI using GPUs has

[25:32] Tranium, but even OpenAI using GPUs has um has Triton which they're like we need

[25:35] um has Triton which they're like we need our own kernels. So they've um down to

[25:38] our own kernels. So they've um down to CUDA C++ they've instead of using Kublas

[25:41] CUDA C++ they've instead of using Kublas and Nickel and everything they've got

[25:43] and Nickel and everything they've got their own stack which compiles to other

[25:45] their own stack which compiles to other accelerators as well. Um and so if most

[25:47] accelerators as well. Um and so if most of your customers can can and do make

[25:51] of your customers can can and do make replacements for CUDA to what extent is

[25:53] replacements for CUDA to what extent is CUDA really the thing that is going to

[25:55] CUDA really the thing that is going to make Frontier AI happen on Nvidia? CUDA.

[25:59] make Frontier AI happen on Nvidia? CUDA. CUDA is um is a a rich ecosystem and so

[26:04] CUDA is um is a a rich ecosystem and so if you want to build on any computer

[26:06] if you want to build on any computer first, building on CUDA first is

[26:09] first, building on CUDA first is incredibly smart and because the

[26:12] incredibly smart and because the ecosystem is so rich uh we support every

[26:15] ecosystem is so rich uh we support every framework. uh if you want to create

[26:17] framework. uh if you want to create custom kernels uh if you need for

[26:20] custom kernels uh if you need for example we contribute enormously to

[26:22] example we contribute enormously to Triton and so the back end of Triton um

[26:25] Triton and so the back end of Triton um huge amounts of NVIDIA technology

[26:28] huge amounts of NVIDIA technology we're delighted to help every framework

[26:30] we're delighted to help every framework uh become as great as it can be and

[26:33] uh become as great as it can be and there's lots and lots of frameworks

[26:34] there's lots and lots of frameworks there's Triton there's VLM there's SG

[26:36] there's Triton there's VLM there's SG lang and then there's more right and now

[26:38] lang and then there's more right and now there's there's a whole bunch of new

[26:40] there's there's a whole bunch of new reinforcement learning frameworks coming

[26:42] reinforcement learning frameworks coming out you know you got Verl you got Nemo

[26:44] out you know you got Verl you got Nemo RL you got a whole bunch of new and then

[26:45] RL you got a whole bunch of new and then the the now with with with post-

[26:48] the the now with with with post- trainining and reinforcement learning

[26:50] trainining and reinforcement learning that entire area is just exploding right

[26:53] that entire area is just exploding right and so if you want to build on on an

[26:55] and so if you want to build on on an architecture building on a CUDA makes

[26:57] architecture building on a CUDA makes the most sense because you know that the

[26:58] the most sense because you know that the ecosystem is great you know that if

[27:01] ecosystem is great you know that if something happens it's more likely in

[27:03] something happens it's more likely in your code and not in the mountain of

[27:05] your code and not in the mountain of code underneath you know don't forget

[27:07] code underneath you know don't forget the amount of code that you're dealing

[27:08] the amount of code that you're dealing with when you're building these systems

[27:11] with when you're building these systems when something doesn't work was it you

[27:14] when something doesn't work was it you or was it the computer, you would like

[27:16] or was it the computer, you would like it always to be you and to to be able to

[27:19] it always to be you and to to be able to trust the computer and and you know,

[27:21] trust the computer and and you know, obviously we still have lots and lots of

[27:22] obviously we still have lots and lots of lots and lots of bugs ourselves, but but

[27:25] lots and lots of bugs ourselves, but but our system is so well rung out that you

[27:29] our system is so well rung out that you could at least build on top of the

[27:30] could at least build on top of the foundation. So that's number one is that

[27:32] foundation. So that's number one is that the richness of the ecosystem, the

[27:34] the richness of the ecosystem, the programmability of it, the capability of

[27:35] programmability of it, the capability of it. The second thing is is um if you

[27:38] it. The second thing is is um if you were a developer and you were building

[27:40] were a developer and you were building anything at all, the single most

[27:42] anything at all, the single most important thing you want more than

[27:43] important thing you want more than anything is install base. You want the

[27:45] anything is install base. You want the software that you run to run on a whole

[27:47] software that you run to run on a whole bunch of other computers. You don't want

[27:49] bunch of other computers. You don't want to build a software. You're not building

[27:50] to build a software. You're not building software just for yourself. You're

[27:52] software just for yourself. You're building software for your fleet or for

[27:54] building software for your fleet or for everybody else's fleet because you're a

[27:55] everybody else's fleet because you're a framework builder. And Nvidia's CUDA

