The Supply and Demand of AI Tokens | Dylan Patel Interview

https://www.youtube.com/watch?v=LF3aUIM57uw

[00:00] What used to matter a lot was execution
[00:02] What used to matter a lot was execution was very very difficult and
[00:04] was very very difficult and ideas were cheap.
[00:07] ideas were cheap. Now ideas are cheap
[00:09] and plentiful but execution is very easy.
[00:11] So really only the good ideas are the ones that can justify the spend on
[00:13] the ones that can justify the spend on super cheap implementation.
[00:30] You told me this incredible story about
[00:32] You told me this incredible story about how your own team's use of tokens has
[00:35] how your own team's use of tokens has changed dramatically this year.
[00:37] Yeah.
[00:37] Retell that story and what it is teaching you about what's going on in
[00:39] teaching you about what's going on in the world.
[00:41] the world.
[00:41] Last year we thought we were heavy users
[00:42] Last year we thought we were heavy users of AI.
[00:45] Everyone's using chat GPT.
[00:45] Everyone's using cloud.
[00:47] Everyone's got you know I'm providing whatever
[00:48] you know I'm providing whatever subscriptions anyone wants on the order
[00:50] subscriptions anyone wants on the order of spend of like tens of thousands of
[00:52] of spend of like tens of thousands of dollars for our firm.
[00:54] This year the spend has just skyrocketed and and it
[00:57] spend has just skyrocketed and and it really started in late December with
[00:59] really started in late December with Opus that included Doug who's president
[01:01] Opus that included Doug who's president uh Doug Olaflin.
[01:03] He's very much like uh Doug Olaflin.
[01:04] He's very much like leading the charge in the sense of like leading the charge in the sense of like non-technical people using uh AI for coding.
[01:08] non-technical people using uh AI for coding.
[01:11] Um and so he's basically pled the whole firm slowly over time.
[01:13] I think he's been the the leader in doing that.
[01:14] Obviously the engineers were using it anyways but spend in January just started to inflect and rocket and rocket and rocket and rocket.
[01:18] started to inflect and rocket and rocket and rocket and rocket.
[01:20] Um, we signed, and rocket and rocket.
[01:23] Um, we signed, you know, an enterprise contract with Anthropic and it's gone to the point where now, um, I think when I last talked to you it was 5 million spend rate.
[01:26] Anthropic and it's gone to the point where now, um, I think when I last talked to you it was 5 million spend rate.
[01:28] where now, um, I think when I last talked to you it was 5 million spend rate.
[01:30] talked to you it was 5 million spend rate.
[01:31] It's actually 7 million spend right now.
[01:32] right now.
[01:33] That was last week, by the way.
[01:35] A lot of that is just the usage, right?
[01:37] What's what's really, you know, people people who are have never coded before are using cloud code and spending thousands of dollars sometimes a day,
[01:39] people who are have never coded before are using cloud code and spending thousands of dollars sometimes a day,
[01:41] are using cloud code and spending thousands of dollars sometimes a day,
[01:44] thousands of dollars sometimes a day, but across a firm, we're spending $7 million a year now on cloud code at the current rate.
[01:46] but across a firm, we're spending $7 million a year now on cloud code at the current rate.
[01:49] million a year now on cloud code at the current rate.
[01:52] um versus our salary expense being in the neighborhood of $25 million.
[01:54] expense being in the neighborhood of $25 million.
[01:56] So, you know, we're north of 25% of spend on cloud code as a percentage of salary.
[01:59] 25% of spend on cloud code as a percentage of salary.
[02:01] And if this trajectory continues, then you know,
[02:02] trajectory continues, then you know, we'll spend more than 100% by the end of
[02:03] we'll spend more than 100% by the end of the year.
[02:03] Uh which is a bit terrifying.
[02:06] the year. Uh which is a bit terrifying.
[02:08] Thankfully, I don't have to decide
[02:10] between people and AI because our company's growing so fast.
[02:11] It's, you know, more so like, okay, well, I don't
[02:13] know, more so like, okay, well, I don't have to hire nearly as fast and I can
[02:15] have to hire nearly as fast and I can spend a lot more on AI and it works and
[02:16] spend a lot more on AI and it works and we just grow faster.
[02:18] But I think other folks will start to reckon with the fact
[02:20] folks will start to reckon with the fact that, huh, if this person can do the
[02:23] that, huh, if this person can do the work of five to 10 to 15 people uh using
[02:26] work of five to 10 to 15 people uh using cloud code, then all of a sudden I
[02:29] cloud code, then all of a sudden I should probably cut people.
[02:30] should probably cut people. But right now, I think the use cases are so broad.
[02:33] now, I think the use cases are so broad. For example, one thing is we have a
[02:35] For example, one thing is we have a reverse engineering lab in Oregon that
[02:36] reverse engineering lab in Oregon that we've been building for a year and a
[02:37] we've been building for a year and a half. We have a bunch of, you know,
[02:39] half. We have a bunch of, you know, fancy microscopes, scanning electron
[02:41] fancy microscopes, scanning electron microscopes. The whole purpose of this
[02:42] microscopes. The whole purpose of this is you reverse engineer chips. You get
[02:44] is you reverse engineer chips. You get uh the architecture out of it. you get
[02:46] uh the architecture out of it. you get the materials that they're using to
[02:47] the materials that they're using to manufacture and this is some of the data
[02:48] manufacture and this is some of the data we sell. This is a very slow process of
[02:50] we sell. This is a very slow process of analyzing that data. Instead, um one
[02:53] analyzing that data. Instead, um one person on the team, they've been able to
[02:55] person on the team, they've been able to spend with a couple thousand dollars of
[02:56] spend with a couple thousand dollars of cloud tokens. They've been able to
[02:57] cloud tokens. They've been able to create this application that is GPU
[02:59] create this application that is GPU accelerated runs on a server that we
[03:02] accelerated runs on a server that we have at Coreweave and anytime we send it
[03:04] have at Coreweave and anytime we send it an image, it's able to take the picture
[03:05] an image, it's able to take the picture of the chip and overlay where every
[03:07] of the chip and overlay where every single material is.
[03:09] single material is. Oh, this part is copper.
[03:11] copper. Oh, this part of the gate is uh tantelum.
[03:13] tantelum. This part of the gate is germanium.
[03:14] germanium. This part of the gate is cobalt.
[03:15] cobalt. And so you can do a finite element analysis of the entire stackup
[03:17] element analysis of the entire stackup of the chip very very quickly visual
[03:20] of the chip very very quickly visual with a dashboard guey it's everything
[03:22] with a dashboard guey it's everything few thousand dollars would took claude
[03:23] few thousand dollars would took claude the person previously worked at Intel
[03:25] the person previously worked at Intel and he said that was an entire team's
[03:27] and he said that was an entire team's job to build that and maintain that now
[03:29] job to build that and maintain that now rack that up across you know the entire
[03:31] rack that up across you know the entire firm it's it's insane another example
[03:33] firm it's it's insane another example that I think is super fun is Malcolm
[03:36] that I think is super fun is Malcolm who's an economist at a major bank
[03:38] who's an economist at a major bank before um their economist department was
[03:41] before um their economist department was like 100 or 200 people what he built was
[03:44] like 100 or 200 people what he built was the most incredible thing ever.
[03:46] the most incredible thing ever. He piped all of this different data, you know,
[03:48] all of this different data, you know, FRED data and all these other data,
[03:49] FRED data and all these other data, right? Employment reports and all these
[03:51] right? Employment reports and all these other things from various APIs. We
[03:53] other things from various APIs. We signed a couple contracts with folks to
[03:54] signed a couple contracts with folks to get API access to data. Pulled it all
[03:56] get API access to data. Pulled it all in, started running regression, started
[03:58] in, started running regression, started looking at the impact of various
[04:00] looking at the impact of various economic revolutions on the economy um
[04:03] economic revolutions on the economy um from a deflationary inflationary
[04:05] from a deflationary inflationary perspective.
[04:08] The BLS has this entire um Bureau of Labor Statistics has this entire like set of like 2,000 tasks.
[04:12] And so he did that with AI, which ones can be done by AI, which ones cannot, and grading them across a rubric.
[04:17] You know, about 3% are doable now with AI.
[04:20] Um, and so he's created this like metric so that you can measure things that can be done by AI, what what the massive deflationary uh, you know, what the cost of being able to do those with AI and therefore the deflationary aspect of it.
[04:31] You know, output can go up.
[04:32] It's called phantom GDP is what he called it.
[04:33] Phantom GDP.
[04:35] Output can go up, but because cost falls so much, actually GDP theoretically shrinks.
[04:39] So he created this whole analysis and a brand new benchmark of uh language models um a set of evals across 2,000 different evals.
[04:46] Right.
[04:47] This all by himself.
[04:47] This is all by himself.
[04:49] Yeah.
[04:49] And he's like dude this would have taken the team of 200 economist a year.
[04:53] He's just like he's like completely cracked out on claude.
[04:54] He's like everything has changed.
[04:56] How do you think about as a business owner going from close to zero to 25% accelerating towards whatever percent of total spend?
[05:01] Like at what point are you like, whoa, I need to put the brakes on
[05:06] like, whoa, I need to put the brakes on this and be careful how much we're spending.
[05:10] Maybe we don't need to spend on the most cutting it on Opus 4.7, which came out today.
[05:13] Maybe I can throttle it back to something that's a little bit cheaper.
[05:17] Ultimately, like I'm in the information business, right?
[05:18] That that is, you know, we sell analysis, we sell, we do consulting, we create data sets.
[05:22] I don't see why this wouldn't be completely commoditized on a pretty rapid basis if I'm not constantly improving.
[05:29] my first product that I was selling as a data set actually it is you know like there's more people trying to do it now.
[05:33] we've made it constantly better and better and better and more detailed and so therefore it sells a market.
[05:39] but the way we were doing it in 2023 is not terribly different than you know is it's it's basically what everyone else is doing now.
[05:45] if I don't move up the bar then I will be commoditized.
[05:50] if I don't move fast enough I will also lose my edge so the question is yes AI commoditizes things.
[05:55] just like it commoditizes software those who can move fast and keep control of their customers and keep providing them an awesome service and keep improving the service won't shrink.
[06:05] They'll grow. They'll grow faster. Those
[06:07] They'll grow.
[06:07] They'll grow faster.
[06:08] Those who are incumbent and not doing anything, they're going to lose.
[06:10] And so, it's a bit of an existential like if I don't adopt AI, someone else will and they will beat me.
[06:13] Uh, another easy example is the energy space.
[06:16] So, we've had a few energy analysts for a couple for like a year now.
[06:18] We've been trying to build out this energy model.
[06:20] It's very complex.
[06:21] Energy's data services market is something like $900 million.
[06:23] So obviously a huge market for me to try and break into but it has you know we really hadn't broken into the energy data services business despite a year of having multiple people on the team.
[06:25] Um then cloud code psychosis hits one of the people who leads the data center energy and industrial sort of business at semi analysis uh Jeremy hits him and now all of a sudden in 3 weeks um he spent a lot he was spending like $6,000 a day.
[06:27] It was an insane amount but he scraped every single power plant in the US every single transmission line above a certain voltage.
[06:29] um and created this entire mapping of the entire US grid as well as a lot of demand sources all from various public sources of data.
[06:30] Um and we've shown it to and and we built and it's got like this dashboard where you
[07:08] it's got like this dashboard where you can view and check you can see all the micro regions of the US where there's power deficits and surpluses.
[07:14] Um all of these details built in a handful of weeks we started showing some of our customers who buy our data center data set but are energy like traders.
[07:19] We showed some of them and they're like wow how long did this take you?
[07:25] This is really good. this is better than XYZ company and then we like get dig deeper.
[07:29] XYZ company has 100 people and have been working on this for a decade now.
[07:32] Obviously our thing is not fully robust as robust but in some ways it is better.
[07:36] I'm going to commoditize these energy services companies, data services company.
[07:39] Who's going to come commoditize me if I don't move faster?
[07:40] And so the question from a business owner's perspective is yeah I'm spending a lot but what does that spend getting me?
[07:46] Is it getting more revenue? Yeah.
[07:50] >> Most software companies try to maximize your time on their app to juice engagement.
[07:53] RAMP does the exact opposite.
[07:55] RAMP understands that no one wants to spend hours filing expense reports, reviewing expense reports, and checking for policy violations.
[08:01] So, they built their tools to give that time back, using AI to automate 85% of expense reviews with 99% accuracy.
[08:04] And
[08:10] expense reviews with 99% accuracy.
[08:12] And since RAMP saves companies 5%, it's no wonder that Shopify runs on RAM, Stripe runs on RAM, and my business does, too.
[08:18] To see what happens when you eliminate the busy work, check out ramp.com/invest.
[08:22] OpenAI, Cursor, Enthropic, Perplexity, and Verscell all have something in common.
[08:26] They all use Work OS.
[08:28] And here's why.
[08:30] To achieve enterprise adoption at scale, you have to deliver on core capabilities like SSO, SKIM, Arbback, and audit logs.
[08:37] That's where work OS comes in.
[08:38] Instead of spending months building these missionritical capabilities yourself, you can just use work OS APIs to gain all of them on day zero.
[08:46] That's why so many of the top AI teams you hear about already run on work OS.
[08:51] Work OS is the fastest way to become enterprise ready and stay focused on what matters most, your product.
[08:54] Visit works.com to get started.
[08:58] Felix by Rogo is a personal finance agent that turns a single prompt into finished clientready work using your firm's own templates, context, and standards.
[09:06] Send Felix an email like, "Take these comments and turn them for me, or update my tracker with the context of these emails, or run
[09:12] with the context of these emails, or run the ability to pay math on this buyer.
[09:14] the ability to pay math on this buyer. And Felix sends back finished PowerPoint
[09:16] And Felix sends back finished PowerPoint decks, Excel models, and sourced
[09:18] decks, Excel models, and sourced research. Felix works the way your team
[09:20] research. Felix works the way your team already does, delivering work quickly
[09:21] already does, delivering work quickly and accurately around the clock. Learn
[09:24] and accurately around the clock. Learn more at rogo.ai/felix.
[09:27] more at rogo.ai/felix. Are you worried that in the limit the
[09:29] Are you worried that in the limit the people that control capital and
[09:30] people that control capital and investing capital who are often hiring
[09:32] you for for what you do will just say,
[09:35] you for for what you do will just say, "Well, we have analysts too who are
[09:36] "Well, we have analysts too who are really smart about this. Like, we'll
[09:37] really smart about this. Like, we'll just build this ourselves." Like if it's
[09:39] just build this ourselves." Like if it's getting that easy, at what point does it
[09:41] getting that easy, at what point does it just all pull into the investment firms
[09:44] just all pull into the investment firms that stand to gain the most because they
[09:45] that stand to gain the most because they have the most leverage on top of the
[09:47] have the most leverage on top of the data or the insights that that they
[09:49] data or the insights that that they glean?
[09:49] glean? >> First of all, any information services
[09:51] >> First of all, any information services business, obviously I don't generate as
[09:54] business, obviously I don't generate as much value as my customer does from such
[09:55] much value as my customer does from such information. Uh because if I sell you
[09:57] information. Uh because if I sell you information for a dollar, you're only
[09:59] information for a dollar, you're only buying it for a dollar because you know
[10:01] buying it for a dollar because you know that information helps you make a
[10:02] that information helps you make a decision that lets you make more than
[10:04] decision that lets you make more than $1. And so therefore, you have you have
[10:06] $1. And so therefore, you have you have arbit you you have made more money off
[10:08] arbit you you have made more money off of me than I did from the information
[10:09] of me than I did from the information myself. An investment fund, these
[10:11] myself. An investment fund, these investment funds all have their own
[10:13] Investment funds all have their own information services, you know.
[10:14] Information services, you know, especially like the super like the Jane Streets of the world and the Citadels.
[10:17] Streets of the world and the Citadels.
[10:19] They're they're really detailed on their data.
[10:21] And yet, um, these sort of folks also purchase data from us and continue to do so and continue to grow with us.
[10:25] Because I think there's just some some it factor, right?
[10:27] We move faster, we're more nimble, we're a smaller team that's focused on just one specific thing.
[10:34] uh AI infrastructure and and the huge revolution that causes in AI um and tokconomics and all these things and and we sort of really see where it's headed.
[10:43] And so we're moving faster and building faster.
[10:45] Um I think investment professionals just would you know yes they'll try and build some of the stuff we do and um more likely they'll just buy the data from us and it's cheaper for them to buy the data from us and then to build and then build on top of it than it is to build it themselves.
[11:01] But ultimately some may try.
[11:02] I feel like every conversation I have with you, what I'm always getting at is just supply and demand of tokens like that's the thing that's interesting to me in the world right now.
[11:09] What has this experience taught you about the demand?
[11:11] Has it
[11:13] Taught you about the demand?
[11:14] Has it changed your view on the demand side of that equation?
[11:16] Just feeling it viscerally yourself.
[11:17] If we take a step back and look at the macro lens, right?
[11:20] Enthropic has gone from 9 billion revenue to what they're at 3540 billion now.
[11:24] Probably by the time this airs 40 45 billion, who does ARR?
[11:30] Their compute has not grown to the same degree.
[11:32] Um, and if you do the calculations and you assume they didn't decrease their research and development compute, they clearly didn't.
[11:35] Their release, they have Mythos, they have up is 4.7.
[11:40] So they clearly didn't decrease their research compute spend.
[11:43] Um, so ultimately what they've done, even if you assume all incremental compute they've gotten has gone towards inference, their margins are at a floor of 72%.
[11:51] In reality, some of that incremental compute they've got probably went to research and development.
[11:54] It may be higher than 72% gross margins.
[11:57] To be clear, at the start of the year, they started uh there was um there was a leak by someone from their funding some some of their funding round docs.
[12:03] Someone leaked it 30 something% gross margins.
[12:08] Where on earth does a business like this grow margins like that?
[12:10] And it's in principle, right?
[12:11] Their demand is so high.
[12:11] They're able to cut back on usage
[12:13] high.
[12:13] They're able to cut back on usage limits, rate limits, all these things.
[12:16] limits, rate limits, all these things.
[12:18] Um, what really matters is having an anthropic rep and having an enterprise
[12:19] anthropic rep and having an enterprise contract with them and getting the rate
[12:21] contract with them and getting the rate limit increases that you need because
[12:23] limit increases that you need because otherwise tokens are ultimately super
[12:25] otherwise tokens are ultimately super super in demand.
[12:28] super in demand. Whoever whoever can pay for them anthropic has the same problem,
[12:29] for them anthropic has the same problem, right?
[12:31] right? Like I mean not problem, it's it's just the reality of how capitalism
[12:32] it's just the reality of how capitalism works.
[12:35] works. Yes, people are spending sending them $40 billion AR in tokens and but
[12:38] them $40 billion AR in tokens and but those tokens are generating way more
[12:39] those tokens are generating way more than $40 billion in value.
[12:42] than $40 billion in value. Various businesses will have different value
[12:44] businesses will have different value generation per token. But as we get more
[12:46] generation per token. But as we get more and more intelligent, what really
[12:48] and more intelligent, what really matters is access to these most
[12:49] matters is access to these most intelligent tokens and leveraging them
[12:51] intelligent tokens and leveraging them at things. You as a person deciding what
[12:54] at things. You as a person deciding what is the best way to leverage these tokens
[12:56] is the best way to leverage these tokens to grow business and generate value
[12:58] to grow business and generate value because a lot of folks will want tokens
[13:00] and generate tokens. Uh but the shitty SAS startup and and and and SF who is
[13:02] SAS startup and and and and SF who is using Claude to generate, you know,
[13:06] using Claude to generate, you know, their software product is not
[13:08] their software product is not necessarily actually creating a ton of
[13:09] necessarily actually creating a ton of value and therefore they're going to get
[13:11] value and therefore they're going to get priced out of tokens uh soon enough.
[13:15] priced out of tokens uh soon enough.
[13:16] Are you at all surprised that I I had this experience just today where on the flight here I got rate limited out on something I saw 4.7 came out and what I immediately wanted was like to be on 4.7 that second and I was it just I couldn't think about using 4.6 anymore.
[13:29] or not.
[13:31] This 47 is out.
[13:33] I was perfectly happy with 4.6 for the last many weeks.
[13:35] It's amazing.
[13:37] Are you surprised that people are so insistent on going to the most expensive leading edge thing to the degree they are?
[13:42] Without a doubt.
[13:44] One of my funniest memories in the past month and a half is myself and a buddy of mine, Leopold, being on our knees in front of an anthropic co-founder begging him for access to Methos and then pretending it doesn't exist cuz we knew it existed.
[14:00] were like, "Please give us access."
[14:02] And he's like, "I don't know what you're talking about."
[14:04] What was your reaction to that rate card or that eval card coming out?
[14:08] It was rumored in the Bay Area.
[14:09] Everyone, you know, we sort of like knew it was supposed to be really good, but um if you just look at the benchmarks and obviously benchmarks change over
[14:15] and obviously benchmarks change over time, Mythos is potentially the biggest
[14:18] time, Mythos is potentially the biggest step up in model capabilities in like 2
[14:21] step up in model capabilities in like 2 years. I think that's really really an
[14:23] years. I think that's really really an an important detail that you know it
[14:25] an important detail that you know it it's so good that they're like don't
[14:27] it's so good that they're like don't want to release it even though they're
[14:28] want to release it even though they're they they already announced the price to
[14:31] their people that they did a selective release for cyber for and it's like five
[14:33] their people that they did a selective release for cyber for and it's like five or 10x the token cost. They just don't