[27:58] framework builder. And Nvidia's CUDA ecosystem is ultimately its great

[28:01] ecosystem is ultimately its great treasure. We are now I don't know how

[28:04] treasure. We are now I don't know how many several hundred million GPUs. Every

[28:07] many several hundred million GPUs. Every cloud has it goes back to A10, A100,

[28:11] cloud has it goes back to A10, A100, H100, H200,

[28:14] H100, H200, you know, the L series, the P series. I

[28:18] you know, the L series, the P series. I mean, there's a whole bunch of them and

[28:21] mean, there's a whole bunch of them and and they're they're they're in all kinds

[28:23] and they're they're they're in all kinds of sizes and shapes. And if you're a

[28:24] of sizes and shapes. And if you're a robotics company, you want that CUDA

[28:26] robotics company, you want that CUDA stack to actually run in the CUDA in the

[28:28] stack to actually run in the CUDA in the robot itself. We're literally

[28:29] robot itself. We're literally everywhere. And so the install base says

[28:32] everywhere. And so the install base says that once you develop the software, once

[28:34] that once you develop the software, once you develop the model, it's going to be

[28:36] you develop the model, it's going to be useful everywhere. And so the install

[28:38] useful everywhere. And so the install base is just too incredibly valuable.

[28:41] base is just too incredibly valuable. And then lastly, the fact that we're in

[28:43] And then lastly, the fact that we're in every single cloud makes us genuinely

[28:46] every single cloud makes us genuinely unique because you're an AI company and

[28:49] unique because you're an AI company and you're an AI developer. You're not

[28:51] you're an AI developer. You're not exactly sure which CSP you're going to

[28:53] exactly sure which CSP you're going to partner with and where you would like to

[28:54] partner with and where you would like to run it. And we'd run it everywhere,

[28:56] run it. And we'd run it everywhere, including on prem for you if you like.

[28:58] including on prem for you if you like. And so so I think that that the the

[29:02] And so so I think that that the the richness of the ecosystem, the

[29:05] richness of the ecosystem, the expansiveness of the of the of the

[29:08] expansiveness of the of the of the install base and the versatility of

[29:11] install base and the versatility of where where where we are, that

[29:13] where where where we are, that combination is is uh makes CUDA

[29:15] combination is is uh makes CUDA invaluable.

[29:16] invaluable. >> That makes a lot of sense. I guess I I

[29:17] >> That makes a lot of sense. I guess I I guess the thing I'm curious about is um

[29:20] guess the thing I'm curious about is um whether those advantages matter a lot to

[29:24] whether those advantages matter a lot to your main customers. um like there

[29:28] your main customers. um like there there's many people who who they might

[29:29] there's many people who who they might matter for for the kind of person who

[29:30] matter for for the kind of person who can actually build their own software

[29:32] can actually build their own software stack who are make up most of your

[29:33] stack who are make up most of your revenue um especially if you go to a

[29:35] revenue um especially if you go to a world where AI is getting especially

[29:37] world where AI is getting especially good at the things which have tight

[29:39] good at the things which have tight verification loops where you can RL on

[29:41] verification loops where you can RL on them and then this question of how do

[29:43] them and then this question of how do you write a kernel that does attention

[29:46] you write a kernel that does attention or MLP the most efficiently across a

[29:48] or MLP the most efficiently across a scale up it's a very verifiable sort of

[29:51] scale up it's a very verifiable sort of feedback loop and so oh can everybody

[29:54] feedback loop and so oh can everybody can all the hyperscalers write these

[29:55] can all the hyperscalers write these custom kernel for themselves. Um, and

[29:58] custom kernel for themselves. Um, and they might still Nvidia has uh still has

[30:01] they might still Nvidia has uh still has great price performance. So, they might

[30:02] great price performance. So, they might still prefer to use Nvidia. But then the

[30:04] still prefer to use Nvidia. But then the question is does it just become a

[30:05] question is does it just become a question of who is offering the best

[30:08] question of who is offering the best specs, the best um flops and memory and

[30:11] specs, the best um flops and memory and memory bandwidth for a given dollar

[30:13] memory bandwidth for a given dollar where historically Nvidia has just had

[30:14] where historically Nvidia has just had and still has you know the best margins

[30:17] and still has you know the best margins in all of AI across hardware and

[30:18] in all of AI across hardware and software 70% plus because of this CUDA

[30:21] software 70% plus because of this CUDA mode. And the question is, oh, can you

[30:23] mode. And the question is, oh, can you sustain those margins if for most of

[30:26] sustain those margins if for most of your customers they can actually afford

[30:28] your customers they can actually afford to build

[30:31] to build build instead of the CUDA mode. The

[30:34] build instead of the CUDA mode. The number of engineers we have assigned to

[30:35] number of engineers we have assigned to these AI labs is insane.

[30:38] these AI labs is insane. working with them, optimizing their

[30:39] working with them, optimizing their stack. And the reason for that is

[30:42] stack. And the reason for that is because because um nobody knows our

[30:44] because because um nobody knows our architecture better than we do. And

[30:46] architecture better than we do. And these architectures are not not as

[30:49] these architectures are not not as general purpose as a CPU. The reason the

[30:52] general purpose as a CPU. The reason the reason why a CPU is so, you know, a CPU

[30:54] reason why a CPU is so, you know, a CPU is kind of like like a Cadillac, you

[30:56] is kind of like like a Cadillac, you know, it's it just always, you know, it

[30:59] know, it's it just always, you know, it it's a nice cruiser. It never goes too

[31:01] it's a nice cruiser. It never goes too fast.

[31:03] fast. Everybody drives it pretty well. You

[31:05] Everybody drives it pretty well. You know, it's got cruise control. you know,

[31:08] know, it's got cruise control. you know, and everything is easy. But in a lot of

[31:11] and everything is easy. But in a lot of ways, Nvidia's GPUs are accelerators are

[31:14] ways, Nvidia's GPUs are accelerators are kind of like F1 racers. And yeah, I I

[31:18] kind of like F1 racers. And yeah, I I could imagine everybody's able to drive

[31:20] could imagine everybody's able to drive it at 100 100 miles an hour, but it

[31:23] it at 100 100 miles an hour, but it takes quite a bit of expertise to be

[31:24] takes quite a bit of expertise to be able to push it to the limit. And we use

[31:27] able to push it to the limit. And we use we use a ton of AI to create the kernels

[31:30] we use a ton of AI to create the kernels that we have. And um I'm pretty sure

[31:34] that we have. And um I'm pretty sure we're going to still be needed for quite

[31:35] we're going to still be needed for quite some time. And so our expertise um helps

[31:38] some time. And so our expertise um helps our our our um uh our AI labs partners

[31:43] our our our um uh our AI labs partners get another 2x out of their stack

[31:47] get another 2x out of their stack easily. Often times it's not unusual

[31:50] easily. Often times it's not unusual that we you know by the time that we're

[31:51] that we you know by the time that we're done optimizing their stack or

[31:53] done optimizing their stack or optimizing a particular kernel their

[31:55] optimizing a particular kernel their model sped up by 3x 2x 50%.

[32:00] model sped up by 3x 2x 50%. Um, that's a huge number, especially

[32:04] Um, that's a huge number, especially when you're talking about the installed

[32:05] when you're talking about the installed base of the fleet that they have of all

[32:07] base of the fleet that they have of all the hoppers and black walls that they

[32:09] the hoppers and black walls that they have. When you increase it by a factor

[32:11] have. When you increase it by a factor of two, that doubles the revenues. That

[32:15] of two, that doubles the revenues. That directly translates to revenues.

[32:17] directly translates to revenues. Nvidia's computing stack is the best

[32:20] Nvidia's computing stack is the best performance per TCO in the world, bar

[32:22] performance per TCO in the world, bar none.

[32:24] none. Nobody can demonstrate to me that any

[32:27] Nobody can demonstrate to me that any single platform in the world today has

[32:30] single platform in the world today has better performance TCO ratio. Not one

[32:33] better performance TCO ratio. Not one company. And in fact in fact the the uh

[32:36] company. And in fact in fact the the uh the benchmarks are out there uh Dylan's

[32:39] the benchmarks are out there uh Dylan's right inference max is sitting out there

[32:41] right inference max is sitting out there for everybody to to use and not one TPU

[32:44] for everybody to to use and not one TPU won't come trrenium won't come. I I

[32:47] won't come trrenium won't come. I I encourage them to

[32:50] encourage them to use inference max and demonstrate their

[32:53] use inference max and demonstrate their incredible

[32:54] incredible inference cost. It's really really hard.

[32:57] inference cost. It's really really hard. Uh not nobody wants to show up. Uh ML

[33:00] Uh not nobody wants to show up. Uh ML Perf I would I would welcome Trrenium to

[33:04] Perf I would I would welcome Trrenium to demonstrate their 40% that they claim

[33:06] demonstrate their 40% that they claim all the time. I would I would love to to

[33:08] all the time. I would I would love to to hear them demonstrate the the uh cost

[33:11] hear them demonstrate the the uh cost advantage of TPUs. It makes no sense in

[33:13] advantage of TPUs. It makes no sense in my mind. it makes absolutely zero sense

[33:16] my mind. it makes absolutely zero sense on first principles. It makes no sense.