[14:34] or 10x the token cost. They just don't want to release it um because they're
[14:36] want to release it um because they're worried about the like impact on the
[14:38] worried about the like impact on the world and they're releasing a shitty
[14:39] world and they're releasing a shitty worse version of open 47 to us and they
[14:42] worse version of open 47 to us and they explicitly said in the model card hey we
[14:45] explicitly said in the model card hey we actually preferentially made it worse at
[14:47] actually preferentially made it worse at cyber. I don't know if you read that.
[14:49] cyber. I don't know if you read that. whoever you are, if you have enough
[14:50] whoever you are, if you have enough capital, you should get a freaking
[14:52] capital, you should get a freaking enterprise cloud uh enterprise anthropic
[14:54] enterprise cloud uh enterprise anthropic subscription where you pay per token,
[14:56] subscription where you pay per token, not with these like subscriptions
[14:58] not with these like subscriptions because then you won't get rate limited
[14:59] because then you won't get rate limited much. And then you must you need to
[15:01] much. And then you must you need to figure out how to leverage those tokens
[15:02] figure out how to leverage those tokens to the highest value task um and make
[15:03] to the highest value task um and make money off of it because ultimately what
[15:05] money off of it because ultimately what you're doing maybe maybe like a year
[15:07] you're doing maybe maybe like a year from now or two years from now the
[15:09] from now or two years from now the business is actually just arbitrageing
[15:10] business is actually just arbitrageing tokens, right? The tokens are amazing,
[15:11] tokens, right? The tokens are amazing, but let's figure out what direction to
[15:13] but let's figure out what direction to point them in and then three or four
[15:16] point them in and then three or four years from now the model will know
[15:17] years from now the model will know, you know, what to do with the tokens and how
[15:18] know, what to do with the tokens and how to make the most value.
[15:20] to make the most value. You know, you can look at this retroactively.
[15:21] can look at this retroactively. Pick any benchmark.
[15:24] benchmark. The cost to hit a certain capability tier used to cost X and now
[15:27] capability tier used to cost X and now it cost 1/100th or 1/ 1,000th of that.
[15:30] it cost 1/100th or 1/ 1,000th of that. Deepseek, for example, on GPD4 was
[15:33] Deepseek, for example, on GPD4 was 1/600th the cost. And since then, the
[15:36] 1/600th the cost. And since then, the costs have fallen further for GPD4 class
[15:38] costs have fallen further for GPD4 class models. Of course, no one gives a crap
[15:41] models. Of course, no one gives a crap about GP4 class models. They want the
[15:43] about GP4 class models. They want the frontier because the frontier lets them
[15:44] frontier because the frontier lets them create the economically valuable things.
[15:46] create the economically valuable things. But GP4 class models can still be used
[15:48] But GP4 class models can still be used in like stuff and so people are using
[15:50] in like stuff and so people are using them in some like tiny use cases. It's
[15:52] them in some like tiny use cases. It's just the cost have fallen so fast. It's
[15:54] just the cost have fallen so fast. It's it's not really what's driving the
[15:55] it's not really what's driving the demand. What's driving the demand is is
[15:57] demand. What's driving the demand is is all these new use cases. Yeah. Current
[16:00] all these new use cases. Yeah. Current 4.6 opus or 4.7 opus tier models a year
[16:04] from now my spend for the same exact quality of the model would probably be
[16:07] from now my spend for the same exact quality of the model would probably be like 70k.
[16:10] quality of the model would probably be like 70k. I bet you it'll be 100 times
[16:13] like 70k. I bet you it'll be 100 times cheaper. irrelevant because I'm going to
[16:15] cheaper. irrelevant because I'm going to be using a way way way better model
[16:17] be using a way way way better model which can do way way better things.
[16:18] which can do way way better things.
[16:20] Enthropic mythos is more expensive as a model but it spends a lot less tokens to
[16:22] model but it spends a lot less tokens to do the thing and therefore it is
[16:24] do the thing and therefore it is actually cheaper in most tasks than 46
[16:26] actually cheaper in most tasks than 46 opus because it's just way more
[16:28] opus because it's just way more efficient even though each individual
[16:29] efficient even though each individual token is smarter.
[16:30] token is smarter.
[16:32] When I last saw you Methos had just come out maybe the day before or something or
[16:34] out maybe the day before or something or the the card had just come out and you
[16:36] the the card had just come out and you said something like uh it actually made
[16:38] said something like uh it actually made you feel like a little scared it was so
[16:40] you feel like a little scared it was so good. What did you mean by that?
[16:41] good. What did you mean by that?
[16:46] Anthropic's whole like goal in 2025 was and and even a lot of 2024 they're like
[16:48] and and even a lot of 2024 they're like hey by the end of 2025 we need an L4
[16:51] hey by the end of 2025 we need an L4 software engineer uh in our model and
[16:54] software engineer uh in our model and and they by and large achieved that with
[16:55] and they by and large achieved that with 46 Opus. What they didn't say is that
[16:57] 46 Opus. What they didn't say is that you know and if you look at Mythos and
[16:59] you know and if you look at Mythos and if you compare like the benchmarks it's
[17:01] if you compare like the benchmarks it's like an L6 engineer. So L4 is like
[17:04] like an L6 engineer. So L4 is like pretty new. L6 is like quite well
[17:06] pretty new. L6 is like quite well experienced. I think Anthropic said that
[17:08] experienced. I think Anthropic said that the model internally was available in
[17:10] the model internally was available in February. So in two months they've gone
[17:13] February. So in two months they've gone from L4 engineer to L6 engineer. Uh
[17:16] from L4 engineer to L6 engineer. Uh what's next? Um you know when when you
[17:19] what's next?
[17:21] Um you know when when you think about the model progress it's only accelerated.
[17:23] Enthropic release cadence has compressed.
[17:25] Open's release cadence has compressed.
[17:27] Why? Because these models generally to make a better model you need a few things right.
[17:28] You need amazing compute.
[17:30] Compute is very expensive and it has a time scale that we you know we track and it's like you know it's growing but like you know it's it's sort of set in stone for the next you know short short term.
[17:38] it's like kind of set in stone what you've already signed.
[17:39] Um there will be delays and shifts and some somehow you can find a little more but it's generally pretty set in stone.
[17:44] There's amazing researchers that people are paying tens of millions of dollars for.
[17:47] And then lastly there's implementation.
[17:49] Historically has been very difficult.
[17:51] If I have an idea now I have to implement it.
[17:52] Implementing is hard.
[17:54] Now ideas are there.
[17:57] Implementation is very easy.
[18:00] It's expensive but it's very easy.
[18:02] So how do you how does one decide what ideas to implement?
[18:05] And it turns out if your implementation is just so much easier now you can just implement more ideas and move on the treadmill faster and faster and faster.
[18:10] Whether that is AI model research and so now your model release cadence is shrunk to down to 2 months from where it was 6 months before
[18:19] months from where it was 6 months before or hey I want to I want to take every power plant in the US and every transmission line and model it and run regressions and see the micro supply and demand.
[18:26] I can also do that. The idea is cheap. You know which idea makes sense? which idea is worth the capital that you have to spend on the tokens because the implementation is there.
[18:35] It's it's that's the I think the key learning and if implementation costs continue to tank which they are um we don't even have mythos yet.
[18:46] It's only been you know a handful of hours since Opus 47 launched but you know my team is pretty excited about it internally.
[18:50] What now comes to the world uh it's a complete reordering of how like economies work.
[18:57] What used to matter a lot was execution was very very difficult and ideas were cheap.
[19:02] Now ideas are cheap and plentiful but execution is very easy.
[19:07] So really only the good ideas are worth are the ones that can justify the spend on super cheap implementation.
[19:12] So are you actually scared or are you just is it just does it just introduce an uncertainty that's hard to grapple with?
[19:18] Uncertainty is there. Um but I do I do
[19:22] Uncertainty is there.
[19:22] Um but I do I do think that causes some fear in terms of think that causes some fear in terms of how does society reform itself?
[19:29] How does one one exist in a world where actually any you exist in a world where actually any you know your ability to implement something know your ability to implement something is not actually that important.
[19:37] Your ability to choose the correct idea for AI to implement and then your ability to sell that idea or sell what the AI has implemented is what matters.
[19:45] Your ability to garner capital towards that is what matters.
[19:49] And going back to the point of like it's very important to have the newest model always.
[19:52] Who's going to have access to the newest model?
[19:53] Anthropics project.
[19:55] I know it's not called earwig, but I troll anthropic people by calling it earwig.
[19:59] Um, glasswig anthropic earwig, you know, where they only release mythos to certain companies for cyber.
[20:05] That's just going to be something that continues.
[20:07] Models will have less broad and less broad deployment.
[20:09] I know I know Open AI and Enthropic and all these people are like, we want to have great AI for everyone.
[20:15] AI is very [ __ ] expensive.
[20:18] Who's going to pay for the trillion dollars of infrastructure?
[20:19] People who have money and can can build useful
[20:22] have money and can can build useful things with AI. And then you don't want
[20:24] things with AI. And then you don't want people to distill your models. So you
[20:25] people to distill your models. So you don't release them broadly. Uh you
[20:27] don't release them broadly. Uh you release them to a fewer and fewer set of
[20:29] release them to a fewer and fewer set of customers. Those customers are also now
[20:31] customers. Those customers are also now wrestling over the tokens unless
[20:33] wrestling over the tokens unless anthropic jacks them. You know, they
[20:34] anthropic jacks them. You know, they could double their pricing on Opus and I
[20:36] could double their pricing on Opus and I would continue to pay and I bet most
[20:37] would continue to pay and I bet most users would continue to pay.
[20:38] users would continue to pay. >> I bet that wouldn't solve their
[20:40] >> I bet that wouldn't solve their humongous capacity problem that they
[20:42] humongous capacity problem that they have. So then the question becomes where
[20:44] have. So then the question becomes where does this cycle end where you know token
[20:47] does this cycle end where you know token usage and therefore the benefits of
[20:49] usage and therefore the benefits of those tokens the additional value
[20:51] those tokens the additional value generated on top of those tokens
[20:52] generated on top of those tokens aggregates among fewer and fewer and
[20:54] aggregates among fewer and fewer and fewer companies. I don't have mythos.
[20:56] fewer companies. I don't have mythos. You know who has mythos? Top freaking
[20:58] You know who has mythos? Top freaking banks. Um now they're only using it for
[21:00] banks. Um now they're only using it for cyber security. But at some point I can
[21:02] cyber security. But at some point I can envision a world where hey maybe I
[21:04] envision a world where hey maybe I because I have an enterprise enthropic
[21:05] because I have an enterprise enthropic contract and because enthropic people
[21:07] contract and because enthropic people kind of like me they're willing to give
[21:09] kind of like me they're willing to give us like slightly earlier access or
[21:11] us like slightly earlier access or slightly higher rate limits or something
[21:13] slightly higher rate limits or something for a model. I hope that's what happens.
[21:15] for a model. I hope that's what happens. And then my competitor whoever that is
[21:18] And then my competitor whoever that is doesn't have that and I'm able to
[21:19] doesn't have that and I'm able to [ __ ] crush them. There are people who
[21:21] [ __ ] crush them. There are people who are like Ken Griffin of Citadel is like
[21:23] are like Ken Griffin of Citadel is like super well-connected and super rich and
[21:25] super well-connected and super rich and he's like he he just signs, you know,
[21:27] he's like he he just signs, you know, who knows? He goes and signs a deal with
[21:28] who knows? He goes and signs a deal with Open Arenthropic that's like, "Yeah, I'm
[21:30] Open Arenthropic that's like, "Yeah, I'm going to get access to your models. Um,
[21:32] going to get access to your models. Um, and I'll buy the first $10 billion worth
[21:35] and I'll buy the first $10 billion worth of tokens each year. So, whenever you
[21:36] of tokens each year. So, whenever you release the model, you know, I'll spend
[21:38] release the model, you know, I'll spend the first 10 billion tokens and then
[21:39] the first 10 billion tokens and then everyone else can get the model after
[21:40] everyone else can get the model after that." And it's like, okay, well, now
[21:42] that." And it's like, okay, well, now what does that do? Well, now he's going
[21:43] what does that do? Well, now he's going to crush everyone in the market. That's
[21:44] to crush everyone in the market. That's just an example. Could be cyber like
[21:46] just an example. Could be cyber like Anthropic is worried about, oh, now I
[21:47] Anthropic is worried about, oh, now I can hack people. could be information
[21:49] can hack people. could be information services business like myself where I
[21:50] services business like myself where I crush someone else. I think you know it
[21:52] crush someone else. I think you know it it's it's such a broad base. We don't
[21:54] it's it's such a broad base. We don't know what these models can do. Anthropic
[21:55] know what these models can do. Anthropic doesn't know what these models can do.
[21:56] doesn't know what these models can do. No one knows what these models can do.
[21:57] No one knows what these models can do. It's up to the end user to figure out
[21:59] It's up to the end user to figure out where they can leverage the tokens to
[22:00] where they can leverage the tokens to see what they can build and imagine
[22:02] see what they can build and imagine which is tremendously productive and
[22:04] which is tremendously productive and uplifting for humanity. But then what
[22:06] uplifting for humanity. But then what happens to the concentration of
[22:07] happens to the concentration of resources and usage of it?
[22:08] resources and usage of it? >> Presumably right now robotics or robots
[22:11] >> Presumably right now robotics or robots consume relatively zero tokens versus
[22:14] consume relatively zero tokens versus everything else. Do you see what's your
[22:16] everything else. Do you see what's your view of that? If that's like a second
[22:18] view of that? If that's like a second demand curve that could start to
[22:19] demand curve that could start to ratchet, there's a new startup every
[22:21] ratchet, there's a new startup every single day, you know, within a mile of
[22:23] single day, you know, within a mile of here trying to build something
[22:24] here trying to build something interesting in robotics.
[22:25] interesting in robotics. >> So there's this concept of software only
[22:27] >> So there's this concept of software only singularity, which is that the world
[22:29] singularity, which is that the world has, you know, AI singularity, but only
[22:31] has, you know, AI singularity, but only in software. And now what about the rest
[22:33] in software. And now what about the rest of the world? Vast majority of the world
[22:35] of the world? Vast majority of the world is physical. You can see the world
[22:38] is physical. You can see the world orient around hardware, not software.
[22:40] orient around hardware, not software. That's actually why I think software
[22:42] That's actually why I think software only singularity is like just a blip and
[22:44] only singularity is like just a blip and not like a you know we we do get
[22:45] not like a you know we we do get everything else because once software is
[22:47] everything else because once software is super easy what makes robots really hard
[22:49] super easy what makes robots really hard it's like programming microcontrollers
[22:51] it's like programming microcontrollers and actuators and controlling all this
[22:52] and actuators and controlling all this stuff is very difficult right now the
[22:55] stuff is very difficult right now the interesting thing about models AI models
[22:57] interesting thing about models AI models is they're actually really inefficient
[22:59] is they're actually really inefficient in learning it's just we're able to give
[23:01] in learning it's just we're able to give them so much data that they're able to
[23:03] them so much data that they're able to learn and surpass us in certain ways
[23:05] learn and surpass us in certain ways robots currently the robot models um
[23:07] robots currently the robot models um VA's uh vision language action models
[23:10] VA's uh vision language action models which is very popular right now is
[23:12] which is very popular right now is probably not going to be the thing that
[23:14] probably not going to be the thing that ultimately scales beyond. They are
[23:16] ultimately scales beyond. They are inefficient in data um and we can't
[23:18] inefficient in data um and we can't scale the data for them fast enough.
[23:20] scale the data for them fast enough. There is going to be some way to large
[23:22] There is going to be some way to large scale pre-train robot models where just
[23:24] scale pre-train robot models where just like humans see all this data throughout
[23:26] like humans see all this data throughout their lives. And what's interesting is
[23:28] their lives. And what's interesting is humans the reason why we're so good is
[23:30] humans the reason why we're so good is we're sample efficient. One example, two
[23:32] we're sample efficient. One example, two example, we're good. And so applying
[23:33] example, we're good. And so applying that to robotics. So once you once you
[23:35] that to robotics. So once you once you have this software only singularity
[23:37] have this software only singularity implementation is super cheap. anyone
[23:38] implementation is super cheap. anyone can start to build these mo people can
[23:40] can start to build these mo people can start to build models that now robots
[23:43] start to build models that now robots are actually useful and so I think in
[23:45] are actually useful and so I think in the next six to 18 months we'll start
[23:46] the next six to 18 months we'll start seeing real breakthroughs in robotics
[23:49] seeing real breakthroughs in robotics that enable few shot learning i.e.
[23:52] that enable few shot learning i.e. there's a pre-trained robot model and
[23:54] there's a pre-trained robot model and now there's a robot that you have hired
[23:56] now there's a robot that you have hired or bought or whatever. You show it a few
[23:58] or bought or whatever. You show it a few examples and it's able to do it. You
[23:59] examples and it's able to do it. You tell it to stack these two things or you
[24:01] tell it to stack these two things or you tell it, hey, this can can actually like
[24:03] tell it, hey, this can can actually like balance perfectly, you know, and and it
[24:05] balance perfectly, you know, and and it starts doing these things.
[24:06] starts doing these things. >> Nicely done.
[24:07] >> Nicely done. >> One shot.
[24:09] >> One shot. >> No, trust me, I've spilled many of
[24:11] >> No, trust me, I've spilled many of times.
[24:12] times. >> So, I think I think robots will get fot
[24:15] >> So, I think I think robots will get fot learning right now. Now, you know,
[24:16] learning right now. Now, you know, there's a lot of companies doing robots
[24:18] there's a lot of companies doing robots for like, you know, advertisement or
[24:19] for like, you know, advertisement or robots for like simple stuff like that,
[24:21] robots for like simple stuff like that, but it'll be like, oh, folding clothes,
[24:23] but it'll be like, oh, folding clothes, but it's going to get really niche like
[24:24] but it's going to get really niche like robots just for cleaning chalkboards.
[24:26] robots just for cleaning chalkboards. Um, and it's a rental service or, you
[24:28] Um, and it's a rental service or, you know, it'll be it'll be a model package
[24:30] know, it'll be it'll be a model package that you download onto your standard
[24:31] that you download onto your standard robot that then does that, right? And
[24:33] robot that then does that, right? And and you pay for that. And anyways, there
[24:35] and you pay for that. And anyways, there will be a huge explosion in physical
[24:36] will be a huge explosion in physical good acceleration and and deflationary
[24:39] good acceleration and and deflationary effects there. But and and so that's
[24:41] effects there. But and and so that's that's ultimately going to keep token
[24:43] that's ultimately going to keep token demand going crazy. I I don't think
[24:45] demand going crazy. I I don't think token demand slows down personally.
[24:46] token demand slows down personally. >> Did you learn anything else about the
[24:48] >> Did you learn anything else about the world based on Mythos's results and how
[24:51] world based on Mythos's results and how it was built? My way of asking like the
[24:53] it was built? My way of asking like the you know if you break down the the
[24:54] you know if you break down the the components of the scaling laws like the
[24:56] components of the scaling laws like the >> So Methos is a materially larger model
[24:58] >> So Methos is a materially larger model than prior models and so yes it is a
[25:01] than prior models and so yes it is a much larger model. Now whether or not
[25:03] much larger model. Now whether or not it's it's what chip it's trained on is
[25:05] it's it's what chip it's trained on is not really relevant. It's the scale and
[25:07] not really relevant. It's the scale and obviously you know to a 100,000 black
[25:09] obviously you know to a 100,000 black wells is equivalent to hundreds of
[25:11] wells is equivalent to hundreds of thousands of prior generation chips.
[25:12] thousands of prior generation chips. TPUs and tranium have their different
[25:14] TPUs and tranium have their different release cadence. So it's not exactly
[25:15] release cadence. So it's not exactly like mirrored one to one. Um but
[25:17] like mirrored one to one. Um but ultimately yes mythos is a significantly
[25:19] ultimately yes mythos is a significantly larger model. It's proof that the
[25:20] larger model. It's proof that the scaling laws still work. Um everything
[25:22] scaling laws still work. Um everything about it shows the trend line continues
[25:24] about it shows the trend line continues of models. More compute into model makes
[25:26] of models. More compute into model makes model better. And along the whole way
[25:28] model better. And along the whole way it's not just more compute into model
[25:29] it's not just more compute into model makes model better. along the whole way
[25:31] makes model better. along the whole way we're also getting these compute
[25:32] we're also getting these compute efficiency wins which are you know as as
[25:35] efficiency wins which are you know as as all this research compute that the labs
[25:37] all this research compute that the labs are spending is actually turning into if
[25:39] are spending is actually turning into if I want x capability tier model every 6
[25:42] I want x capability tier model every 6 months that cost or every two months
[25:43] months that cost or every two months that cost is dramatically decreasing but
[25:45] that cost is dramatically decreasing but then if I scale it up massively I get a
[25:47] then if I scale it up massively I get a humongous capability jump as well and so
[25:50] humongous capability jump as well and so yes it's it's proof that this is still
[25:52] yes it's it's proof that this is still happening Google and anthropic are not
[25:53] happening Google and anthropic are not heavy heavy users of GPUs on the
[25:55] heavy heavy users of GPUs on the training side but openai they'll they'll