[33:18] on first principles. It makes no sense. And so I I think the I think the the the

[33:21] And so I I think the I think the the the reason why we're so successful is simply

[33:24] reason why we're so successful is simply because our TCO is so great. There's a

[33:27] because our TCO is so great. There's a second you say um 60% of our customers

[33:31] second you say um 60% of our customers are the top five but most of that

[33:34] are the top five but most of that business is external. For example, most

[33:37] business is external. For example, most of AWS is most of Nvidia in AWS is for

[33:40] of AWS is most of Nvidia in AWS is for external customers not internal use.

[33:42] external customers not internal use. Most of our customers at Azure,

[33:44] Most of our customers at Azure, obviously all of our customers are

[33:45] obviously all of our customers are external. All of our customers at OCI

[33:47] external. All of our customers at OCI are external, not internal use. The

[33:49] are external, not internal use. The reason why they they favor us is because

[33:52] reason why they they favor us is because our reach is so great. We can bring them

[33:56] our reach is so great. We can bring them all of the great customers in the world.

[33:57] all of the great customers in the world. They're all built on Nvidia. And the

[33:59] They're all built on Nvidia. And the reason why all these C companies are

[34:00] reason why all these C companies are built on Nvidia is because our reach and

[34:02] built on Nvidia is because our reach and our versatility is so great. And so so I

[34:06] our versatility is so great. And so so I think I think the flywheel

[34:08] think I think the flywheel is is really install base the

[34:11] is is really install base the programmability of our architecture the

[34:14] programmability of our architecture the richness of our ecosystem and the fact

[34:16] richness of our ecosystem and the fact that there's so many AI companies in the

[34:18] that there's so many AI companies in the world there's tens of thousands of them

[34:20] world there's tens of thousands of them now

[34:21] now >> and if you were one of those AI startups

[34:24] >> and if you were one of those AI startups what architecture would you would you

[34:25] what architecture would you would you choose you would choose an architecture

[34:27] choose you would choose an architecture that's most abundant where the most

[34:29] that's most abundant where the most abundant in the world

[34:31] abundant in the world >> the one has the largest installed base

[34:33] >> the one has the largest installed base where the

[34:34] where the largest installed base and one that has

[34:36] largest installed base and one that has a rich ecosystem. And so that's the

[34:38] a rich ecosystem. And so that's the flywheel that that's the reason why

[34:39] flywheel that that's the reason why between the combination of one, our perf

[34:43] between the combination of one, our perf per dollar is so great um that that uh

[34:47] per dollar is so great um that that uh uh they have the lowest cost tokens.

[34:49] uh they have the lowest cost tokens. Second, our perf per watt is the highest

[34:52] Second, our perf per watt is the highest in the world. And so if if uh uh one of

[34:55] in the world. And so if if uh uh one of these companies if our partners built a

[34:58] these companies if our partners built a 1 gawatt data center that 1 gawatt data

[35:01] 1 gawatt data center that 1 gawatt data center better deliver the maximum amount

[35:04] center better deliver the maximum amount of revenues that and number of tokens

[35:07] of revenues that and number of tokens which directly translates to revenues

[35:09] which directly translates to revenues you wanted to generate as many tokens as

[35:10] you wanted to generate as many tokens as possible maximize the revenues for that

[35:12] possible maximize the revenues for that data center. We have the highest tokens

[35:15] data center. We have the highest tokens per watt architecture in the world. And

[35:17] per watt architecture in the world. And then lastly if your goal is to rent the

[35:19] then lastly if your goal is to rent the infrastructure we have the most

[35:21] infrastructure we have the most customers in the world. M and so that's

[35:23] customers in the world. M and so that's the reason why the flywheel works.

[35:25] the reason why the flywheel works. >> Interesting. I I I guess the question

[35:27] >> Interesting. I I I guess the question comes down to what is the actual market

[35:30] comes down to what is the actual market structure here because even if there's

[35:31] structure here because even if there's other companies there could have been a

[35:33] other companies there could have been a world where there's tens of thousands of

[35:34] world where there's tens of thousands of AI companies uh that have roughly equal

[35:37] AI companies uh that have roughly equal share of compute but if even through

[35:39] share of compute but if even through these five hyperscalers really the

[35:41] these five hyperscalers really the people on Amazon using the computer

[35:44] people on Amazon using the computer anthropic openai um and these big big

[35:47] anthropic openai um and these big big foundation labs who who can themselves

[35:49] foundation labs who who can themselves afford and have the ability to make

[35:52] afford and have the ability to make different accelerators work um

[35:54] different accelerators work um >> no I I I think your your your assumption

[35:57] >> no I I I think your your your assumption is is um premise is wrong.

[35:58] is is um premise is wrong. >> Maybe um let me let me let me ask you a

[36:00] >> Maybe um let me let me let me ask you a slightly different question which is

[36:01] slightly different question which is >> come back and make me correct your your

[36:03] >> come back and make me correct your your your um your premise.

[36:04] your um your premise. >> Okay, let me just ask a different

[36:06] >> Okay, let me just ask a different question which is okay if everything

[36:08] question which is okay if everything >> but still make sure that make me come

[36:09] >> but still make sure that make me come back and okay and fix because it's just

[36:11] back and okay and fix because it's just too important to AI it's too important

[36:14] too important to AI it's too important to the future of science is too

[36:16] to the future of science is too important to the future of the industry

[36:18] important to the future of the industry that that premise

[36:20] that that premise >> the premise look let me just finish the

[36:22] >> the premise look let me just finish the question and then we can address it

[36:24] question and then we can address it together. Yeah.

[36:25] together. Yeah. >> So what do you think if if all these

[36:29] >> So what do you think if if all these things are true about uh price

[36:31] things are true about uh price performance and performance per watt etc

[36:33] performance and performance per watt etc are true why why do you think it is the

[36:35] are true why why do you think it is the case that say um anthropic for example

[36:38] case that say um anthropic for example just announced a couple days ago they

[36:40] just announced a couple days ago they have a multi- gigawatt deal with

[36:41] have a multi- gigawatt deal with Broadcom and uh Google for TPUs and

[36:44] Broadcom and uh Google for TPUs and majority of their compute obviously for

[36:46] majority of their compute obviously for Google it's um TPU majority comput so if

[36:48] Google it's um TPU majority comput so if I look at these big AI companies it

[36:50] I look at these big AI companies it seems like a lot of their there was some

[36:52] seems like a lot of their there was some point where it was all Nvidia

[36:54] point where it was all Nvidia and now it's not. And so I'm curious how

[36:58] and now it's not. And so I'm curious how to square

[37:00] to square if these things are true on paper, why

[37:01] if these things are true on paper, why are they going with other accelerators?

[37:03] are they going with other accelerators? >> Yeah, anthropic is is an is a unique

[37:06] >> Yeah, anthropic is is an is a unique instance um and not a trend. Uh without

[37:09] instance um and not a trend. Uh without an anthropic, why would there be any TPU

[37:12] an anthropic, why would there be any TPU growth at all?

[37:14] growth at all? It's 100% anthropic. Without anthropic,

[37:17] It's 100% anthropic. Without anthropic, why would there be any tranium growth at

[37:19] why would there be any tranium growth at all? It's 100% anthropic. And I think

[37:21] all? It's 100% anthropic. And I think that's fairly wellnown and well

[37:23] that's fairly wellnown and well understood. It's not that it's not that

[37:25] understood. It's not that it's not that there's an abundance of ASIC

[37:27] there's an abundance of ASIC opportunities.

[37:29] opportunities. There's only one anthropic,

[37:31] There's only one anthropic, >> but OpenAI deals with AMD. They're

[37:33] >> but OpenAI deals with AMD. They're building their own Titan accelerator.

[37:35] building their own Titan accelerator. >> Yeah. But they're mostly I we could all

[37:36] >> Yeah. But they're mostly I we could all acknowledge they're vastly Nvidia and

[37:39] acknowledge they're vastly Nvidia and and we're going to still do a lot of

[37:41] and we're going to still do a lot of work together.

[37:42] work together. >> Yeah. And we're not we're not I'm not

[37:46] >> Yeah. And we're not we're not I'm not offended by other people using something

[37:48] offended by other people using something else and trying things. If they don't

[37:51] else and trying things. If they don't try these other things, how would they

[37:52] try these other things, how would they know how good ours is, you know? And

[37:55] know how good ours is, you know? And sometimes you got to be reminded of it

[37:57] sometimes you got to be reminded of it and and um we we got to and we have to

[38:00] and and um we we got to and we have to continuously earn earn um uh the

[38:02] continuously earn earn um uh the position that we're in. Uh you there

[38:05] position that we're in. Uh you there always claims and look at the number of

[38:07] always claims and look at the number of AS6 that have been cancelled. Just

[38:10] AS6 that have been cancelled. Just because you're going to build an ASIC,

[38:11] because you're going to build an ASIC, you still have to build something

[38:12] you still have to build something better. than Nvidia.