[25:58] training side but openai they'll they'll start having their new class of models I
[26:00] start having their new class of models I think they're taking a more sensible
[26:01] think they're taking a more sensible principled approach to scaling uh in
[26:04] principled approach to scaling uh in small steps. Enthropic really went for a
[26:06] small steps. Enthropic really went for a huge jump. We'll see better and better
[26:07] huge jump. We'll see better and better models throughout the year and the
[26:08] models throughout the year and the release cadence is only going to get
[26:10] release cadence is only going to get faster.
[26:10] faster. >> We've gone a long way in the
[26:11] >> We've gone a long way in the conversation with saying almost nothing
[26:13] conversation with saying almost nothing about OpenAI which would have been so
[26:14] about OpenAI which would have been so strange.
[26:15] strange. >> So, so this is this is the interesting
[26:16] >> So, so this is this is the interesting thing. Everyone's like, okay, so
[26:18] thing. Everyone's like, okay, so Anthropics just won, right? You know,
[26:19] Anthropics just won, right? You know, they had Methos in February. They never
[26:21] they had Methos in February. They never even released it cuz they didn't feel
[26:22] even released it cuz they didn't feel the need to. They're already sold out.
[26:23] the need to. They're already sold out. Their revenue is already adding $10
[26:25] Their revenue is already adding $10 billion a month. Um and then you've got
[26:27] billion a month. Um and then you've got Opus 47 today all before open eyes you
[26:31] Opus 47 today all before open eyes you know um alleged Spud release which you
[26:34] know um alleged Spud release which you know media such as the information and
[26:36] know media such as the information and others have have posted about. So
[26:38] others have have posted about. So clearly Anthropic is in the lead right
[26:40] clearly Anthropic is in the lead right and OpenAI is cooked. What's interesting
[26:42] and OpenAI is cooked. What's interesting is because Anthropic has such bounds on
[26:45] is because Anthropic has such bounds on compute and they can only grow it so
[26:48] compute and they can only grow it so fast and sort of to the point of you
[26:49] fast and sort of to the point of you know you know Daria Daria used to gloat
[26:51] know you know Daria Daria used to gloat about how OpenAI was being too
[26:54] about how OpenAI was being too aggressive on compute and Anthropic was
[26:56] aggressive on compute and Anthropic was more sensible in their scaling and now
[26:58] more sensible in their scaling and now Enthropic is like [ __ ] we should have I
[27:00] Enthropic is like [ __ ] we should have I wish we had a lot more compute. OpenAI
[27:01] wish we had a lot more compute. OpenAI is able to pay the bills perfectly fine.
[27:04] is able to pay the bills perfectly fine. In fact, they've raised a ton of money
[27:05] In fact, they've raised a ton of money to get incremental compute in addition
[27:08] to get incremental compute in addition to the irresponsible levels of compute
[27:10] to the irresponsible levels of compute that they were buying from Oracle and
[27:11] that they were buying from Oracle and Core and SoftBank and all these people
[27:13] Core and SoftBank and all these people and Microsoft uh you know such as
[27:15] and Microsoft uh you know such as Tranium. Now they're getting tranium as
[27:16] Tranium. Now they're getting tranium as well from Amazon. Um so so they've done
[27:19] well from Amazon. Um so so they've done this like insane thing on compute and
[27:21] this like insane thing on compute and they need know they also know they need
[27:22] they need know they also know they need more. But what's interesting is if you
[27:24] more. But what's interesting is if you were to say Opus 46, you know, let's
[27:27] were to say Opus 46, you know, let's ignore models getting better over time.
[27:29] ignore models getting better over time. Let's just take diffusion of this
[27:31] Let's just take diffusion of this technology. You and I may get jump on
[27:33] technology. You and I may get jump on the model immediately day one, but other
[27:35] the model immediately day one, but other businesses take time and it takes time
[27:37] businesses take time and it takes time for people to learn and the spark of oh
[27:40] for people to learn and the spark of oh [ __ ] claude psychosis moment doesn't hit
[27:42] [ __ ] claude psychosis moment doesn't hit everyone at the same time. And so by the
[27:45] everyone at the same time. And so by the end of the year, let's say a 46 opus
[27:46] end of the year, let's say a 46 opus tier model the economy would spend
[27:48] tier model the economy would spend $und00 billion on. I don't think that's
[27:50] $und00 billion on. I don't think that's unreasonable. It's spending $40 billion
[27:51] unreasonable. It's spending $40 billion right now.
[27:52] right now. >> That's like a linear extrapolation.
[27:54] >> That's like a linear extrapolation. >> It's a linear extrapolation, not a not
[27:55] >> It's a linear extrapolation, not a not an exponential. To get the exponential,
[27:57] an exponential. To get the exponential, you need the better models. Enthropic
[27:59] you need the better models. Enthropic won't have enough compute to do that.
[28:00] won't have enough compute to do that. And so and and presumably OpenAI and
[28:03] And so and and presumably OpenAI and Google will hit that tier soon enough.
[28:05] Google will hit that tier soon enough. Whoever hits that tier next, sure,
[28:07] Whoever hits that tier next, sure, Enthropic may get to charge 70 plus%
[28:09] Enthropic may get to charge 70 plus% gross margins, but if OpenAI hits it
[28:11] gross margins, but if OpenAI hits it next, they charge 50% gross margins.
[28:14] next, they charge 50% gross margins. They still get all of this incremental
[28:15] They still get all of this incremental demand. And probably they also won't
[28:17] demand. And probably they also won't have enough compute to serve all the
[28:18] have enough compute to serve all the users. And so, sure, maybe Mythos is a
[28:22] users. And so, sure, maybe Mythos is a model where if the world had enough
[28:23] model where if the world had enough compute, it'd be $500 billion of revenue
[28:26] compute, it'd be $500 billion of revenue or something crazy. There is such demand
[28:28] or something crazy. There is such demand for these tokens and such limitations on
[28:30] for these tokens and such limitations on compute, you know, and we see this with
[28:32] compute, you know, and we see this with H100 prices skyrocketing and the useful
[28:34] H100 prices skyrocketing and the useful life of these GPUs continue to extend.
[28:36] life of these GPUs continue to extend. It's pretty clear even the tier 2 lab is
[28:38] It's pretty clear even the tier 2 lab is going to be sold out of tokens, let
[28:39] going to be sold out of tokens, let alone the tier one lab. The tier one lab
[28:41] alone the tier one lab. The tier one lab will have better margins, but the tier
[28:43] will have better margins, but the tier two lab will be sold out and probably
[28:45] two lab will be sold out and probably the tier three lab will also be close to
[28:46] the tier three lab will also be close to sold out. Economic value that the best
[28:48] sold out. Economic value that the best model can deliver is growing faster than
[28:50] model can deliver is growing faster than our ability to actually serve those
[28:52] our ability to actually serve those tokens to people via the infrastructure.
[28:54] tokens to people via the infrastructure. And so this gap will continue to grow
[28:55] And so this gap will continue to grow and the model labs will continue to have
[28:57] and the model labs will continue to have expanding margins until people in the
[28:59] expanding margins until people in the hardware supply chain infrastructure
[29:00] hardware supply chain infrastructure supply chain are like wait no why don't
[29:01] supply chain are like wait no why don't I just jack up my margins. So suffice to
[29:03] I just jack up my margins. So suffice to say I think the assessment today or your
[29:05] say I think the assessment today or your assessment of the demand side is
[29:07] assessment of the demand side is completely explosive in your own
[29:08] completely explosive in your own particular example here at semi analysis
[29:10] particular example here at semi analysis but just more broadly that as people
[29:12] but just more broadly that as people fall in you call it AI psychosis as
[29:14] fall in you call it AI psychosis as people fall into this experience of what
[29:16] people fall into this experience of what they can do the implementation
[29:18] they can do the implementation difficulty going completely away I I've
[29:20] difficulty going completely away I I've certainly felt that you know my own
[29:22] certainly felt that you know my own token spend is just through the absolute
[29:23] token spend is just through the absolute roof just in the matter of weeks so that
[29:26] roof just in the matter of weeks so that that feels like a pretty good assessment
[29:27] that feels like a pretty good assessment anything we're missing on the demand
[29:28] anything we're missing on the demand side
[29:29] side >> if you don't use more tokens you'll
[29:30] >> if you don't use more tokens you'll never escape the permanent underclass
[29:32] never escape the permanent underclass just expand on that.
[29:33] just expand on that. >> So either either you use more tokens and
[29:35] >> So either either you use more tokens and you generate economic value outsized
[29:37] you generate economic value outsized economic value for the use of those
[29:39] economic value for the use of those tokens. Um a lot of people are doing it
[29:40] tokens. Um a lot of people are doing it the boring lazy way. Oh, I guess I'll
[29:42] the boring lazy way. Oh, I guess I'll just work one hour a day instead of
[29:43] just work one hour a day instead of eight hours a day and I'll have AI do
[29:45] eight hours a day and I'll have AI do most of my job. That's the boring way.
[29:47] most of my job. That's the boring way. The cool way is I'll still work eight
[29:49] The cool way is I'll still work eight hours a day and I'll I'll do 8x the work
[29:51] hours a day and I'll I'll do 8x the work and maybe I'll make 5x the money. Um
[29:54] and maybe I'll make 5x the money. Um maybe not you can't do this with a job
[29:55] maybe not you can't do this with a job obviously. There's people who have
[29:57] obviously. There's people who have multiple jobs. Um there's people who
[29:58] multiple jobs. Um there's people who like start companies and start selling
[30:00] like start companies and start selling stuff. get that economic value on on
[30:02] stuff. get that economic value on on this AI before everyone is using it and
[30:04] this AI before everyone is using it and it's table stakes. Uh because it's still
[30:06] it's table stakes. Uh because it's still not table stakes if you don't use more
[30:08] not table stakes if you don't use more tokens and generate the value from them
[30:10] tokens and generate the value from them and capture that value. These there's
[30:11] and capture that value. These there's three different problems here. Using
[30:12] three different problems here. Using more tokens, generating value from those
[30:14] more tokens, generating value from those tokens and capturing value from those
[30:16] tokens and capturing value from those tok uh from the value that you created
[30:17] tok uh from the value that you created from the tokens. Uh if you don't do
[30:19] from the tokens. Uh if you don't do these three things, you'll never escape
[30:20] these three things, you'll never escape the permanent underclass i.e. as models
[30:23] the permanent underclass i.e. as models continue to skyrocket in capability and
[30:25] continue to skyrocket in capability and the concentration of resources
[30:26] the concentration of resources potentially happens.
[30:28] potentially happens. >> Okay, let's talk about supply. what is
[30:29] >> Okay, let's talk about supply. what is going on like how would you describe the
[30:31] going on like how would you describe the frontier of what's changing or what is
[30:33] frontier of what's changing or what is changing at the frontier of supplying
[30:35] changing at the frontier of supplying the the entire stack that's required to
[30:37] the the entire stack that's required to serve all these tokens as the demand
[30:39] serve all these tokens as the demand curve explodes
[30:40] curve explodes >> as demand skyrockets prices are going up
[30:42] >> as demand skyrockets prices are going up for everything on the supply side um
[30:45] for everything on the supply side um whether it be the NGPUs
[30:47] whether it be the NGPUs uh their prices are going up in addition
[30:50] uh their prices are going up in addition their useful life is extending
[30:51] their useful life is extending >> H100 prices look like this
[30:53] >> H100 prices look like this >> yeah exactly there's people who have
[30:54] >> yeah exactly there's people who have argued GPU's full lives are less than 5
[30:56] argued GPU's full lives are less than 5 years complete nonsense
[30:58] years complete nonsense Um there are clusters now resigning
[31:01] Um there are clusters now resigning three or foury old hopper clusters
[31:02] three or foury old hopper clusters resigning for 3 or four more years. Um
[31:05] resigning for 3 or four more years. Um there's A100 clusters that are resigning
[31:07] there's A100 clusters that are resigning for another couple years. So the useful
[31:08] for another couple years. So the useful life is clearly not 5 years. It's maybe
[31:10] life is clearly not 5 years. It's maybe even seven or eight years. Um arguably
[31:12] even seven or eight years. Um arguably we we don't know yet. We'll see. We'll
[31:14] we we don't know yet. We'll see. We'll see when Hopper gets there, but it it's
[31:16] see when Hopper gets there, but it it's clearly not 5 years. So the useful life
[31:17] clearly not 5 years. So the useful life is extending and the prices are going up
[31:19] is extending and the prices are going up on that renewal. So in effect the gross
[31:22] on that renewal. So in effect the gross margin was not 35% on a cluster, it's
[31:25] margin was not 35% on a cluster, it's beyond that. Um so margins are expanding
[31:27] beyond that. Um so margins are expanding in the in the cloud layer. Margins are
[31:30] in the in the cloud layer. Margins are um extremely healthy on the hardware
[31:33] um extremely healthy on the hardware layer with you know Nvidia still
[31:34] layer with you know Nvidia still charging 75 or whatever percent gross
[31:36] charging 75 or whatever percent gross margin as we move down the stack. Memory
[31:38] margin as we move down the stack. Memory obviously margins have skyrocketed
[31:40] obviously margins have skyrocketed there. Places like optics and logic
[31:44] there. Places like optics and logic there are large prepayments um and
[31:46] there are large prepayments um and margins are growing slowly um more so
[31:49] margins are growing slowly um more so the companies that are making chips like
[31:50] the companies that are making chips like Nvidia are paying huge prepayments. So
[31:53] Nvidia are paying huge prepayments. So in effect the cast of capital or timing
[31:55] in effect the cast of capital or timing of cash flow return on invested capital
[31:57] of cash flow return on invested capital is going up even if the gross margin
[31:58] is going up even if the gross margin isn't. And you see this across the whole
[32:00] isn't. And you see this across the whole supply chain. You see ASML is completely
[32:02] supply chain. You see ASML is completely sold out and they need Carl Zeiss to
[32:04] sold out and they need Carl Zeiss to expand faster. Everywhere along the
[32:06] expand faster. Everywhere along the chain
[32:07] chain everyone's either sold out and margins
[32:09] everyone's either sold out and margins are going up or they're getting
[32:10] are going up or they're getting prepayments increases the return on
[32:12] prepayments increases the return on invested capital because the invested
[32:13] invested capital because the invested capital is lower. And so this is like a
[32:15] capital is lower. And so this is like a consistent trend across any part. It's
[32:17] consistent trend across any part. It's it's even like you know a PCB to make a
[32:19] it's even like you know a PCB to make a PCB requires copper foil and that copper
[32:22] PCB requires copper foil and that copper foil is sold out and people are making
[32:23] foil is sold out and people are making prepayments for it. It's like anything
[32:25] prepayments for it. It's like anything and everything that like has a pulse and
[32:28] and everything that like has a pulse and is like sold out. People are like
[32:29] is like sold out. People are like jumping to get more incremental supply
[32:31] jumping to get more incremental supply and fighting over the supply for the
[32:33] and fighting over the supply for the years after.
[32:34] years after. >> As your business scales up, everything
[32:35] >> As your business scales up, everything gets more complex, especially your
[32:37] gets more complex, especially your compliance and security needs. With so
[32:39] compliance and security needs. With so many tools offering band-aids and
[32:40] many tools offering band-aids and patches, it's unfortunately far too easy
[32:42] patches, it's unfortunately far too easy for something to slip through the
[32:43] for something to slip through the cracks. Fortunately, Vanta is a powerful
[32:45] cracks. Fortunately, Vanta is a powerful tool designed to simplify and automate
[32:47] tool designed to simplify and automate your security work and deliver a single
[32:49] your security work and deliver a single source of truth for compliance and risk.
[32:51] source of truth for compliance and risk. There's a reason that Ramp, Cursor, and
[32:53] There's a reason that Ramp, Cursor, and Snowflake all use Vanta. It frees them
[32:55] Snowflake all use Vanta. It frees them to focus on building amazing
[32:56] to focus on building amazing differentiated products, knowing that
[32:58] differentiated products, knowing that compliance and security are under
[32:59] compliance and security are under control. Learn more at vanta.com/invest.
[33:03] control. Learn more at vanta.com/invest. I know firsthand how complex the tech
[33:05] I know firsthand how complex the tech stack is for asset management firms. And
[33:07] stack is for asset management firms. And seemingly every new tool and data source
[33:09] seemingly every new tool and data source makes the problem even worse. Adding
[33:11] makes the problem even worse. Adding more complexity, more headcount, and
[33:12] more complexity, more headcount, and more risk. Ridgeline offers a better way
[33:15] more risk. Ridgeline offers a better way forward. One unified platform that
[33:16] forward. One unified platform that automates away that that automates away
[33:18] automates away that that automates away that complexity across portfolio
[33:20] that complexity across portfolio accounting, reconciliation, reporting,
[33:22] accounting, reconciliation, reporting, trading, compliance, and more. All at
[33:24] trading, compliance, and more. All at scale. Ridgeline is revolutionizing
[33:26] scale. Ridgeline is revolutionizing investment management, helping ambitious
[33:28] investment management, helping ambitious firms scale faster, operate smarter, and
[33:30] firms scale faster, operate smarter, and stay ahead of the curve. See what
[33:31] stay ahead of the curve. See what Ridgeline can unlock for your firm.
[33:33] Ridgeline can unlock for your firm. Schedule a demo at ridgeline.ai.
[33:36] Schedule a demo at ridgeline.ai. What do you think are the most important
[33:37] What do you think are the most important bottlenecks? Like typically in economic
[33:39] bottlenecks? Like typically in economic history when there's this kind of
[33:41] history when there's this kind of demand, supply reorients and rises very
[33:44] demand, supply reorients and rises very very quickly to meet the demand. It
[33:47] very quickly to meet the demand. It seems like it's almost impossible for
[33:48] seems like it's almost impossible for supply right now in this moment to keep
[33:50] supply right now in this moment to keep up. You know, famous last words, every
[33:51] up. You know, famous last words, every every shortage is followed by a glut
[33:53] every shortage is followed by a glut historically. But what are the most
[33:55] historically. But what are the most interesting bottlenecks to you on across
[33:57] interesting bottlenecks to you on across the supply side?
[33:58] the supply side? >> Supply chains are usually very fast to
[34:00] >> Supply chains are usually very fast to react. Um, one unique thing is that our
[34:03] react. Um, one unique thing is that our supply chains now are more complex than
[34:05] supply chains now are more complex than ever. and the things we're building are
[34:06] ever. and the things we're building are more complex than ever and therefore the
[34:07] more complex than ever and therefore the lead times are longer. Um, and it's not
[34:10] lead times are longer. Um, and it's not like we haven't seen 18-monthl long lead
[34:12] like we haven't seen 18-monthl long lead times in other industries. It's just
[34:15] times in other industries. It's just building incremental supply didn't take
[34:17] building incremental supply didn't take years. Um, and this is the case with
[34:19] years. Um, and this is the case with memory, right? Memory can only grow
[34:22] memory, right? Memory can only grow capacity, you know, low double digit
[34:24] capacity, you know, low double digit percentages a year, right? 20s 30% a
[34:27] percentages a year, right? 20s 30% a year. Um, even less for NAND, a little
[34:29] year. Um, even less for NAND, a little bit higher for DRM. Even though the
[34:30] bit higher for DRM. Even though the demand signal was very strong at the end
[34:31] demand signal was very strong at the end of 2025, the memory companies
[34:33] of 2025, the memory companies immediately sort of started reacting.
[34:35] immediately sort of started reacting. None of that incremental capacity really
[34:37] None of that incremental capacity really gets here until the second that they've
[34:38] gets here until the second that they've decided to do in addition to the typical
[34:40] decided to do in addition to the typical 20 to 30%. You know, they can stretch a
[34:43] 20 to 30%. You know, they can stretch a little bit, but really the true
[34:44] little bit, but really the true incremental supply doesn't come till 28,
[34:46] incremental supply doesn't come till 28, which is a very unique thing. Even if
[34:48] which is a very unique thing. Even if they wanted to build as fast as
[34:49] they wanted to build as fast as possible, it doesn't come till 28 uh
[34:51] possible, it doesn't come till 28 uh early late 27 at best. And so the result
[34:54] early late 27 at best. And so the result is memory prices have, you know, gone
[34:57] is memory prices have, you know, gone through the roof. And guess what?
[34:58] through the roof. And guess what? they're going to double and triple
[34:59] they're going to double and triple again. Um, at least on DRAM especially,
[35:02] again. Um, at least on DRAM especially, people are like, "Oh, the memory storage
[35:03] people are like, "Oh, the memory storage is overplayed. Everyone gets it." And
[35:04] is overplayed. Everyone gets it." And it's like, "No, no, no. You don't get
[35:05] it's like, "No, no, no. You don't get it." DM will double or triple from here
[35:08] it." DM will double or triple from here still because that's that's how much
[35:11] still because that's that's how much capacity is required and they have to
[35:13] capacity is required and they have to steal capacity from somewhere else. And
[35:15] steal capacity from somewhere else. And the only way to steal capacity from
[35:16] the only way to steal capacity from somewhere else in a in a capitalist
[35:18] somewhere else in a in a capitalist economy is demand destruction via higher
[35:20] economy is demand destruction via higher pricing. We're not like rationing stuff
[35:21] pricing. We're not like rationing stuff here. And so ultimately, that's what's
[35:23] here. And so ultimately, that's what's going to happen. And so margins continue
[35:24] going to happen. And so margins continue to go up. Um, I think Logic also has
[35:28] to go up. Um, I think Logic also has humongous uh capacity problems. TSMC
[35:30] humongous uh capacity problems. TSMC just had their earnings. Uh, they keep
[35:32] just had their earnings. Uh, they keep upping capex. Ultimately, you know, it
[35:34] upping capex. Ultimately, you know, it takes them quite some time to build
[35:35] takes them quite some time to build fabs. Um, they're trying to do
[35:36] fabs. Um, they're trying to do everything they can to squeeze every
[35:38] everything they can to squeeze every little output out of every fab that they