[38:14] better. than Nvidia. And it's not that easy building

[38:16] And it's not that easy building something better than Nvidia. It's not

[38:17] something better than Nvidia. It's not sensible actually, you know. It's we

[38:20] sensible actually, you know. It's we Nvidia's got to be missing something.

[38:22] Nvidia's got to be missing something. Seriously, you know, and because our our

[38:24] Seriously, you know, and because our our scale, our velocity, we're the only

[38:27] scale, our velocity, we're the only company in the world that's cranking it

[38:28] company in the world that's cranking it out every single year. Big leaps every

[38:31] out every single year. Big leaps every single year.

[38:32] single year. >> I guess their logic is that, hey, it

[38:33] >> I guess their logic is that, hey, it doesn't need to be better. It just needs

[38:34] doesn't need to be better. It just needs to be not more than 70% worse because

[38:37] to be not more than 70% worse because they're paying you 70% margins.

[38:39] they're paying you 70% margins. >> No, no, no. Don't forget uh even an AS6

[38:42] >> No, no, no. Don't forget uh even an AS6 margin is really quite high. Nvidia's

[38:44] margin is really quite high. Nvidia's margin 6 70% let's say but an ASIC

[38:47] margin 6 70% let's say but an ASIC margin is 65.

[38:49] margin is 65. What are you really saving?

[38:51] What are you really saving? >> Oh, you mean from Broadcom or something?

[38:52] >> Oh, you mean from Broadcom or something? >> Yeah, sure.

[38:54] >> Yeah, sure. >> You got to pay somebody.

[38:55] >> You got to pay somebody. >> Yeah.

[38:56] >> Yeah. >> And so so I think the the ASIC margins

[38:58] >> And so so I think the the ASIC margins are are incredibly good from what I can

[39:01] are are incredibly good from what I can tell and and they believe it. They

[39:03] tell and and they believe it. They believe it so too. And so they're

[39:05] believe it so too. And so they're they're quite proud of their their

[39:07] they're quite proud of their their incredible ASIC margins. And so you ask

[39:10] incredible ASIC margins. And so you ask the question why.

[39:12] the question why. A long time ago we just didn't have the

[39:14] A long time ago we just didn't have the ability to do it.

[39:17] ability to do it. And and this is this is this is and at

[39:19] And and this is this is this is and at the time I at the time I didn't deeply

[39:23] the time I at the time I didn't deeply internalize how difficult it would be to

[39:27] internalize how difficult it would be to build a a foundation AI lab

[39:30] build a a foundation AI lab >> like OpenAI and Anthropic.

[39:33] >> like OpenAI and Anthropic. uh and the the fact that they needed

[39:37] uh and the the fact that they needed huge investments from the supplier

[39:39] huge investments from the supplier themselves. Uh we just weren't in a

[39:42] themselves. Uh we just weren't in a position to make the multi-billion

[39:43] position to make the multi-billion dollar investment into anthropic so that

[39:46] dollar investment into anthropic so that they could use our use our compute but

[39:49] they could use our use our compute but Google and and AWS were and they put in

[39:52] Google and and AWS were and they put in huge investments in the beginning so

[39:54] huge investments in the beginning so that anthropic um in return use their

[39:57] that anthropic um in return use their compute. uh we we just weren't in a

[39:59] compute. uh we we just weren't in a position to do so uh at the time. Nor

[40:02] position to do so uh at the time. Nor nor did I I would say my mistake is I

[40:06] nor did I I would say my mistake is I didn't deeply internalize that they they

[40:08] didn't deeply internalize that they they really had no other options that that

[40:11] really had no other options that that that a VC would never put in 510 billion

[40:15] that a VC would never put in 510 billion of investment into an AI lab with the

[40:18] of investment into an AI lab with the with the hopes of it turning out to be

[40:20] with the hopes of it turning out to be anthropic. And so that was my miss. Uh

[40:24] anthropic. And so that was my miss. Uh but even if I understood it, I don't

[40:26] but even if I understood it, I don't think we would have been in a position

[40:27] think we would have been in a position to do that at the time. But um I'm not

[40:30] to do that at the time. But um I'm not going to make that same mistake again.

[40:32] going to make that same mistake again. And and um uh I'm delighted to invest in

[40:35] And and um uh I'm delighted to invest in OpenAI and and um I'm delighted to to uh

[40:39] OpenAI and and um I'm delighted to to uh help them scale and I believe it's

[40:41] help them scale and I believe it's essential to do so. And then and then

[40:44] essential to do so. And then and then when um uh when I was able to uh anth

[40:47] when um uh when I was able to uh anth when Anthropic came to us, I'm delighted

[40:49] when Anthropic came to us, I'm delighted to be an investor, delighted to help

[40:52] to be an investor, delighted to help them scale and um uh but we just weren't

[40:55] them scale and um uh but we just weren't at at the time able to do so.

[40:57] at at the time able to do so. >> If I if I could uh rewind everything, uh

[41:01] >> If I if I could uh rewind everything, uh Nvid Nvidia could have been as big back

[41:03] Nvid Nvidia could have been as big back then as we are now, I would have been

[41:05] then as we are now, I would have been more than happy to do it. This is this

[41:07] more than happy to do it. This is this is actually quite interesting which is

[41:09] is actually quite interesting which is um for many years Nvidia has been this

[41:12] um for many years Nvidia has been this um the company in AI making money making

[41:16] um the company in AI making money making lots of money and um now you're

[41:19] lots of money and um now you're investing it it's been reported that

[41:21] investing it it's been reported that you've done up to 30 billion in open AI

[41:23] you've done up to 30 billion in open AI and 10 billion in um anthropic um but

[41:27] and 10 billion in um anthropic um but now their valuations have increased and

[41:28] now their valuations have increased and I'm sure they'll continue to increase um

[41:30] I'm sure they'll continue to increase um and so if over overall these many years

[41:33] and so if over overall these many years you know you were giving them the

[41:34] you know you were giving them the compute you saw where yeah was headed

[41:36] compute you saw where yeah was headed and then they were worth like onetenth

[41:38] and then they were worth like onetenth what they are now a couple years ago or

[41:39] what they are now a couple years ago or even a year ago in some cases um and you

[41:42] even a year ago in some cases um and you had all this cash

[41:46] there there's a world where either

[41:47] there there's a world where either Nvidia themselves becomes a foundation

[41:49] Nvidia themselves becomes a foundation lab um does the huge investment to make

[41:52] lab um does the huge investment to make that possible or has made the deals

[41:54] that possible or has made the deals you've made now at current valuations

[41:56] you've made now at current valuations much earlier on um and you had the cash

[41:58] much earlier on um and you had the cash to do it so I am curious actually why

[42:00] to do it so I am curious actually why not have done it earlier

[42:02] not have done it earlier >> we did it as soon as we could

[42:05] >> we did it as soon as we could We did it as soon as we could have and

[42:07] We did it as soon as we could have and and and um if I could have, I would have

[42:10] and and um if I could have, I would have done it even earlier. Um at the time

[42:13] done it even earlier. Um at the time that Anthropic needed us to do it, we

[42:15] that Anthropic needed us to do it, we just weren't in a position to do it. It

[42:17] just weren't in a position to do it. It wasn't it wasn't, you know, it wasn't in

[42:19] wasn't it wasn't, you know, it wasn't in our sensibility to do so. How's that

[42:21] our sensibility to do so. How's that like a cash thing or just

[42:23] like a cash thing or just >> Yeah, the level of investment, you know,

[42:25] >> Yeah, the level of investment, you know, we never invested outside the company at

[42:27] we never invested outside the company at the time and not that much and um

[42:32] the time and not that much and um and we didn't realize we needed to,

[42:35] and we didn't realize we needed to, you know, I always I always thought that

[42:37] you know, I always I always thought that they could just go raise VCs for God's

[42:39] they could just go raise VCs for God's sakes like like all companies do. Um but

[42:43] sakes like like all companies do. Um but but um uh what they were trying to what

[42:46] but um uh what they were trying to what they were were trying to do uh couldn't

[42:49] they were were trying to do uh couldn't have been done through VCs. What OpenAI

[42:52] have been done through VCs. What OpenAI wanted to do couldn't have been done

[42:53] wanted to do couldn't have been done through VCs. And and I recognize that

[42:55] through VCs. And and I recognize that now. I didn't know it then, you know,

[42:57] now. I didn't know it then, you know, but that's their genius. That's why

[42:59] but that's their genius. That's why they're smart,

[43:00] they're smart, >> you know, and so so they realized they

[43:02] >> you know, and so so they realized they realized it then that they had to do

[43:03] realized it then that they had to do something like that. And I'm delighted

[43:05] something like that. And I'm delighted that they did, you know, and and even

[43:07] that they did, you know, and and even though even though um we we caused

[43:11] though even though um we we caused Anthropic to have to go to somebody

[43:13] Anthropic to have to go to somebody else, um I'm still happy that it

[43:15] else, um I'm still happy that it happened. Anthropic's existence is great

[43:18] happened. Anthropic's existence is great for the world. I'm I'm delighted for it.