[35:40] little output out of every fab that they have. But ultimately, they're not
[35:42] have. But ultimately, they're not raising prices fast because they're good
[35:43] raising prices fast because they're good people. It seems like, um, you know,
[35:45] people. It seems like, um, you know, singledigit price increases instead of,
[35:47] singledigit price increases instead of, you know, tripledigit price increases
[35:49] you know, tripledigit price increases like the memory guys have had. And so
[35:50] like the memory guys have had. And so you ultimately have like this like
[35:52] you ultimately have like this like market where yeah TSMC is a great
[35:53] market where yeah TSMC is a great company but are they are they actually
[35:55] company but are they are they actually going to extract all the value? I
[35:56] going to extract all the value? I mentioned things like copper foil, glass
[35:58] mentioned things like copper foil, glass fibers for PCBs, lasers. These are
[36:01] fibers for PCBs, lasers. These are things that are like well understood and
[36:02] things that are like well understood and niche supply chains but they're very
[36:04] niche supply chains but they're very very tight. Um and ultimately upstream
[36:07] very tight. Um and ultimately upstream the semiconductor wafer fabrication
[36:09] the semiconductor wafer fabrication equipment supply chain is one that like
[36:10] equipment supply chain is one that like I still think is it's gone up a lot but
[36:12] I still think is it's gone up a lot but it's still very underappreciated. TSMC
[36:15] it's still very underappreciated. TSMC capex this year they say 56. Uh we've
[36:17] capex this year they say 56. Uh we've had 57.4 4 billion since January. Um,
[36:20] had 57.4 4 billion since January. Um, and we may up it slightly more just
[36:22] and we may up it slightly more just because we see some some ways that they
[36:24] because we see some some ways that they can get incremental capex. But what
[36:26] can get incremental capex. But what people aren't focusing on is what does
[36:27] people aren't focusing on is what does that mean next year and what does that
[36:28] that mean next year and what does that mean the year after? And it turns out 3
[36:31] mean the year after? And it turns out 3 years from now TSMC is going to spend
[36:32] years from now TSMC is going to spend hundred billion on capex. U maybe two
[36:35] hundred billion on capex. U maybe two years from now, right? Maybe 28.
[36:36] years from now, right? Maybe 28. Sincerely, they may spend $und00 billion
[36:38] Sincerely, they may spend $und00 billion on capex in 2028. And people like just
[36:41] on capex in 2028. And people like just can't fathom that. But what does that
[36:42] can't fathom that. But what does that mean for their downstream supply chains?
[36:44] mean for their downstream supply chains? um you know companies like Lamb Research
[36:46] um you know companies like Lamb Research or Applied Materials or ASML or their
[36:48] or Applied Materials or ASML or their further downstream supply chains like
[36:50] further downstream supply chains like MKSI and and all these other companies
[36:52] MKSI and and all these other companies the tail whip just gets whipped harder
[36:54] the tail whip just gets whipped harder and harder and harder and ultimately
[36:57] and harder and harder and ultimately that's a shortage if you know TSMC wants
[36:58] that's a shortage if you know TSMC wants to spend $100 billion in 2028 which is a
[37:01] to spend $100 billion in 2028 which is a real possibility I think people would
[37:03] real possibility I think people would think that's insane but that's a real
[37:04] think that's insane but that's a real real possibility
[37:05] real possibility >> what about other parts of the chip
[37:06] >> what about other parts of the chip ecosystem where GPUs have been
[37:08] ecosystem where GPUs have been completely dominant what about like CPUs
[37:10] completely dominant what about like CPUs or AS6 or things that start to pop out
[37:12] or AS6 or things that start to pop out as both opportunities and bottlenecks
[37:14] as both opportunities and bottlenecks beyond just like Nvidia's GPU dominance.
[37:17] beyond just like Nvidia's GPU dominance. >> Yeah, I mean AS6 are obviously taking
[37:18] >> Yeah, I mean AS6 are obviously taking off, but I'll sort of pivot away from AI
[37:20] off, but I'll sort of pivot away from AI chips to talk about these other things.
[37:22] chips to talk about these other things. There's a project we did on FPGAAS and
[37:24] There's a project we did on FPGAAS and there turns out there's 120 FPGAs per
[37:26] there turns out there's 120 FPGAs per per um next generation rack um AI rack
[37:30] per um next generation rack um AI rack and then like what about all the FPGA
[37:32] and then like what about all the FPGA names CPU wise all these reinforcement
[37:34] names CPU wise all these reinforcement learning environments plus all the slop
[37:36] learning environments plus all the slop code you and I are generating that is
[37:37] code you and I are generating that is now running on some you know Versell
[37:39] now running on some you know Versell instance or whatever it is um or some
[37:42] instance or whatever it is um or some AWS instant or some bucket that we've
[37:43] AWS instant or some bucket that we've spun up all of that requires CPU and so
[37:46] spun up all of that requires CPU and so CPUs are completely sold out and demand
[37:48] CPUs are completely sold out and demand is skyrocketing there
[37:49] is skyrocketing there >> help people understand the role that CPU
[37:51] >> help people understand the role that CPU plays and everything.
[37:52] plays and everything. >> Yeah. So there's two there's two main
[37:53] >> Yeah. So there's two there's two main reasons why you need tons of CPUs. One
[37:55] reasons why you need tons of CPUs. One is when you're doing reinforcement
[37:56] is when you're doing reinforcement learning um the CPU is very critical to
[38:00] learning um the CPU is very critical to that. So so before you would throw all
[38:02] that. So so before you would throw all the internet's data into the model,
[38:03] the internet's data into the model, train it, spit it spits and it it spits
[38:05] train it, spit it spits and it it spits some stuff out. Now you train all the
[38:07] some stuff out. Now you train all the world's internet you put all the
[38:08] world's internet you put all the internet data into the model. Then you
[38:10] internet data into the model. Then you put it in this environment. This
[38:11] put it in this environment. This environment is like hey model try this
[38:13] environment is like hey model try this out and it tries stuff out. It tries a
[38:15] out and it tries stuff out. It tries a bunch of different things and in the end
[38:17] bunch of different things and in the end there is an environment which scores
[38:20] there is an environment which scores whether or not what it tried out is
[38:22] whether or not what it tried out is successful and it grades it. And these
[38:23] successful and it grades it. And these environments can be anything. It can be,
[38:25] environments can be anything. It can be, hey, check if the text was outputed in
[38:27] hey, check if the text was outputed in the right way, structured outputs. It
[38:29] the right way, structured outputs. It can be very simple stuff. It can be very
[38:30] can be very simple stuff. It can be very complex stuff. Um, and people are
[38:32] complex stuff. Um, and people are starting to get into very complex
[38:33] starting to get into very complex things, right? Like, hey, I want you to
[38:36] things, right? Like, hey, I want you to open this file, change it, edit it,
[38:38] open this file, change it, edit it, update it, submit it to this website. I
[38:40] update it, submit it to this website. I want you to open up this physics
[38:41] want you to open up this physics simulation from Seammens and edit this
[38:43] simulation from Seammens and edit this CAD model. So the environments can get
[38:45] CAD model. So the environments can get more and more complex and those
[38:46] more and more complex and those environments run on CPUs. They don't run
[38:48] environments run on CPUs. They don't run on GPUs. They don't run on AS6. The AS6
[38:50] on GPUs. They don't run on AS6. The AS6 run the model that takes the input data
[38:53] run the model that takes the input data from the environment, runs it through
[38:55] from the environment, runs it through the model. The model creates outputs of
[38:57] the model. The model creates outputs of various different trajectories, right?
[38:59] various different trajectories, right? Ways that it think it could solve it um
[39:01] Ways that it think it could solve it um in different instances. those
[39:04] in different instances. those trajectories are graded slashscored and
[39:06] trajectories are graded slashscored and the ones that are successful you train
[39:07] the ones that are successful you train on and you update and you reiterate and
[39:10] on and you update and you reiterate and you iterate iterate iterate and so CPUs
[39:12] you iterate iterate iterate and so CPUs are very useful for that one and then
[39:14] are very useful for that one and then once you have these great models and
[39:16] once you have these great models and you're deploying them those models are
[39:17] you're deploying them those models are generating code they're generating
[39:19] generating code they're generating useful output that useful output it
[39:21] useful output that useful output it doesn't go from a GPU straight to the
[39:23] doesn't go from a GPU straight to the human brain um it goes from a GPU or an
[39:26] human brain um it goes from a GPU or an ASIC through to you know a deployed app
[39:29] ASIC through to you know a deployed app that you're deploying somewhere that
[39:30] that you're deploying somewhere that actually just runs on CPUs so that's
[39:32] actually just runs on CPUs so that's another area where there's a lot of
[39:33] another area where there's a lot of demand and and things are sold out um in
[39:36] demand and and things are sold out um in a large large way.
[39:37] a large large way. >> As you continue to assess and try to be
[39:39] >> As you continue to assess and try to be the world's best informed person on both
[39:41] the world's best informed person on both the trajectory of supply and demand,
[39:43] the trajectory of supply and demand, what are things that you wish you knew
[39:45] what are things that you wish you knew to make that understanding that you
[39:47] to make that understanding that you don't know?
[39:47] don't know? >> I think the hardest area for us um and
[39:52] >> I think the hardest area for us um and for everyone is understanding
[39:54] for everyone is understanding tokconomics, economics of tokens. Um, I
[39:56] tokconomics, economics of tokens. Um, I think we have a really tremendously like
[39:58] think we have a really tremendously like good insight into how much it costs to
[40:00] good insight into how much it costs to run infrastructure, what the cost of
[40:02] run infrastructure, what the cost of tokens are, what the cost of models are,
[40:04] tokens are, what the cost of models are, what the margins of these labs are, but
[40:06] what the margins of these labs are, but the usage and adoption is what's really
[40:08] the usage and adoption is what's really difficult to model, you know,
[40:10] difficult to model, you know, continuously, right? We we have these
[40:12] continuously, right? We we have these like we had like crazy in January, we
[40:13] like we had like crazy in January, we had crazy estimates for February,
[40:14] had crazy estimates for February, anthropic smashed them. How do we
[40:16] anthropic smashed them. How do we calibrate this model? What are the data
[40:17] calibrate this model? What are the data sources for this? February, uh, we had
[40:20] sources for this? February, uh, we had crazy assumptions for March and then
[40:21] crazy assumptions for March and then they smashed them. And everyone sees the
[40:23] they smashed them. And everyone sees the number of 10 billion and they're like
[40:24] number of 10 billion and they're like what the how do they add 10 billion in
[40:26] what the how do they add 10 billion in revenue? Who is using all these tokens?
[40:28] revenue? Who is using all these tokens? Why are they using them? What are they
[40:29] Why are they using them? What are they building with them? And then more
[40:30] building with them? And then more importantly with what they're building
[40:32] importantly with what they're building with these tokens, how is that actually
[40:34] with these tokens, how is that actually diffusing into the economy? And what
[40:35] diffusing into the economy? And what value is that generating? Because it's
[40:37] value is that generating? Because it's not really something that you can
[40:38] not really something that you can capture in any any GDP statistic, right?
[40:41] capture in any any GDP statistic, right? all of the value of the tokens that I
[40:43] all of the value of the tokens that I use get transformed into better
[40:44] use get transformed into better information which I then sell at a
[40:47] information which I then sell at a discount to what people used to sell
[40:48] discount to what people used to sell information for relatively because um
[40:51] information for relatively because um and therefore that information is now
[40:53] and therefore that information is now making its way throughout the economy
[40:55] making its way throughout the economy and and people are making better
[40:56] and and people are making better investment decisions or better
[40:59] investment decisions or better competitive decisions if they're a semi
[41:00] competitive decisions if they're a semi company or data center company or
[41:02] company or data center company or hyperscaler and now how how much what
[41:04] hyperscaler and now how how much what what is the value of this and what has
[41:05] what is the value of this and what has that what has that done to the economy
[41:07] that what has that done to the economy it's clearly by every subjective metric
[41:10] it's clearly by every subjective metric amazing Amazing. But where is the
[41:12] amazing Amazing. But where is the phantom GDP? What is the phantom GDP?
[41:15] phantom GDP? What is the phantom GDP? How do we track the real economic?
[41:17] How do we track the real economic? Because because the GDP metrics are not,
[41:20] Because because the GDP metrics are not, you know, accurate if you were to say
[41:22] you know, accurate if you were to say what is the GDP that Dylan Patel is
[41:23] what is the GDP that Dylan Patel is making. It's tiny compared to what the
[41:25] making. It's tiny compared to what the value that I think is being created. And
[41:27] value that I think is being created. And so ultimately, what is the value being
[41:29] so ultimately, what is the value being created by these tokens? Not on a basis
[41:31] created by these tokens? Not on a basis of, you know, just simple, you know,
[41:33] of, you know, just simple, you know, what is the knock-on effect, right? What
[41:35] what is the knock-on effect, right? What is the knock-on effect of all the things
[41:36] is the knock-on effect of all the things that these things are doing? I think
[41:37] that these things are doing? I think that's the real uh question and
[41:39] that's the real uh question and challenge uh that's hard to measure. I
[41:42] challenge uh that's hard to measure. I think we've got a tremendous, you know,
[41:44] think we've got a tremendous, you know, reading on the supply side of things. I
[41:46] reading on the supply side of things. I think we've got a tremendous reading on
[41:47] think we've got a tremendous reading on even a lot of the demand side signals,
[41:49] even a lot of the demand side signals, but it's it's what is the value these
[41:50] but it's it's what is the value these tokens are generating. That's hard to
[41:52] tokens are generating. That's hard to quantify and measure.
[41:54] quantify and measure. >> I hope we get a chance to do this like
[41:55] >> I hope we get a chance to do this like every 3 months because this changes so
[41:56] every 3 months because this changes so quickly. What do you think is going to
[41:58] quickly. What do you think is going to happen next? Like when I when I come
[42:00] happen next? Like when I when I come back 3 months from now and we're in San
[42:01] back 3 months from now and we're in San Francisco together again, what do you
[42:02] Francisco together again, what do you expect?
[42:03] expect? >> Large scale protests.
[42:05] >> Large scale protests. >> Really?
[42:05] >> Really? >> Yeah. Yeah, I think there will be a
[42:06] >> Yeah. Yeah, I think there will be a large scale protest against anthropic
[42:09] large scale protest against anthropic >> and open AI.
[42:09] >> and open AI. >> Expand on that a little more.
[42:10] >> Expand on that a little more. >> Um, people hate AI. Um, AI is less
[42:14] >> Um, people hate AI. Um, AI is less popular than ICE, less popular than
[42:17] popular than ICE, less popular than politicians. Confused how Pew surveyed
[42:19] politicians. Confused how Pew surveyed this, but apparently AI is less popular
[42:21] this, but apparently AI is less popular than politicians. You know, with
[42:22] than politicians. You know, with Enthropic adding so much revenue, that's
[42:25] Enthropic adding so much revenue, that's going to start causing business changes
[42:26] going to start causing business changes downstream. People are going to get more
[42:28] downstream. People are going to get more and more scared of AI. they'll start
[42:30] and more scared of AI. they'll start blaming more and more of their own
[42:31] blaming more and more of their own problems and things that are, you know,
[42:34] problems and things that are, you know, global, you know, have been deep-seated
[42:36] global, you know, have been deep-seated problems for a long time. Those will
[42:37] problems for a long time. Those will bubble up and be blamed on AI. Um,
[42:41] bubble up and be blamed on AI. Um, probably some politician or some social
[42:43] probably some politician or some social media people will start to be able to
[42:44] media people will start to be able to take uh influencer will be able to start
[42:46] take uh influencer will be able to start taking and weaponizing AI against
[42:49] taking and weaponizing AI against people. You look at the comments of news
[42:51] people. You look at the comments of news articles where Sam Alman had a Molotov
[42:53] articles where Sam Alman had a Molotov cocktail thrown in his house twice in
[42:55] cocktail thrown in his house twice in like two weeks. They're like, people are
[42:57] like two weeks. They're like, people are cheering it on. Uh, and this is just the
[43:00] cheering it on. Uh, and this is just the beginning. So, I think I think we'll see
[43:01] beginning. So, I think I think we'll see large scale protests against AI in three
[43:03] large scale protests against AI in three months.
[43:04] months. >> What is the counterwe to that? Like, how
[43:06] >> What is the counterwe to that? Like, how should the AI industry head that off?
[43:08] should the AI industry head that off? >> First of all, Sam Alman and Dario have
[43:10] >> First of all, Sam Alman and Dario have to stop getting on interviews. They're
[43:11] to stop getting on interviews. They're so uncarismatic.
[43:13] so uncarismatic. I don't know what they're doing. Every
[43:15] I don't know what they're doing. Every interview they do is like, wow, normal
[43:17] interview they do is like, wow, normal people are going to hate you even more.
[43:19] people are going to hate you even more. Like, Sam being on Tucker Carlson
[43:21] Like, Sam being on Tucker Carlson probably made all Republicans hate
[43:22] probably made all Republicans hate OpenAI. And same with Dario. They just
[43:24] OpenAI. And same with Dario. They just have no charisma. I think that's first.
[43:26] have no charisma. I think that's first. Two, they need to start showing
[43:28] Two, they need to start showing uplifting things that can be done with
[43:30] uplifting things that can be done with AI. Um, three, they need to stop talking
[43:32] AI. Um, three, they need to stop talking about how the capabilities are going to
[43:34] about how the capabilities are going to change the whole world constantly
[43:35] change the whole world constantly because then people are going to get
[43:36] because then people are going to get fear of that capability because they
[43:38] fear of that capability because they have no connection.
[43:39] have no connection. >> They don't know how to use it. Yeah.
[43:40] >> They don't know how to use it. Yeah. >> There's no connection to it either. It's
[43:41] >> There's no connection to it either. It's like the average person doesn't know an
[43:43] like the average person doesn't know an anthropic employee. The average person
[43:45] anthropic employee. The average person doesn't know an open eye employee.
[43:46] doesn't know an open eye employee. average person doesn't know who these
[43:48] average person doesn't know who these people are, what their goals are, and
[43:49] people are, what their goals are, and they just view them as like this like
[43:51] they just view them as like this like sneaky cobball of like 5,000 people at
[43:54] sneaky cobball of like 5,000 people at this company that are going to change
[43:54] this company that are going to change the world and automate all the jobs and
[43:56] the world and automate all the jobs and and destroy society. That's what they
[43:58] and destroy society. That's what they view it as. And and as people who are
[44:00] view it as. And and as people who are funding the building of all these data
[44:02] funding the building of all these data centers and and power plants that are
[44:04] centers and and power plants that are going to pollute the world, right? They
[44:05] going to pollute the world, right? They don't quite understand what's happening.
[44:06] don't quite understand what's happening. You know, they have to stop talking
[44:07] You know, they have to stop talking about the future thing that's going to
[44:09] about the future thing that's going to happen and only talk about present, how
[44:10] happen and only talk about present, how uplifting AI is. I think it's a huge
[44:13] uplifting AI is. I think it's a huge reorg and rebranding that needs to be
[44:14] reorg and rebranding that needs to be done.
[44:15] done. >> I love doing this with you. Thanks for
[44:16] >> I love doing this with you. Thanks for your time.
[44:16] your time. >> Awesome. Thanks.
[44:22] >> Your finance team isn't losing money on
[44:24] >> Your finance team isn't losing money on big mistakes. It's leaking through a
[44:25] big mistakes. It's leaking through a thousand tiny decisions nobody's
[44:27] thousand tiny decisions nobody's watching. Ramp puts guard rails on
[44:29] watching. Ramp puts guard rails on spending before it happens. Real-time
[44:31] spending before it happens. Real-time limits, automatic rules, zero
[44:32] limits, automatic rules, zero firefighting. Try it at ramp.com/invest.
[44:37] firefighting. Try it at ramp.com/invest. As your business grows, Vant scales with
[44:39] As your business grows, Vant scales with you, automating compliance, and giving
[44:40] you, automating compliance, and giving you a single source of truth for
[44:42] you a single source of truth for security and risk. Learn more at
[44:44] security and risk. Learn more at vanta.com/invest.
[44:47] Ridgeline is redefining asset management
[44:49] Ridgeline is redefining asset management technology as a true partner, not just a
[44:51] technology as a true partner, not just a software vendor. They've helped firms 5x
[44:53] software vendor. They've helped firms 5x and scale, enabling faster growth,
[44:55] and scale, enabling faster growth, smarter operations, and a competitive
[44:57] smarter operations, and a competitive edge. Visit ridgelineapps.com
[44:59] edge. Visit ridgelineapps.com to see what they can unlock for your
[45:01] to see what they can unlock for your firm.
[45:02] firm. Every investment firm is unique, and
[45:04] Every investment firm is unique, and generic AI doesn't understand your
[45:05] generic AI doesn't understand your process. Rogo does. It's an AI platform
[45:08] process. Rogo does. It's an AI platform built specifically for Wall Street,
[45:09] built specifically for Wall Street, connected to your data, understanding
[45:11] connected to your data, understanding your process, and producing real
[45:12] your process, and producing real outputs. Check them out at
[45:14] outputs. Check them out at rogo.ai/invest.
[45:17] The best AI and software companies from
[45:19] The best AI and software companies from OpenAI to Cursor to Perplexity use work
[45:21] OpenAI to Cursor to Perplexity use work OS to become enterprise ready overnight,
[45:23] OS to become enterprise ready overnight, not in months. Visit works.com to skip
[45:26] not in months. Visit works.com to skip the unglamorous infrastructure work and
[45:28] the unglamorous infrastructure work and focus on your product.