[43:21] for the world. I'm I'm delighted for it. >> Uh I guess you still are making a ton of

[43:23] >> Uh I guess you still are making a ton of money and you're making way more money

[43:24] money and you're making way more money um quarter after quarter.

[43:25] um quarter after quarter. >> It's still okay to have regrets. Um so

[43:29] >> It's still okay to have regrets. Um so then the question still arises okay well

[43:31] then the question still arises okay well now that we're here and you have all

[43:33] now that we're here and you have all this money that you keep making um what

[43:35] this money that you keep making um what should Nvidia be doing with it and

[43:37] should Nvidia be doing with it and there's one answer which says look

[43:38] there's one answer which says look there's this whole middleman ecosystem

[43:40] there's this whole middleman ecosystem that has popped up for converting um

[43:43] that has popped up for converting um capex into opex for these labs so that

[43:46] capex into opex for these labs so that they can rent compute um because the

[43:48] they can rent compute um because the chips are really expensive they make a

[43:50] chips are really expensive they make a lot of money over their lifetime through

[43:51] lot of money over their lifetime through because the models are getting better

[43:53] because the models are getting better the value that they generate their

[43:54] the value that they generate their tokens is increasing but they're

[43:55] tokens is increasing but they're expensive to set up Nvidia has the money

[43:58] expensive to set up Nvidia has the money to do the capex. So, and in fact, you

[44:00] to do the capex. So, and in fact, you are

[44:02] are you're it's been reported you're back

[44:03] you're it's been reported you're back stoping core. We have up to 6.3 billion

[44:05] stoping core. We have up to 6.3 billion and have invested 2B. Um, but yeah, why

[44:08] and have invested 2B. Um, but yeah, why why doesn't Nvidia become

[44:10] why doesn't Nvidia become a cloud themselves? Why doesn't become a

[44:12] a cloud themselves? Why doesn't become a hyperscaler themselves and run this

[44:13] hyperscaler themselves and run this computer out? You have all this cash to

[44:14] computer out? You have all this cash to do it.

[44:15] do it. >> This is a philosophy of the company and

[44:17] >> This is a philosophy of the company and and I think is wise. We should do as

[44:19] and I think is wise. We should do as much as needed as little as possible.

[44:23] much as needed as little as possible. And and what that means is the the work

[44:26] And and what that means is the the work that we do with building our our

[44:28] that we do with building our our computing platform. If we don't if we

[44:30] computing platform. If we don't if we don't do it, I genuinely believe it

[44:33] don't do it, I genuinely believe it doesn't get done. If we didn't take the

[44:35] doesn't get done. If we didn't take the risk that we take, if we didn't build

[44:37] risk that we take, if we didn't build MVLink the way we built, if we didn't

[44:38] MVLink the way we built, if we didn't build the whole stack, if we didn't

[44:40] build the whole stack, if we didn't create the ecosystem the way we did it,

[44:42] create the ecosystem the way we did it, if we didn't dedicate ourselves to 20

[44:44] if we didn't dedicate ourselves to 20 years of CUDA while losing money most of

[44:47] years of CUDA while losing money most of that time, if we didn't do it, nobody

[44:49] that time, if we didn't do it, nobody else would have done it.

[44:52] else would have done it. If we didn't create all the CUDA X

[44:53] If we didn't create all the CUDA X libraries so that they're all domain

[44:55] libraries so that they're all domain specific, you know, this is several a

[44:58] specific, you know, this is several a decade and a half ago, we pushed into

[45:01] decade and a half ago, we pushed into domain specific libraries because we

[45:03] domain specific libraries because we realized that if we didn't create these

[45:04] realized that if we didn't create these domain specific libraries, whether it's

[45:06] domain specific libraries, whether it's for ray tracing or image generation or

[45:09] for ray tracing or image generation or even the early works of AI, these

[45:11] even the early works of AI, these models, if we didn't create them for

[45:13] models, if we didn't create them for data processing, structure data

[45:14] data processing, structure data processing or vector data process, if we

[45:16] processing or vector data process, if we didn't create them, nobody would. And I

[45:19] didn't create them, nobody would. And I am completely certain of that. We