The Supply and Demand of AI Tokens | Dylan Patel Interview

https://www.youtube.com/watch?v=LF3aUIM57uw
Translation: zh-TW

[00:00] 过去很重要的事情是执行
[00:02] 过去很重要的事情是执行非常非常困难，而且
[00:04] 非常非常困难，而且想法很廉价。
[00:07] 想法很廉价。现在想法很廉价
[00:09] 而且丰富，但执行非常容易。
[00:11] 所以真的只有好的想法才能证明在...上的花费是合理的
[00:13] 那些可以证明在超级廉价的实施上的花费是合理的。
[00:30] 你给我讲了一个关于...的不可思议的故事
[00:32] 你给我讲了一个关于你自己的团队今年使用代币的方式发生了巨大变化的不可思议的故事。
[00:35] 你自己的团队今年使用代币的方式发生了巨大变化。
[00:37] 是的。
[00:37] 重述那个故事，以及它在教你关于正在发生的什么事情
[00:39] 在世界上正在发生的事情。
[00:41] 世界上。
[00:41] 去年我们认为我们是重度用户
[00:42] 去年我们认为我们是人工智能的重度用户。
[00:45] 每个人都在使用聊天GPT。
[00:45] 每个人都在使用云。
[00:47] 每个人都有，你知道我提供任何
[00:48] 你知道我提供任何任何人想要的订阅，大约
[00:50] 订阅，任何人想要的，花费大约是几万美元
[00:52] 花费是几万美元，对我们公司来说。
[00:54] 今年花费只是飙升了，而且它
[00:57] 花费只是飙升了，而且它真的始于十二月下旬，与
[00:59] 真的始于十二月下旬，与包括道格在内的Opus，他是总裁
[01:01] Opus，其中包括道格，他是道格·奥拉夫林的总裁。
[01:03] 他非常像道格·奥拉夫林。
[01:04] 他非常像在引领潮流，就像非技术人员使用人工智能进行编码一样。
[01:08] 非技术人员使用人工智能进行编码。
[01:11] 嗯，所以他基本上是慢慢地、随着时间的推移，说服了整个公司。
[01:13] 我认为他一直是这样做的领导者。
[01:14] 显然，工程师们无论如何都在使用它，但在 1 月份的支出刚刚开始发生变化并迅速增长。
[01:18] 开始发生变化并迅速增长。
[01:20] 嗯，我们签署了，并且迅速增长。
[01:23] 嗯，我们签署了，你知道，与 Anthropic 签订了一份企业合同，现在已经到了这个地步，嗯，我想我上次和你谈话时，支出率为 500 万。
[01:26] Anthropic，现在已经到了这个地步，嗯，我想我上次和你谈话时，支出率为 500 万。
[01:28] 现在，嗯，我想我上次和你谈话时，支出率为 500 万。
[01:30] 和你谈话时，支出率为 500 万。
[01:31] 现在实际上是 700 万的支出。
[01:32] 现在。
[01:33] 顺便说一句，那是上周。
[01:35] 其中很多只是使用量，对吧？
[01:37] 真正重要的是，你知道，那些从未写过代码的人正在使用云代码，有时一天花费数千美元，
[01:39] 那些从未写过代码的人正在使用云代码，有时一天花费数千美元，
[01:41] 正在使用云代码，有时一天花费数千美元，
[01:44] 有时一天花费数千美元，但整个公司，我们现在每年在云代码上的支出为 700 万美元，按当前费率计算。
[01:46] 但整个公司，我们现在每年在云代码上的支出为 700 万美元，按当前费率计算。
[01:49] 每年数百万美元，按当前费率计算。
[01:52] 嗯，相比之下，我们的薪资支出约为 2500 万美元。
[01:54] 支出约为 2500 万美元。
[01:56] 所以，你知道，我们云代码的支出占薪资的 25% 以上。
[01:59] 云代码支出占薪资的 25%。
[02:01] 如果这种轨迹继续下去，那么你知道，
[02:02] 轨迹继续，那么你知道，到年底我们将花费超过100%。
[02:03] 到年底我们将花费超过100%。
[02:03] 呃，这有点可怕。
[02:06] 年。呃，这有点可怕。
[02:08] 谢天谢地，我不必决定
[02:10] 在人和人工智能之间，因为我们的公司发展如此之快。
[02:11] 它是，你知道，更像是，好吧，我不需要
[02:13] 知道，更像是，好吧，我不需要像以前那样快地招聘，而且我可以
[02:15] 需要像以前那样快地招聘，而且我可以花更多的钱在人工智能上，而且它奏效了，而且
[02:16] 花更多的钱在人工智能上，而且它奏效了，而且我们只是增长得更快。
[02:18] 但我想其他人会开始认识到这个事实
[02:20] 人们会开始认识到这个事实，嗯，如果这个人能做
[02:23] 那个，嗯，如果这个人能做五个到十个到十五个人的工作，呃，使用
[02:26] 五个到十个到十五个人的工作，呃，使用云代码，那么突然之间我
[02:29] 云代码，那么突然之间我可能应该裁掉一些人。
[02:30] 可能应该裁掉一些人。但现在，我认为用例非常广泛。
[02:33] 现在，我认为用例非常广泛。例如，有一件事是我们有一个
[02:35] 例如，有一件事是我们有一个在俄勒冈州的逆向工程实验室，它
[02:36] 在俄勒冈州的逆向工程实验室，我们已经建造了一年半了。
[02:37] 我们已经建造了一年半了。我们有一堆，你知道，
[02:39] 半。我们有一堆，你知道，花哨的显微镜，扫描电子
[02:41] 花哨的显微镜，扫描电子显微镜。这个的整个目的是
[02:42] 显微镜。这个的整个目的是你逆向工程芯片。你得到
[02:44] 是你逆向工程芯片。你从中获得呃架构。你得到
[02:46] 呃从中获得架构。你得到他们正在使用的材料来
[02:47] 他们正在使用的材料来制造，这是我们销售的一些数据。
[02:48] 制造，这是我们销售的一些数据。这是一个非常缓慢的分析数据的过程。
[02:50] 我们销售。这是一个非常缓慢的分析数据的过程。相反，嗯，一个人
[02:53] 分析这些数据。相反，嗯，团队中的一个人，他们已经能够
[02:55] 团队中的一个人，他们能够花费几千美元的
[02:56] 花费几千美元的云令牌。他们已经能够
[02:57] 云令牌。他们已经能够创建这个应用程序，它是 GPU 加速的
[02:59] 创建这个应用程序，它是 GPU 加速的，运行在我们拥有的服务器上
[03:02] 加速的，运行在我们拥有的 Coreweave 服务器上，并且任何时候我们发送它
[03:04] 在 Coreweave 有，并且任何时候我们发送一张图片，它都能拍下这张照片
[03:05] 一张图片，它都能拍下芯片的照片并覆盖每一个
[03:07] 芯片，并覆盖每一种材料所在的位置。
[03:09] 材料是。哦，这部分是铜。
[03:11] 铜。哦，这个栅极的部分是呃钽。
[03:13] 钽。这个栅极的部分是锗。
[03:14] 锗。这个栅极的部分是钴。
[03:15] 钴。所以你可以对整个堆叠进行有限元分析
[03:17] 对芯片的整个堆叠进行非常非常快速的视觉分析
[03:20] 芯片的非常非常快速的视觉分析，带有一个仪表板界面，它是一切
[03:22] 带有一个仪表板界面，它是一切，几千美元就能做到克劳德
[03:23] 几千美元就能做到克劳德，这个人以前在英特尔工作
[03:25] 这个人以前在英特尔工作，他说那是一个整个团队的
[03:27] 他说那是一个整个团队的工作来构建和维护它，现在
[03:29] 工作来构建和维护它，现在将其推广到你知道的整个
[03:31] 推广到你知道的整个公司，这太疯狂了，另一个例子
[03:33] 公司，这太疯狂了，另一个我认为非常有趣的例子是马尔科姆
[03:36] 我认为非常有趣的例子是马尔科姆，他是一家主要银行的经济学家
[03:38] 他是一家主要银行的经济学家，在此之前，呃，他们的经济学部门是
[03:41] 在此之前，呃，他们的经济学部门有大约 100 或 200 人，他构建了
[03:44] 大约 100 或 200 人，他构建了有史以来最令人难以置信的东西。
[03:46] 有史以来最令人难以置信的东西。他接入了所有这些不同的数据，你知道，
[03:48] 所有这些不同的数据，你知道，FRED 数据和所有这些其他数据，
[03:49] FRED 数据和所有这些其他数据，对吧？就业报告和所有这些
[03:51] 对吧？就业报告和所有这些来自各种 API 的其他东西。我们
[03:53] 来自各种 API 的其他东西。我们与一些人签订了几份合同，以
[03:54] 签订了几份合同，以获取数据 API 访问权限。将所有数据拉取进来，
[03:56] 获取数据 API 访问权限。将所有数据拉取进来，开始运行回归分析，开始
[03:58] 进来，开始运行回归分析，开始查看各种
[04:00] 查看各种经济革命对经济的影响，呃
[04:03] 经济革命对经济的影响，呃，从通缩通胀的角度
[04:05] 從通縮和通貨膨脹的角度來看。
[04:08] 勞工統計局擁有這整個，嗯，勞工統計局擁有這整個像是約 2,000 項任務的集合。
[04:12] 所以他用 AI 做了這件事，哪些可以由 AI 完成，哪些不行，並根據標準評分。
[04:17] 你知道，大約有 3% 現在可以用 AI 完成。
[04:20] 嗯，所以他創建了這個指標，以便你可以衡量那些可以用 AI 完成的事情，以及大規模通縮的呃，你知道，用 AI 完成這些事情的成本，以及因此帶來的通縮方面。
[04:31] 你知道，產出可以增加。
[04:32] 這就是他所說的「幻象 GDP」。
[04:33] 幻象 GDP。
[04:35] 產出可以增加，但由於成本大幅下降，實際上 GDP 理論上會萎縮。
[04:39] 所以他創建了這個完整的分析和一個全新的呃語言模型基準，嗯，一套跨越 2,000 個不同評估的評估。
[04:46] 對。
[04:47] 這一切都是他一個人完成的。
[04:47] 這一切都是他一個人完成的。
[04:49] 是的。
[04:49] 他就像是，老兄，這需要 200 名經濟學家團隊一年的時間。
[04:53] 他就像是，他就像是完全沉迷於 Claude。
[04:54] 他就像是，一切都改變了。
[04:56] 作為一個企業主，你如何看待從接近零到 25% 的加速增長，朝著總支出的任何百分比發展？
[05:01] 就像在什麼時候你會覺得，哇，我需要踩剎車
[05:06] 就像，哇，我需要刹住这一点，并小心我们花了多少钱。
[05:10] 也许我们不需要花钱在 Opus 4.7 上，它今天出来了。
[05:13] 也许我可以把它降到更便宜的东西。
[05:17] 最终，就像我在信息行业一样，对吧？
[05:18] 那就是，你知道，我们销售分析，我们销售，我们做咨询，我们创建数据集。
[05:22] 我不知道为什么这不会在相当快的速度下完全商品化，如果我不断改进的话。
[05:29] 我卖的第一个产品是数据集，实际上你知道现在有更多的人在尝试做这个。
[05:33] 我们不断地把它做得越来越好，越来越详细，因此它会卖得很好。
[05:39] 但我们在 2023 年的做法与你知道的没有什么太大区别，基本上是现在其他人都在做的事情。
[05:45] 如果我不提高标准，我就会被商品化。
[05:50] 如果我行动不够快，我也会失去优势，所以问题是，是的，人工智能会商品化事物。
[05:55] 就像它商品化软件一样，那些能够快速行动并控制客户并继续为他们提供优质服务并不断改进服务的人不会萎缩。
[06:05] 他们会成长。他们会成长得更快。那些
[06:07] 他們會成長。
[06:07] 他們會成長得更快。
[06:08] 那些現任者卻無所作為的人，他們將會輸掉。
[06:10] 所以，這有點像一種生存危機，如果我不採用人工智能，別人就會採用，他們就會打敗我。
[06:13] 呃，另一個簡單的例子是能源領域。
[06:16] 所以，我們已經有幾位能源分析師工作了一年左右了。
[06:18] 我們一直在努力建立這個能源模型。
[06:20] 它非常複雜。
[06:21] 能源數據服務市場大約有九億美元。
[06:23] 所以顯然這是一個巨大的市場，我試圖進入，但你知道，儘管團隊中有幾個人工作了一年，我們實際上並沒有進入能源數據服務業務。
[06:25] 嗯，然後雲代碼精神病擊中了半分析公司中負責數據中心能源和工業業務的傑里，現在突然在三週內，他花了很多錢，他每天花費約六千美元。
[06:27] 這是一筆瘋狂的數額，但他抓取了美國的每一個發電廠，以及每一條高於一定電壓的輸電線路。
[06:29] 嗯，並創建了整個美國電網的完整地圖，以及許多來自各種公共數據來源的需求來源。
[06:30] 嗯，我們已經向...展示了，而且我們已經構建了，它有一個儀表板，在那裡你
[07:08] 它有一个仪表板，您可以在其中查看和检查，可以看到美国所有存在电力赤字和盈余的微观区域。
[07:14] 嗯，所有这些细节都在几周内完成，我们开始向一些购买我们数据中心数据集但从事能源交易的客户展示。
[07:19] 我们向他们展示了一些，他们说哇，这花了你多长时间？
[07:25] 这真的很好。这比 XYZ 公司好，然后我们深入挖掘。
[07:29] XYZ 公司有 100 人，并且已经为此工作了十年。
[07:32] 显然，我们的东西并不完全健壮，但有些方面它更好。
[07:36] 我将把这些能源服务公司、数据服务公司商品化。
[07:39] 如果我不加快速度，谁会来商品化我？
[07:40] 所以从企业主的角度来看，问题是是的，我花了很多钱，但这些花费能给我带来什么？
[07:46] 它能带来更多收入吗？是的。
[07:50] >>大多数软件公司试图最大化您在他们应用程序上的时间以提高参与度。
[07:53] RAMP 则完全相反。
[07:55] RAMP 明白没有人想花几个小时填写费用报告、审查费用报告和检查政策违规行为。
[08:01] 因此，他们构建了工具来将这些时间还给用户，利用人工智能以 99% 的准确率自动处理 85% 的费用审查。
[08:04] 和
[08:10] 費用審核準確率高達 99%。
[08:12] 而且由於 RAMP 為公司節省了 5% 的費用，難怪 Shopify 使用 RAM，Stripe 使用 RAM，我的公司也是如此。
[08:18] 想看看消除繁瑣工作後會發生什麼，請造訪 ramp.com/invest。
[08:22] OpenAI、Cursor、Enthropic、Perplexity 和 Verscell 都有一個共同點。
[08:26] 它們都使用 Work OS。
[08:28] 原因如下。
[08:30] 要實現大規模的企業採用，您必須提供 SSO、SKIM、Arbback 和審計日誌等核心功能。
[08:37] 這就是 Work OS 的用武之地。
[08:38] 與其花費數月時間自行建構這些關鍵任務功能，不如使用 Work OS API 在第一天就獲得所有這些功能。
[08:46] 這就是為什麼您聽說過的許多頂級 AI 團隊都已經在使用 Work OS。
[08:51] Work OS 是做好企業準備並專注於最重要事項（您的產品）的最快方法。
[08:54] 請造訪 works.com 開始使用。
[08:58] Felix by Rogo 是一款個人財務代理，可將單一提示轉換為使用您公司自己的範本、內容和標準完成的客戶就緒工作。
[09:06] 向 Felix 發送一封電子郵件，例如：「將這些評論為我轉換，或根據這些電子郵件的內容更新我的追蹤器，或執行
[09:12] 在这些电子邮件的背景下，或者对这位买家进行支付能力计算。
[09:14] 对这位买家进行支付能力计算。然后 Felix 发回了完成的 PowerPoint
[09:16] 然后 Felix 发回了完成的 PowerPoint 演示文稿、Excel 模型和搜集到的
[09:18] 演示文稿、Excel 模型和搜集到的研究。Felix 的工作方式与您的团队
[09:20] 研究一样。Felix 的工作方式与您的团队已有的方式相同，能够快速地
[09:21] 并且准确地全天候交付工作。了解
[09:24] 并且准确地全天候交付工作。在 rogo.ai/felix 了解更多信息。
[09:27] 在 rogo.ai/felix 了解更多信息。您是否担心最终控制资本和
[09:29] 您是否担心最终控制资本和投资资本的人，他们经常因为您所做的事情而雇佣您，
[09:30] 投资资本的人，他们经常因为您所做的事情而雇佣您，
[09:32] 就会说，
[09:35] 就会说，“嗯，我们也有分析师，他们在这方面非常聪明。比如，我们
[09:36] “嗯，我们也有分析师，他们在这方面非常聪明。比如，我们
[09:37] 就会自己来构建这个。”就像如果
[09:39] 变得如此容易，那么在什么时间点，所有的一切都会
[09:41] 都会拉入那些最能从中获利的投资公司，
[09:44] 因为它们拥有对数据或它们所提取的见解的最大影响力，
[09:45] 因为它们拥有对数据或它们所提取的见解的最大影响力，
[09:47] 拥有对数据或它们所提取的见解的最大影响力，
[09:49] 它们所提取的见解？
[09:49] 提取？>> 首先，任何信息服务
[09:51] >> 首先，任何信息服务业务，显然我从客户那里获得的信息价值不如客户自己从信息中获得的价值高。
[09:54] 业务，显然我从客户那里获得的信息价值不如客户自己从信息中获得的价值高。
[09:55] 信息。嗯，因为如果我卖给你
[09:57] 信息。嗯，因为如果我卖给你信息，价格是一美元，你只
[09:59] 信息，价格是一美元，你只买它是因为你知道
[10:01] 那信息能帮助你做出一个决定，让你赚取的利润超过
[10:02] 那信息能帮助你做出一个决定，让你赚取的利润超过
[10:04] 一美元。因此，你从我这里赚取的利润比我从信息本身获得的利润要多。
[10:06] 一美元。因此，你从我这里赚取的利润比我从信息本身获得的利润要多。
[10:08] 你从我这里赚取的利润比我从信息本身获得的利润要多。
[10:09] 信息本身。投资基金，这些
[10:11] 投资基金，这些投资基金都有自己的
[10:13] 投資基金都有自己的信息服務，你知道的。
[10:14] 信息服務，你知道的，尤其是像像簡街這樣的公司和像 Citadel 這樣的公司。
[10:17] 世界各地的街道和 Citadel。
[10:19] 他們對他們數據的細節非常詳盡。
[10:21] 然而，嗯，這類人也從我們這裡購買數據，並繼續這樣做，並與我們一起成長。
[10:25] 因為我認為這就是一些一些關鍵因素，對吧？
[10:27] 我們行動更快，我們更靈活，我們是一個更小的團隊，專注於一件特定的事情。
[10:34] 呃人工智能基礎設施以及它在人工智能中引起的巨大革命，嗯，以及代幣經濟學和所有這些事情，我們 sort of 真的看到了它的發展方向。
[10:43] 所以我們行動更快，構建也更快。
[10:45] 嗯，我認為投資專業人士只是，你知道的，是的，他們會嘗試構建我們做的一些東西，嗯，更有可能的是，他們會直接從我們這裡購買數據，對他們來說，從我們這裡購買數據然後構建，然後在其之上構建比他們自己構建更便宜。
[11:01] 但最終有些人可能會嘗試。
[11:02] 我覺得我每次和你談話，我總是在談論代幣的供需，就像這是我現在在這個世界上感興趣的事情一樣。
[11:09] 這次經歷讓你對需求學到了什麼？
[11:11] 它是否
[11:13] 關於需求，我教過你嗎？
[11:14] 這是否改變了你對該方程式中需求方面的看法？
[11:16] 你自己親身感受一下。
[11:17] 如果我們退一步，從宏觀角度來看，對吧？
[11:20] Enthropic 的收入從 90 億美元增長到現在的 3540 億美元。
[11:24] 可能等到這個節目播出時，400 億到 450 億美元，誰能達到 ARR？
[11:30] 他們的計算能力沒有達到相同的程度。
[11:32] 嗯，如果你做計算並假設他們沒有減少研發計算，他們顯然沒有。
[11:35] 他們的發布，他們有 Mythos，他們有 4.7。
[11:40] 所以他們顯然沒有減少研究計算支出。
[11:43] 嗯，所以最終他們所做的，即使你假設他們獲得的所有增量計算都用於推理，他們的利潤率也處於 72% 的地板價。
[11:51] 實際上，他們獲得的一些增量計算可能用於研發。
[11:54] 毛利率可能高於 72%。
[11:57] 需要說明的是，在年初，他們開始時，嗯，有人從他們的融資文件中洩露了一些信息。
[12:03] 有人洩露了 30 多個百分點的毛利率。
[12:08] 像這樣的企業，利潤率到底是如何增長的？
[12:10] 而且這是原則上的，對吧？
[12:11] 他們的市場需求如此之高。
[12:11] 他們能夠削減使用量
[12:13] 高。
[12:13] 他們能夠削減使用限制、速率限制，所有這些東西。
[12:16] 限制、速率限制，所有這些東西。
[12:18] 嗯，真正重要的是要有一個 Anthropic 代表，並有一個企業級
[12:19] Anthropic 代表並與他們簽訂企業合約，並獲得速率
[12:21] 合約，並獲得您需要的速率限制增加，因為
[12:23] 限制增加，因為否則 token 最終會非常
[12:25] 否則 token 會非常非常搶手。
[12:28] 搶手。誰能為他們付費，Anthropic 也有同樣的問題，
[12:29] 為他們付費，Anthropic 也有同樣的問題，對吧？
[12:31] 對吧？我的意思是，不是問題，這只是資本主義運作的現實。
[12:32] 這只是資本主義運作的現實。
[12:35] 運作。是的，人們正在花費 400 億美元購買 token，但是
[12:38] 他們 400 億美元的 token，但是這些 token 產生的價值遠遠超過
[12:39] 400 億美元。各種
[12:42] 價值。各種企業每 token 的價值產生會有所不同。
[12:44] 但隨著我們變得越來越
[12:46] 智能，真正重要的是獲得這些最智能的 token 並加以利用。
[12:48] 智能，真正重要的是獲得這些最智能的 token 並加以利用。
[12:49] 智能的 token 並加以利用。
[12:51] 事物。你作為一個人，決定什麼是利用這些 token 來發展業務和創造價值的最佳方式
[12:54] 事物。你作為一個人，決定什麼是利用這些 token 來發展業務和創造價值的最佳方式
[12:56] 來發展業務和創造價值，因為很多人會想要 token
[12:58] 並產生 token。嗯，但那個糟糕的 SAS 新創公司，在舊金山，使用 Claude 來生成，你知道，
[13:00] 並產生 token。嗯，但那個糟糕的 SAS 新創公司，在舊金山，使用 Claude 來生成，你知道，
[13:02] SAS 新創公司，在舊金山，使用 Claude 來生成，你知道，他們的軟體產品不一定真的創造了很多價值，因此他們很快就會被 token 定價。
[13:06] 你知道，他們的軟體產品不一定真的創造了很多價值，因此他們很快就會被 token 定價。
[13:08] 他們的軟體產品不一定真的創造了很多價值，因此他們很快就會被 token 定價。
[13:09] 價值，因此他們很快就會被 token 定價。
[13:11] 價值，因此他們很快就會被 token 定價。
[13:15] 很快就买不起代币了。
[13:16] 你对我今天遇到的这种情况一点也不感到惊讶吗？在这里的航班上，我因为看到了 4.7 的发布而被限制了，而我立刻就想在下一秒就用上 4.7，但我就是无法再考虑使用 4.6 了。
[13:29] 或者不。
[13:31] 这个 47 已经出来了。
[13:33] 在过去几周里，我一直对 4.6 非常满意。
[13:35] 太棒了。
[13:37] 你是否惊讶于人们如此坚持要使用最昂贵的前沿技术，以至于达到了他们所达到的程度？
[13:42] 毫无疑问。
[13:44] 过去一个半月里我最有趣的回忆之一就是我和我的朋友利奥波德，双膝跪在一位 Anthropic 的联合创始人面前，恳求他给我们 Methos 的访问权限，然后假装它不存在，因为我们知道它存在。
[14:00] 我们就像说：“请给我们访问权限。”
[14:02] 而他说：“我不知道你在说什么。”
[14:04] 你对那个费率卡或那个评估卡出来有什么反应？
[14:08] 这在湾区有传言。
[14:09] 你知道，大家都差不多知道它应该会非常好，但是，嗯，如果你只看基准测试，而且显然基准测试会随着时间而变化
[14:15] 顯然，基準會隨著時間而改變，Mythos 可能是最大的
[14:18] 時間，Mythos 可能是模型功能方面最大的進步，在過去的兩年裡
[14:21] 兩年裡模型功能方面的進步。我認為這是一個非常非常重要的細節，你知道
[14:23] 你知道，它太好了，以至於他們甚至不想發布它，儘管他們已經宣布了價格
[14:25] 對於他們的員工，他們為 Cyber​​ 進行了選擇性發布，而且它的代幣成本是 5 到 10 倍。
[14:27] 他們只是不想發布它，因為他們擔心它對世界的影響，他們正在向我們發布一個糟糕的、更糟糕的開源 47 版本，
[14:28] 他們明確表示，在模型卡中，我們實際上優先使其在網絡安全方面變得更糟。我不知道你是否讀過。
[14:31] 無論你是誰，如果你有足夠的資本，你應該獲得一個該死的企業雲或企業 Anthropic 訂閱，你按代幣付費，
[14:33] 而不是像這些訂閱一樣，因為這樣你就不會受到 كثيرا的速率限制。然後你必須弄清楚如何利用這些代幣來完成最高價值的任務，並從中獲利，因為最終你可能在一年或兩年後
[14:34] 從現在開始，這項業務實際上只是在套利代幣，對嗎？代幣很棒，但讓我們弄清楚將它們指向哪個方向，然後三四年
[14:36] 他們只是不想發布它，因為他們擔心它對世界的影響，他們正在向我們發布一個糟糕的、更糟糕的開源 47 版本，
[14:38] 他們擔心它對世界的影響，他們正在向我們發布一個糟糕的、更糟糕的開源 47 版本，
[14:39] 他們明確表示，在模型卡中，我們實際上優先使其在網絡安全方面變得更糟。我不知道你是否讀過。
[14:42] 他們明確表示，在模型卡中，我們實際上優先使其在網絡安全方面變得更糟。我不知道你是否讀過。
[14:45] 無論你是誰，如果你有足夠的資本，你應該獲得一個該死的企業雲或企業 Anthropic 訂閱，你按代幣付費，
[14:47] 實際上優先使其在網絡安全方面變得更糟。我不知道你是否讀過。
[14:49] 無論你是誰，如果你有足夠的資本，你應該獲得一個該死的企業雲或企業 Anthropic 訂閱，你按代幣付費，
[14:50] 無論你是誰，如果你有足夠的資本，你應該獲得一個該死的企業雲或企業 Anthropic 訂閱，你按代幣付費，
[14:52] 企業雲或企業 Anthropic 訂閱，你按代幣付費，
[14:54] 企業 Anthropic 訂閱，你按代幣付費，而不是像這些訂閱一樣，因為這樣你就不會受到 كثيرا的速率限制。
[14:56] 你按代幣付費，而不是像這些訂閱一樣，因為這樣你就不會受到 كثيرا的速率限制。
[14:58] 而不是像這些訂閱一樣，因為這樣你就不會受到 كثيرا的速率限制。
[14:59] 因為這樣你就不會受到 كثيرا的速率限制。然後你必須弄清楚如何利用這些代幣來完成最高價值的任務，並從中獲利，因為最終你可能在一年或兩年後
[15:01] 然後你必須弄清楚如何利用這些代幣來完成最高價值的任務，並從中獲利，因為最終你可能在一年或兩年後
[15:02] 弄清楚如何利用這些代幣來完成最高價值的任務，並從中獲利，因為最終你可能在一年或兩年後
[15:03] 完成最高價值的任務，並從中獲利，因為最終你可能在一年或兩年後
[15:05] 獲利，因為最終你可能在一年或兩年後
[15:07] 你可能在一年或兩年後
[15:09] 從現在開始，這項業務實際上只是在套利代幣，對嗎？
[15:10] 這項業務實際上只是在套利代幣，對嗎？代幣很棒，但讓我們弄清楚將它們指向哪個方向，然後三四年
[15:11] 代幣很棒，但讓我們弄清楚將它們指向哪個方向，然後三四年
[15:13] 但讓我們弄清楚將它們指向哪個方向，然後三四年
[15:16] 将它们指向其中，然后从现在起三到四年，模型就会知道
[15:17] 从现在起几年后，模型就会知道，你知道，如何处理令牌以及如何
[15:18] 知道，如何处理令牌以及如何创造最大的价值。
[15:20] 创造最大的价值。你知道，你可以回顾性地看待这件事。
[15:21] 可以回顾性地看待这件事。选择任何基准。
[15:24] 基准。达到某个能力等级的成本曾经是 X，而现在
[15:27] 能力等级的成本曾经是 X，而现在成本是它的 1/100 或 1/1000。
[15:30] 成本是它的 1/100 或 1/1000。例如，Deepseek 在 GPD4 上的成本是
[15:33] Deepseek，例如，在 GPD4 上的成本是 1/600。此后，
[15:36] 1/600 的成本。此后，GPD4 级模型的成本进一步下降。
[15:38] GPD4 级模型的成本进一步下降。当然，没有人关心
[15:41] GP4 级模型。他们想要的是前沿模型，因为前沿模型能让他们
[15:43] GP4 级模型。他们想要的是前沿模型，因为前沿模型能让他们
[15:44] 前沿模型，因为前沿模型能让他们创造经济上有价值的东西。
[15:46] 创造经济上有价值的东西。但是 GP4 级模型仍然可以用于
[15:48] 但是 GP4 级模型仍然可以用于类似的东西，所以人们正在使用
[15:50] 在一些类似的小用例中。这只是因为成本下降得太快了。
[15:52] 在一些类似的小用例中。这只是因为成本下降得太快了。这
[15:54] 只是因为成本下降得太快了。这并不是真正驱动
[15:55] 这并不是真正驱动需求。驱动需求的是
[15:57] 需求。驱动需求的是所有这些新的用例。是的。目前的
[16:00] 所有这些新的用例。是的。目前的 4.6 opus 或 4.7 opus 级模型，一年后
[16:04] 从现在起，我为模型相同质量的支出可能会是
[16:07] 从现在起，我为模型相同质量的支出可能会是 7 万。
[16:10] 模型质量的支出可能会是 7 万。我敢打赌它会便宜 100 倍。
[16:13] 7 万。我敢打赌它会便宜 100 倍。无关紧要，因为我将要
[16:15] 便宜。无关紧要，因为我将要使用一个好得多得多的模型
[16:17] 使用一个好得多的模型，它可以做许多更好的事情。
[16:18] 它可以做许多更好的事情。
[16:20] Enthropic mythos作为一个模型更昂贵，但它花费的令牌少得多，用于
[16:22] 模型，但它花费的令牌少得多，用于完成任务，因此它
[16:24] 完成任务，因此它在大多数任务中实际上比 46
[16:26] 实际上比 46 opus 更便宜，因为它效率高得多
[16:28] opus，因为它的效率高得多，即使每个单独的
[16:29] 效率高，即使每个单独的令牌都更智能。
[16:30] 令牌更智能。
[16:32] 我上次见到你时，Methos 可能在前一天或类似的时候刚刚发布，或者
[16:34] 刚刚发布，或者卡刚刚发布，你
[16:36] 刚刚发布，你说的话，好像它实际上让你
[16:38] 感觉有点害怕，它太好了。
[16:40] 你感觉有点害怕，它太好了。你是什么意思？
[16:41] 太好了。你是什么意思？
[16:46] Anthropic 在 2025 年的整个目标是，甚至在 2024 年的很多时候，他们都说
[16:48] 甚至在 2024 年的很多时候，他们都说嘿，到 2025 年底，我们需要一个 L4
[16:51] 嘿，到 2025 年底，我们需要一个 L4 软件工程师在我们的模型中，并且
[16:54] 软件工程师在我们的模型中，并且他们基本上已经实现了，用
[16:55] 并且他们基本上已经用 46 Opus 实现了。他们没有说的是
[16:57] 46 Opus。他们没有说的是，你知道，如果你看看 Methos，并且
[16:59] 你知道，如果你看看 Methos，并且如果你比较一下基准测试，它就像
[17:01] 如果你比较一下基准测试，它就像一个 L6 工程师。所以 L4 就像
[17:04] 一个 L6 工程师。所以 L4 就像一个新手。L6 就像经验丰富。
[17:06] 新手。L6 就像经验丰富。我认为 Anthropic 说
[17:08] 经验丰富。我认为 Anthropic 说模型在内部于
[17:10] 二月可用。所以在两个月内，他们已经从
[17:13] 二月。所以在两个月内，他们已经从 L4 工程师变成了 L6 工程师。嗯
[17:16] 从 L4 工程师变成了 L6 工程师。嗯，接下来是什么？嗯，你知道，当当
[17:19] 接下来是什么？
[17:21] 嗯，你知道，当你想到模型的进展时，它只是加速了。
[17:23] Anthropic 的发布节奏已经压缩。
[17:25] Open 的发布节奏已经压缩。
[17:27] 为什么？因为这些模型通常要制造一个更好的模型，你需要一些正确的东西。
[17:28] 你需要惊人的计算能力。
[17:30] 计算非常昂贵，它有一个时间尺度，你知道我们跟踪它，就像你知道它在增长，但你知道它在接下来的，你知道的短期内，基本上是确定的。
[17:38] 这就像你已经签署的东西一样是确定的。
[17:39] 嗯，会有延迟和变化，不知何故你可以找到更多一点，但它通常是相当确定的。
[17:44] 有很棒的研究人员，人们为此支付数千万美元。
[17:47] 最后是实现。
[17:49] 历史上一直非常困难。
[17:51] 如果我现在有一个想法，我必须实现它。
[17:52] 实现是困难的。
[17:54] 现在想法已经有了。
[17:57] 实现非常容易。
[18:00] 它很昂贵，但它非常容易。
[18:02] 那么你如何决定实现哪些想法呢？
[18:05] 结果是，如果你的实现现在变得如此容易，你就可以实现更多的想法，并且在跑步机上越跑越快。
[18:10] 无论是人工智能模型研究，所以现在你的模型发布节奏已经从之前的 6 个月缩短到 2 个月了。
[18:19] 幾個月前的情況，或者嘿，我想我要把美國的每一個發電廠和每一條輸電線路都建立模型並運行迴歸分析，看看微觀的供需。
[18:26] 我也可以做到。這個想法很便宜。你知道哪個想法有意義嗎？哪個想法值得你花費代幣的資本，因為實施就在那裡。
[18:35] 這是，這是，我認為這是關鍵的學習，如果實施成本繼續下跌，而它們確實如此，嗯，我們甚至還沒有神話。
[18:46] 你知道，自 Opus 47 推出以來才幾個小時，但你知道我的團隊在內部對此感到非常興奮。
[18:50] 現在傳到世界上，呃，這將徹底重塑經濟的運作方式。
[18:57] 過去很重要的東西是執行非常非常困難，而想法很便宜。
[19:02] 現在想法便宜且豐富，但執行非常容易。
[19:07] 所以真正有價值的想法是那些能夠證明花費在超級廉價實施上的價值的想法。
[19:12] 所以你真的害怕嗎，還是你只是，它只是引入了一種難以應對的不確定性？
[19:18] 不確定性是存在的。嗯，但我確實，我確實
[19:22] 存在不確定性。
[19:22] 嗯，但我確實認為這會引起一些恐懼，關於社會如何改革自身？
[19:29] 一個人如何在一個實際上任何你，你存在於一個實際上任何你知道你實施某種能力的世界上，你的實施某種能力的實際上並不那麼重要？
[19:37] 你選擇正確想法讓 AI 實施的能力，然後你推銷該想法或推銷 AI 已實施內容的能力，這才是重要的。
[19:45] 你為此籌集資金的能力才是重要的。
[19:49] 回到重點，擁有最新的模型總是至關重要的。
[19:52] 誰能獲得最新的模型？
[19:53] Anthropic 的項目。
[19:55] 我知道它不叫 earwig，但我通過稱它為 earwig 來嘲弄 Anthropic 的人。
[19:59] 嗯，glasswig anthropic earwig，你知道，他們只向某些公司發布 mythos 以用於網絡安全。
[20:05] 這只會是會持續下去的事情。
[20:07] 模型將具有較少廣泛的部署。
[20:09] 我知道我知道 OpenAI 和 Anthropic 以及所有這些人都說，我們希望為每個人提供偉大的 AI。
[20:15] AI 非常 [ __ ] 昂貴。
[20:18] 誰將支付數萬億美元的基礎設施費用？
[20:19] 有錢並且能夠建立有用的人

The Supply and Demand of AI Tokens | Dylan Patel Interview

https://www.youtube.com/watch?v=LF3aUIM57uw
Translation: zh-TW

[00:00] What used to matter a lot was execution
过去很重要的事情是执行

[00:02] What used to matter a lot was execution was very very difficult and
过去很重要的事情是执行非常非常困难，而且

[00:04] was very very difficult and ideas were cheap.
非常非常困难，而且想法很廉价。

[00:07] ideas were cheap. Now ideas are cheap
想法很廉价。现在想法很廉价

[00:09] and plentiful but execution is very easy.
而且丰富，但执行非常容易。

[00:11] So really only the good ideas are the ones that can justify the spend on
所以真的只有好的想法才能证明在...上的花费是合理的

[00:13] the ones that can justify the spend on super cheap implementation.
那些可以证明在超级廉价的实施上的花费是合理的。

[00:30] You told me this incredible story about
你给我讲了一个关于...的不可思议的故事

[00:32] You told me this incredible story about how your own team's use of tokens has
你给我讲了一个关于你自己的团队今年使用代币的方式发生了巨大变化的不可思议的故事。

[00:35] how your own team's use of tokens has changed dramatically this year.
你自己的团队今年使用代币的方式发生了巨大变化。

[00:37] Yeah.
是的。

[00:37] Retell that story and what it is teaching you about what's going on in
重述那个故事，以及它在教你关于正在发生的什么事情

[00:39] teaching you about what's going on in the world.
在世界上正在发生的事情。

[00:41] the world.
世界上。

[00:41] Last year we thought we were heavy users
去年我们认为我们是重度用户

[00:42] Last year we thought we were heavy users of AI.
去年我们认为我们是人工智能的重度用户。

[00:45] Everyone's using chat GPT.
每个人都在使用聊天GPT。

[00:45] Everyone's using cloud.
每个人都在使用云。

[00:47] Everyone's got you know I'm providing whatever
每个人都有，你知道我提供任何

[00:48] you know I'm providing whatever subscriptions anyone wants on the order
你知道我提供任何任何人想要的订阅，大约

[00:50] subscriptions anyone wants on the order of spend of like tens of thousands of
订阅，任何人想要的，花费大约是几万美元

[00:52] of spend of like tens of thousands of dollars for our firm.
花费是几万美元，对我们公司来说。

[00:54] This year the spend has just skyrocketed and and it
今年花费只是飙升了，而且它

[00:57] spend has just skyrocketed and and it really started in late December with
花费只是飙升了，而且它真的始于十二月下旬，与

[00:59] really started in late December with Opus that included Doug who's president
真的始于十二月下旬，与包括道格在内的Opus，他是总裁

[01:01] Opus that included Doug who's president uh Doug Olaflin.
Opus，其中包括道格，他是道格·奥拉夫林的总裁。

[01:03] He's very much like uh Doug Olaflin.
他非常像道格·奥拉夫林。

[01:04] He's very much like leading the charge in the sense of like leading the charge in the sense of like non-technical people using uh AI for coding.
他非常像在引领潮流，就像非技术人员使用人工智能进行编码一样。

[01:08] non-technical people using uh AI for coding.
非技术人员使用人工智能进行编码。

[01:11] Um and so he's basically pled the whole firm slowly over time.
嗯，所以他基本上是慢慢地、随着时间的推移，说服了整个公司。

[01:13] I think he's been the the leader in doing that.
我认为他一直是这样做的领导者。

[01:14] Obviously the engineers were using it anyways but spend in January just started to inflect and rocket and rocket and rocket and rocket.
显然，工程师们无论如何都在使用它，但在 1 月份的支出刚刚开始发生变化并迅速增长。

[01:18] started to inflect and rocket and rocket and rocket and rocket.
开始发生变化并迅速增长。

[01:20] Um, we signed, and rocket and rocket.
嗯，我们签署了，并且迅速增长。

[01:23] Um, we signed, you know, an enterprise contract with Anthropic and it's gone to the point where now, um, I think when I last talked to you it was 5 million spend rate.
嗯，我们签署了，你知道，与 Anthropic 签订了一份企业合同，现在已经到了这个地步，嗯，我想我上次和你谈话时，支出率为 500 万。

[01:26] Anthropic and it's gone to the point where now, um, I think when I last talked to you it was 5 million spend rate.
Anthropic，现在已经到了这个地步，嗯，我想我上次和你谈话时，支出率为 500 万。

[01:28] where now, um, I think when I last talked to you it was 5 million spend rate.
现在，嗯，我想我上次和你谈话时，支出率为 500 万。

[01:30] talked to you it was 5 million spend rate.
和你谈话时，支出率为 500 万。

[01:31] It's actually 7 million spend right now.
现在实际上是 700 万的支出。

[01:32] right now.
现在。

[01:33] That was last week, by the way.
顺便说一句，那是上周。

[01:35] A lot of that is just the usage, right?
其中很多只是使用量，对吧？

[01:37] What's what's really, you know, people people who are have never coded before are using cloud code and spending thousands of dollars sometimes a day,
真正重要的是，你知道，那些从未写过代码的人正在使用云代码，有时一天花费数千美元，

[01:39] people who are have never coded before are using cloud code and spending thousands of dollars sometimes a day,
那些从未写过代码的人正在使用云代码，有时一天花费数千美元，

[01:41] are using cloud code and spending thousands of dollars sometimes a day,
正在使用云代码，有时一天花费数千美元，

[01:44] thousands of dollars sometimes a day, but across a firm, we're spending $7 million a year now on cloud code at the current rate.
有时一天花费数千美元，但整个公司，我们现在每年在云代码上的支出为 700 万美元，按当前费率计算。

[01:46] but across a firm, we're spending $7 million a year now on cloud code at the current rate.
但整个公司，我们现在每年在云代码上的支出为 700 万美元，按当前费率计算。

[01:49] million a year now on cloud code at the current rate.
每年数百万美元，按当前费率计算。

[01:52] um versus our salary expense being in the neighborhood of $25 million.
嗯，相比之下，我们的薪资支出约为 2500 万美元。

[01:54] expense being in the neighborhood of $25 million.
支出约为 2500 万美元。

[01:56] So, you know, we're north of 25% of spend on cloud code as a percentage of salary.
所以，你知道，我们云代码的支出占薪资的 25% 以上。

[01:59] 25% of spend on cloud code as a percentage of salary.
云代码支出占薪资的 25%。

[02:01] And if this trajectory continues, then you know,
如果这种轨迹继续下去，那么你知道，

[02:02] trajectory continues, then you know, we'll spend more than 100% by the end of
轨迹继续，那么你知道，到年底我们将花费超过100%。

[02:03] we'll spend more than 100% by the end of the year.
到年底我们将花费超过100%。

[02:03] Uh which is a bit terrifying.
呃，这有点可怕。

[02:06] the year. Uh which is a bit terrifying.
年。呃，这有点可怕。

[02:08] Thankfully, I don't have to decide
谢天谢地，我不必决定

[02:10] between people and AI because our company's growing so fast.
在人和人工智能之间，因为我们的公司发展如此之快。

[02:11] It's, you know, more so like, okay, well, I don't
它是，你知道，更像是，好吧，我不需要

[02:13] know, more so like, okay, well, I don't have to hire nearly as fast and I can
知道，更像是，好吧，我不需要像以前那样快地招聘，而且我可以

[02:15] have to hire nearly as fast and I can spend a lot more on AI and it works and
需要像以前那样快地招聘，而且我可以花更多的钱在人工智能上，而且它奏效了，而且

[02:16] spend a lot more on AI and it works and we just grow faster.
花更多的钱在人工智能上，而且它奏效了，而且我们只是增长得更快。

[02:18] But I think other folks will start to reckon with the fact
但我想其他人会开始认识到这个事实

[02:20] folks will start to reckon with the fact that, huh, if this person can do the
人们会开始认识到这个事实，嗯，如果这个人能做

[02:23] that, huh, if this person can do the work of five to 10 to 15 people uh using
那个，嗯，如果这个人能做五个到十个到十五个人的工作，呃，使用

[02:26] work of five to 10 to 15 people uh using cloud code, then all of a sudden I
五个到十个到十五个人的工作，呃，使用云代码，那么突然之间我

[02:29] cloud code, then all of a sudden I should probably cut people.
云代码，那么突然之间我可能应该裁掉一些人。

[02:30] should probably cut people. But right now, I think the use cases are so broad.
可能应该裁掉一些人。但现在，我认为用例非常广泛。

[02:33] now, I think the use cases are so broad. For example, one thing is we have a
现在，我认为用例非常广泛。例如，有一件事是我们有一个

[02:35] For example, one thing is we have a reverse engineering lab in Oregon that
例如，有一件事是我们有一个在俄勒冈州的逆向工程实验室，它

[02:36] reverse engineering lab in Oregon that we've been building for a year and a
在俄勒冈州的逆向工程实验室，我们已经建造了一年半了。

[02:37] we've been building for a year and a half. We have a bunch of, you know,
我们已经建造了一年半了。我们有一堆，你知道，

[02:39] half. We have a bunch of, you know, fancy microscopes, scanning electron
半。我们有一堆，你知道，花哨的显微镜，扫描电子

[02:41] fancy microscopes, scanning electron microscopes. The whole purpose of this
花哨的显微镜，扫描电子显微镜。这个的整个目的是

[02:42] microscopes. The whole purpose of this is you reverse engineer chips. You get
显微镜。这个的整个目的是你逆向工程芯片。你得到

[02:44] is you reverse engineer chips. You get uh the architecture out of it. you get
是你逆向工程芯片。你从中获得呃架构。你得到

[02:46] uh the architecture out of it. you get the materials that they're using to
呃从中获得架构。你得到他们正在使用的材料来

[02:47] the materials that they're using to manufacture and this is some of the data
他们正在使用的材料来制造，这是我们销售的一些数据。

[02:48] manufacture and this is some of the data we sell. This is a very slow process of
制造，这是我们销售的一些数据。这是一个非常缓慢的分析数据的过程。

[02:50] we sell. This is a very slow process of analyzing that data. Instead, um one
我们销售。这是一个非常缓慢的分析数据的过程。相反，嗯，一个人

[02:53] analyzing that data. Instead, um one person on the team, they've been able to
分析这些数据。相反，嗯，团队中的一个人，他们已经能够

[02:55] person on the team, they've been able to spend with a couple thousand dollars of
团队中的一个人，他们能够花费几千美元的

[02:56] spend with a couple thousand dollars of cloud tokens. They've been able to
花费几千美元的云令牌。他们已经能够

[02:57] cloud tokens. They've been able to create this application that is GPU
云令牌。他们已经能够创建这个应用程序，它是 GPU 加速的

[02:59] create this application that is GPU accelerated runs on a server that we
创建这个应用程序，它是 GPU 加速的，运行在我们拥有的服务器上

[03:02] accelerated runs on a server that we have at Coreweave and anytime we send it
加速的，运行在我们拥有的 Coreweave 服务器上，并且任何时候我们发送它

[03:04] have at Coreweave and anytime we send it an image, it's able to take the picture
在 Coreweave 有，并且任何时候我们发送一张图片，它都能拍下这张照片

[03:05] an image, it's able to take the picture of the chip and overlay where every
一张图片，它都能拍下芯片的照片并覆盖每一个

[03:07] of the chip and overlay where every single material is.
芯片，并覆盖每一种材料所在的位置。

[03:09] single material is. Oh, this part is copper.
材料是。哦，这部分是铜。

[03:11] copper. Oh, this part of the gate is uh tantelum.
铜。哦，这个栅极的部分是呃钽。

[03:13] tantelum. This part of the gate is germanium.
钽。这个栅极的部分是锗。

[03:14] germanium. This part of the gate is cobalt.
锗。这个栅极的部分是钴。

[03:15] cobalt. And so you can do a finite element analysis of the entire stackup
钴。所以你可以对整个堆叠进行有限元分析

[03:17] element analysis of the entire stackup of the chip very very quickly visual
对芯片的整个堆叠进行非常非常快速的视觉分析

[03:20] of the chip very very quickly visual with a dashboard guey it's everything
芯片的非常非常快速的视觉分析，带有一个仪表板界面，它是一切

[03:22] with a dashboard guey it's everything few thousand dollars would took claude
带有一个仪表板界面，它是一切，几千美元就能做到克劳德

[03:23] few thousand dollars would took claude the person previously worked at Intel
几千美元就能做到克劳德，这个人以前在英特尔工作

[03:25] the person previously worked at Intel and he said that was an entire team's
这个人以前在英特尔工作，他说那是一个整个团队的

[03:27] and he said that was an entire team's job to build that and maintain that now
他说那是一个整个团队的工作来构建和维护它，现在

[03:29] job to build that and maintain that now rack that up across you know the entire
工作来构建和维护它，现在将其推广到你知道的整个

[03:31] rack that up across you know the entire firm it's it's insane another example
推广到你知道的整个公司，这太疯狂了，另一个例子

[03:33] firm it's it's insane another example that I think is super fun is Malcolm
公司，这太疯狂了，另一个我认为非常有趣的例子是马尔科姆

[03:36] that I think is super fun is Malcolm who's an economist at a major bank
我认为非常有趣的例子是马尔科姆，他是一家主要银行的经济学家

[03:38] who's an economist at a major bank before um their economist department was
他是一家主要银行的经济学家，在此之前，呃，他们的经济学部门是

[03:41] before um their economist department was like 100 or 200 people what he built was
在此之前，呃，他们的经济学部门有大约 100 或 200 人，他构建了

[03:44] like 100 or 200 people what he built was the most incredible thing ever.
大约 100 或 200 人，他构建了有史以来最令人难以置信的东西。

[03:46] the most incredible thing ever. He piped all of this different data, you know,
有史以来最令人难以置信的东西。他接入了所有这些不同的数据，你知道，

[03:48] all of this different data, you know, FRED data and all these other data,
所有这些不同的数据，你知道，FRED 数据和所有这些其他数据，

[03:49] FRED data and all these other data, right? Employment reports and all these
FRED 数据和所有这些其他数据，对吧？就业报告和所有这些

[03:51] right? Employment reports and all these other things from various APIs. We
对吧？就业报告和所有这些来自各种 API 的其他东西。我们

[03:53] other things from various APIs. We signed a couple contracts with folks to
来自各种 API 的其他东西。我们与一些人签订了几份合同，以

[03:54] signed a couple contracts with folks to get API access to data. Pulled it all
签订了几份合同，以获取数据 API 访问权限。将所有数据拉取进来，

[03:56] get API access to data. Pulled it all in, started running regression, started
获取数据 API 访问权限。将所有数据拉取进来，开始运行回归分析，开始

[03:58] in, started running regression, started looking at the impact of various
进来，开始运行回归分析，开始查看各种

[04:00] looking at the impact of various economic revolutions on the economy um
查看各种经济革命对经济的影响，呃

[04:03] economic revolutions on the economy um from a deflationary inflationary
经济革命对经济的影响，呃，从通缩通胀的角度

[04:05] from a deflationary inflationary perspective.
從通縮和通貨膨脹的角度來看。

[04:08] The BLS has this entire um Bureau of Labor Statistics has this entire like set of like 2,000 tasks.
勞工統計局擁有這整個，嗯，勞工統計局擁有這整個像是約 2,000 項任務的集合。

[04:12] And so he did that with AI, which ones can be done by AI, which ones cannot, and grading them across a rubric.
所以他用 AI 做了這件事，哪些可以由 AI 完成，哪些不行，並根據標準評分。

[04:17] You know, about 3% are doable now with AI.
你知道，大約有 3% 現在可以用 AI 完成。

[04:20] Um, and so he's created this like metric so that you can measure things that can be done by AI, what what the massive deflationary uh, you know, what the cost of being able to do those with AI and therefore the deflationary aspect of it.
嗯，所以他創建了這個指標，以便你可以衡量那些可以用 AI 完成的事情，以及大規模通縮的呃，你知道，用 AI 完成這些事情的成本，以及因此帶來的通縮方面。

[04:31] You know, output can go up.
你知道，產出可以增加。

[04:32] It's called phantom GDP is what he called it.
這就是他所說的「幻象 GDP」。

[04:33] Phantom GDP.
幻象 GDP。

[04:35] Output can go up, but because cost falls so much, actually GDP theoretically shrinks.
產出可以增加，但由於成本大幅下降，實際上 GDP 理論上會萎縮。

[04:39] So he created this whole analysis and a brand new benchmark of uh language models um a set of evals across 2,000 different evals.
所以他創建了這個完整的分析和一個全新的呃語言模型基準，嗯，一套跨越 2,000 個不同評估的評估。

[04:46] Right.
對。

[04:47] This all by himself.
這一切都是他一個人完成的。

[04:47] This is all by himself.
這一切都是他一個人完成的。

[04:49] Yeah.
是的。

[04:49] And he's like dude this would have taken the team of 200 economist a year.
他就像是，老兄，這需要 200 名經濟學家團隊一年的時間。

[04:53] He's just like he's like completely cracked out on claude.
他就像是，他就像是完全沉迷於 Claude。

[04:54] He's like everything has changed.
他就像是，一切都改變了。

[04:56] How do you think about as a business owner going from close to zero to 25% accelerating towards whatever percent of total spend?
作為一個企業主，你如何看待從接近零到 25% 的加速增長，朝著總支出的任何百分比發展？

[05:01] Like at what point are you like, whoa, I need to put the brakes on
就像在什麼時候你會覺得，哇，我需要踩剎車

[05:06] like, whoa, I need to put the brakes on this and be careful how much we're spending.
就像，哇，我需要刹住这一点，并小心我们花了多少钱。

[05:10] Maybe we don't need to spend on the most cutting it on Opus 4.7, which came out today.
也许我们不需要花钱在 Opus 4.7 上，它今天出来了。

[05:13] Maybe I can throttle it back to something that's a little bit cheaper.
也许我可以把它降到更便宜的东西。

[05:17] Ultimately, like I'm in the information business, right?
最终，就像我在信息行业一样，对吧？

[05:18] That that is, you know, we sell analysis, we sell, we do consulting, we create data sets.
那就是，你知道，我们销售分析，我们销售，我们做咨询，我们创建数据集。

[05:22] I don't see why this wouldn't be completely commoditized on a pretty rapid basis if I'm not constantly improving.
我不知道为什么这不会在相当快的速度下完全商品化，如果我不断改进的话。

[05:29] my first product that I was selling as a data set actually it is you know like there's more people trying to do it now.
我卖的第一个产品是数据集，实际上你知道现在有更多的人在尝试做这个。

[05:33] we've made it constantly better and better and better and more detailed and so therefore it sells a market.
我们不断地把它做得越来越好，越来越详细，因此它会卖得很好。

[05:39] but the way we were doing it in 2023 is not terribly different than you know is it's it's basically what everyone else is doing now.
但我们在 2023 年的做法与你知道的没有什么太大区别，基本上是现在其他人都在做的事情。

[05:45] if I don't move up the bar then I will be commoditized.
如果我不提高标准，我就会被商品化。

[05:50] if I don't move fast enough I will also lose my edge so the question is yes AI commoditizes things.
如果我行动不够快，我也会失去优势，所以问题是，是的，人工智能会商品化事物。

[05:55] just like it commoditizes software those who can move fast and keep control of their customers and keep providing them an awesome service and keep improving the service won't shrink.
就像它商品化软件一样，那些能够快速行动并控制客户并继续为他们提供优质服务并不断改进服务的人不会萎缩。

[06:05] They'll grow. They'll grow faster. Those
他们会成长。他们会成长得更快。那些

[06:07] They'll grow.
他們會成長。

[06:07] They'll grow faster.
他們會成長得更快。

[06:08] Those who are incumbent and not doing anything, they're going to lose.
那些現任者卻無所作為的人，他們將會輸掉。

[06:10] And so, it's a bit of an existential like if I don't adopt AI, someone else will and they will beat me.
所以，這有點像一種生存危機，如果我不採用人工智能，別人就會採用，他們就會打敗我。

[06:13] Uh, another easy example is the energy space.
呃，另一個簡單的例子是能源領域。

[06:16] So, we've had a few energy analysts for a couple for like a year now.
所以，我們已經有幾位能源分析師工作了一年左右了。

[06:18] We've been trying to build out this energy model.
我們一直在努力建立這個能源模型。

[06:20] It's very complex.
它非常複雜。

[06:21] Energy's data services market is something like $900 million.
能源數據服務市場大約有九億美元。

[06:23] So obviously a huge market for me to try and break into but it has you know we really hadn't broken into the energy data services business despite a year of having multiple people on the team.
所以顯然這是一個巨大的市場，我試圖進入，但你知道，儘管團隊中有幾個人工作了一年，我們實際上並沒有進入能源數據服務業務。

[06:25] Um then cloud code psychosis hits one of the people who leads the data center energy and industrial sort of business at semi analysis uh Jeremy hits him and now all of a sudden in 3 weeks um he spent a lot he was spending like $6,000 a day.
嗯，然後雲代碼精神病擊中了半分析公司中負責數據中心能源和工業業務的傑里，現在突然在三週內，他花了很多錢，他每天花費約六千美元。

[06:27] It was an insane amount but he scraped every single power plant in the US every single transmission line above a certain voltage.
這是一筆瘋狂的數額，但他抓取了美國的每一個發電廠，以及每一條高於一定電壓的輸電線路。

[06:29] um and created this entire mapping of the entire US grid as well as a lot of demand sources all from various public sources of data.
嗯，並創建了整個美國電網的完整地圖，以及許多來自各種公共數據來源的需求來源。

[06:30] Um and we've shown it to and and we built and it's got like this dashboard where you
嗯，我們已經向...展示了，而且我們已經構建了，它有一個儀表板，在那裡你

[07:08] it's got like this dashboard where you can view and check you can see all the micro regions of the US where there's power deficits and surpluses.
它有一个仪表板，您可以在其中查看和检查，可以看到美国所有存在电力赤字和盈余的微观区域。

[07:14] Um all of these details built in a handful of weeks we started showing some of our customers who buy our data center data set but are energy like traders.
嗯，所有这些细节都在几周内完成，我们开始向一些购买我们数据中心数据集但从事能源交易的客户展示。

[07:19] We showed some of them and they're like wow how long did this take you?
我们向他们展示了一些，他们说哇，这花了你多长时间？

[07:25] This is really good. this is better than XYZ company and then we like get dig deeper.
这真的很好。这比 XYZ 公司好，然后我们深入挖掘。

[07:29] XYZ company has 100 people and have been working on this for a decade now.
XYZ 公司有 100 人，并且已经为此工作了十年。

[07:32] Obviously our thing is not fully robust as robust but in some ways it is better.
显然，我们的东西并不完全健壮，但有些方面它更好。

[07:36] I'm going to commoditize these energy services companies, data services company.
我将把这些能源服务公司、数据服务公司商品化。

[07:39] Who's going to come commoditize me if I don't move faster?
如果我不加快速度，谁会来商品化我？

[07:40] And so the question from a business owner's perspective is yeah I'm spending a lot but what does that spend getting me?
所以从企业主的角度来看，问题是是的，我花了很多钱，但这些花费能给我带来什么？

[07:46] Is it getting more revenue? Yeah.
它能带来更多收入吗？是的。

[07:50] >> Most software companies try to maximize your time on their app to juice engagement.
>>大多数软件公司试图最大化您在他们应用程序上的时间以提高参与度。

[07:53] RAMP does the exact opposite.
RAMP 则完全相反。

[07:55] RAMP understands that no one wants to spend hours filing expense reports, reviewing expense reports, and checking for policy violations.
RAMP 明白没有人想花几个小时填写费用报告、审查费用报告和检查政策违规行为。

[08:01] So, they built their tools to give that time back, using AI to automate 85% of expense reviews with 99% accuracy.
因此，他们构建了工具来将这些时间还给用户，利用人工智能以 99% 的准确率自动处理 85% 的费用审查。

[08:04] And
和

[08:10] expense reviews with 99% accuracy.
費用審核準確率高達 99%。

[08:12] And since RAMP saves companies 5%, it's no wonder that Shopify runs on RAM, Stripe runs on RAM, and my business does, too.
而且由於 RAMP 為公司節省了 5% 的費用，難怪 Shopify 使用 RAM，Stripe 使用 RAM，我的公司也是如此。

[08:18] To see what happens when you eliminate the busy work, check out ramp.com/invest.
想看看消除繁瑣工作後會發生什麼，請造訪 ramp.com/invest。

[08:22] OpenAI, Cursor, Enthropic, Perplexity, and Verscell all have something in common.
OpenAI、Cursor、Enthropic、Perplexity 和 Verscell 都有一個共同點。

[08:26] They all use Work OS.
它們都使用 Work OS。

[08:28] And here's why.
原因如下。

[08:30] To achieve enterprise adoption at scale, you have to deliver on core capabilities like SSO, SKIM, Arbback, and audit logs.
要實現大規模的企業採用，您必須提供 SSO、SKIM、Arbback 和審計日誌等核心功能。

[08:37] That's where work OS comes in.
這就是 Work OS 的用武之地。

[08:38] Instead of spending months building these missionritical capabilities yourself, you can just use work OS APIs to gain all of them on day zero.
與其花費數月時間自行建構這些關鍵任務功能，不如使用 Work OS API 在第一天就獲得所有這些功能。

[08:46] That's why so many of the top AI teams you hear about already run on work OS.
這就是為什麼您聽說過的許多頂級 AI 團隊都已經在使用 Work OS。

[08:51] Work OS is the fastest way to become enterprise ready and stay focused on what matters most, your product.
Work OS 是做好企業準備並專注於最重要事項（您的產品）的最快方法。

[08:54] Visit works.com to get started.
請造訪 works.com 開始使用。

[08:58] Felix by Rogo is a personal finance agent that turns a single prompt into finished clientready work using your firm's own templates, context, and standards.
Felix by Rogo 是一款個人財務代理，可將單一提示轉換為使用您公司自己的範本、內容和標準完成的客戶就緒工作。

[09:06] Send Felix an email like, "Take these comments and turn them for me, or update my tracker with the context of these emails, or run
向 Felix 發送一封電子郵件，例如：「將這些評論為我轉換，或根據這些電子郵件的內容更新我的追蹤器，或執行

[09:12] with the context of these emails, or run the ability to pay math on this buyer.
在这些电子邮件的背景下，或者对这位买家进行支付能力计算。

[09:14] the ability to pay math on this buyer. And Felix sends back finished PowerPoint
对这位买家进行支付能力计算。然后 Felix 发回了完成的 PowerPoint

[09:16] And Felix sends back finished PowerPoint decks, Excel models, and sourced
然后 Felix 发回了完成的 PowerPoint 演示文稿、Excel 模型和搜集到的

[09:18] decks, Excel models, and sourced research. Felix works the way your team
演示文稿、Excel 模型和搜集到的研究。Felix 的工作方式与您的团队

[09:20] research. Felix works the way your team already does, delivering work quickly
研究一样。Felix 的工作方式与您的团队已有的方式相同，能够快速地

[09:21] already does, delivering work quickly and accurately around the clock. Learn
并且准确地全天候交付工作。了解

[09:24] and accurately around the clock. Learn more at rogo.ai/felix.
并且准确地全天候交付工作。在 rogo.ai/felix 了解更多信息。

[09:27] more at rogo.ai/felix. Are you worried that in the limit the
在 rogo.ai/felix 了解更多信息。您是否担心最终控制资本和

[09:29] Are you worried that in the limit the people that control capital and
您是否担心最终控制资本和投资资本的人，他们经常因为您所做的事情而雇佣您，

[09:30] people that control capital and investing capital who are often hiring
投资资本的人，他们经常因为您所做的事情而雇佣您，

[09:32] you for for what you do will just say,
就会说，

[09:35] you for for what you do will just say, "Well, we have analysts too who are
就会说，“嗯，我们也有分析师，他们在这方面非常聪明。比如，我们

[09:36] "Well, we have analysts too who are really smart about this. Like, we'll
“嗯，我们也有分析师，他们在这方面非常聪明。比如，我们

[09:37] really smart about this. Like, we'll just build this ourselves." Like if it's
就会自己来构建这个。”就像如果

[09:39] just build this ourselves." Like if it's getting that easy, at what point does it
变得如此容易，那么在什么时间点，所有的一切都会

[09:41] getting that easy, at what point does it just all pull into the investment firms
都会拉入那些最能从中获利的投资公司，

[09:44] just all pull into the investment firms that stand to gain the most because they
因为它们拥有对数据或它们所提取的见解的最大影响力，

[09:45] that stand to gain the most because they have the most leverage on top of the
因为它们拥有对数据或它们所提取的见解的最大影响力，

[09:47] have the most leverage on top of the data or the insights that that they
拥有对数据或它们所提取的见解的最大影响力，

[09:49] data or the insights that that they glean?
它们所提取的见解？

[09:49] glean? >> First of all, any information services
提取？>> 首先，任何信息服务

[09:51] >> First of all, any information services business, obviously I don't generate as
>> 首先，任何信息服务业务，显然我从客户那里获得的信息价值不如客户自己从信息中获得的价值高。

[09:54] business, obviously I don't generate as much value as my customer does from such
业务，显然我从客户那里获得的信息价值不如客户自己从信息中获得的价值高。

[09:55] much value as my customer does from such information. Uh because if I sell you
信息。嗯，因为如果我卖给你

[09:57] information. Uh because if I sell you information for a dollar, you're only
信息。嗯，因为如果我卖给你信息，价格是一美元，你只

[09:59] information for a dollar, you're only buying it for a dollar because you know
信息，价格是一美元，你只买它是因为你知道

[10:01] buying it for a dollar because you know that information helps you make a
那信息能帮助你做出一个决定，让你赚取的利润超过

[10:02] that information helps you make a decision that lets you make more than
那信息能帮助你做出一个决定，让你赚取的利润超过

[10:04] decision that lets you make more than $1. And so therefore, you have you have
一美元。因此，你从我这里赚取的利润比我从信息本身获得的利润要多。

[10:06] $1. And so therefore, you have you have arbit you you have made more money off
一美元。因此，你从我这里赚取的利润比我从信息本身获得的利润要多。

[10:08] arbit you you have made more money off of me than I did from the information
你从我这里赚取的利润比我从信息本身获得的利润要多。

[10:09] of me than I did from the information myself. An investment fund, these
信息本身。投资基金，这些

[10:11] myself. An investment fund, these investment funds all have their own
投资基金，这些投资基金都有自己的

[10:13] Investment funds all have their own information services, you know.
投資基金都有自己的信息服務，你知道的。

[10:14] Information services, you know, especially like the super like the Jane Streets of the world and the Citadels.
信息服務，你知道的，尤其是像像簡街這樣的公司和像 Citadel 這樣的公司。

[10:17] Streets of the world and the Citadels.
世界各地的街道和 Citadel。

[10:19] They're they're really detailed on their data.
他們對他們數據的細節非常詳盡。

[10:21] And yet, um, these sort of folks also purchase data from us and continue to do so and continue to grow with us.
然而，嗯，這類人也從我們這裡購買數據，並繼續這樣做，並與我們一起成長。

[10:25] Because I think there's just some some it factor, right?
因為我認為這就是一些一些關鍵因素，對吧？

[10:27] We move faster, we're more nimble, we're a smaller team that's focused on just one specific thing.
我們行動更快，我們更靈活，我們是一個更小的團隊，專注於一件特定的事情。

[10:34] uh AI infrastructure and and the huge revolution that causes in AI um and tokconomics and all these things and and we sort of really see where it's headed.
呃人工智能基礎設施以及它在人工智能中引起的巨大革命，嗯，以及代幣經濟學和所有這些事情，我們 sort of 真的看到了它的發展方向。

[10:43] And so we're moving faster and building faster.
所以我們行動更快，構建也更快。

[10:45] Um I think investment professionals just would you know yes they'll try and build some of the stuff we do and um more likely they'll just buy the data from us and it's cheaper for them to buy the data from us and then to build and then build on top of it than it is to build it themselves.
嗯，我認為投資專業人士只是，你知道的，是的，他們會嘗試構建我們做的一些東西，嗯，更有可能的是，他們會直接從我們這裡購買數據，對他們來說，從我們這裡購買數據然後構建，然後在其之上構建比他們自己構建更便宜。

[11:01] But ultimately some may try.
但最終有些人可能會嘗試。

[11:02] I feel like every conversation I have with you, what I'm always getting at is just supply and demand of tokens like that's the thing that's interesting to me in the world right now.
我覺得我每次和你談話，我總是在談論代幣的供需，就像這是我現在在這個世界上感興趣的事情一樣。

[11:09] What has this experience taught you about the demand?
這次經歷讓你對需求學到了什麼？

[11:11] Has it
它是否

[11:13] Taught you about the demand?
關於需求，我教過你嗎？

[11:14] Has it changed your view on the demand side of that equation?
這是否改變了你對該方程式中需求方面的看法？

[11:16] Just feeling it viscerally yourself.
你自己親身感受一下。

[11:17] If we take a step back and look at the macro lens, right?
如果我們退一步，從宏觀角度來看，對吧？

[11:20] Enthropic has gone from 9 billion revenue to what they're at 3540 billion now.
Enthropic 的收入從 90 億美元增長到現在的 3540 億美元。

[11:24] Probably by the time this airs 40 45 billion, who does ARR?
可能等到這個節目播出時，400 億到 450 億美元，誰能達到 ARR？

[11:30] Their compute has not grown to the same degree.
他們的計算能力沒有達到相同的程度。

[11:32] Um, and if you do the calculations and you assume they didn't decrease their research and development compute, they clearly didn't.
嗯，如果你做計算並假設他們沒有減少研發計算，他們顯然沒有。

[11:35] Their release, they have Mythos, they have up is 4.7.
他們的發布，他們有 Mythos，他們有 4.7。

[11:40] So they clearly didn't decrease their research compute spend.
所以他們顯然沒有減少研究計算支出。

[11:43] Um, so ultimately what they've done, even if you assume all incremental compute they've gotten has gone towards inference, their margins are at a floor of 72%.
嗯，所以最終他們所做的，即使你假設他們獲得的所有增量計算都用於推理，他們的利潤率也處於 72% 的地板價。

[11:51] In reality, some of that incremental compute they've got probably went to research and development.
實際上，他們獲得的一些增量計算可能用於研發。

[11:54] It may be higher than 72% gross margins.
毛利率可能高於 72%。

[11:57] To be clear, at the start of the year, they started uh there was um there was a leak by someone from their funding some some of their funding round docs.
需要說明的是，在年初，他們開始時，嗯，有人從他們的融資文件中洩露了一些信息。

[12:03] Someone leaked it 30 something% gross margins.
有人洩露了 30 多個百分點的毛利率。

[12:08] Where on earth does a business like this grow margins like that?
像這樣的企業，利潤率到底是如何增長的？

[12:10] And it's in principle, right?
而且這是原則上的，對吧？

[12:11] Their demand is so high.
他們的市場需求如此之高。

[12:11] They're able to cut back on usage
他們能夠削減使用量

[12:13] high.
高。

[12:13] They're able to cut back on usage limits, rate limits, all these things.
他們能夠削減使用限制、速率限制，所有這些東西。

[12:16] limits, rate limits, all these things.
限制、速率限制，所有這些東西。

[12:18] Um, what really matters is having an anthropic rep and having an enterprise
嗯，真正重要的是要有一個 Anthropic 代表，並有一個企業級

[12:19] anthropic rep and having an enterprise contract with them and getting the rate
Anthropic 代表並與他們簽訂企業合約，並獲得速率

[12:21] contract with them and getting the rate limit increases that you need because
合約，並獲得您需要的速率限制增加，因為

[12:23] limit increases that you need because otherwise tokens are ultimately super
限制增加，因為否則 token 最終會非常

[12:25] otherwise tokens are ultimately super super in demand.
否則 token 會非常非常搶手。

[12:28] super in demand. Whoever whoever can pay for them anthropic has the same problem,
搶手。誰能為他們付費，Anthropic 也有同樣的問題，

[12:29] for them anthropic has the same problem, right?
為他們付費，Anthropic 也有同樣的問題，對吧？

[12:31] right? Like I mean not problem, it's it's just the reality of how capitalism
對吧？我的意思是，不是問題，這只是資本主義運作的現實。

[12:32] it's just the reality of how capitalism works.
這只是資本主義運作的現實。

[12:35] works. Yes, people are spending sending them $40 billion AR in tokens and but
運作。是的，人們正在花費 400 億美元購買 token，但是

[12:38] them $40 billion AR in tokens and but those tokens are generating way more
他們 400 億美元的 token，但是這些 token 產生的價值遠遠超過

[12:39] those tokens are generating way more than $40 billion in value.
400 億美元。各種

[12:42] than $40 billion in value. Various businesses will have different value
價值。各種企業每 token 的價值產生會有所不同。

[12:44] businesses will have different value generation per token. But as we get more
但隨著我們變得越來越

[12:46] generation per token. But as we get more and more intelligent, what really
智能，真正重要的是獲得這些最智能的 token 並加以利用。

[12:48] and more intelligent, what really matters is access to these most
智能，真正重要的是獲得這些最智能的 token 並加以利用。

[12:49] matters is access to these most intelligent tokens and leveraging them
智能的 token 並加以利用。

[12:51] intelligent tokens and leveraging them at things. You as a person deciding what
事物。你作為一個人，決定什麼是利用這些 token 來發展業務和創造價值的最佳方式

[12:54] at things. You as a person deciding what is the best way to leverage these tokens
事物。你作為一個人，決定什麼是利用這些 token 來發展業務和創造價值的最佳方式

[12:56] is the best way to leverage these tokens to grow business and generate value
來發展業務和創造價值，因為很多人會想要 token

[12:58] to grow business and generate value because a lot of folks will want tokens
並產生 token。嗯，但那個糟糕的 SAS 新創公司，在舊金山，使用 Claude 來生成，你知道，

[13:00] and generate tokens. Uh but the shitty SAS startup and and and and SF who is
並產生 token。嗯，但那個糟糕的 SAS 新創公司，在舊金山，使用 Claude 來生成，你知道，

[13:02] SAS startup and and and and SF who is using Claude to generate, you know,
SAS 新創公司，在舊金山，使用 Claude 來生成，你知道，他們的軟體產品不一定真的創造了很多價值，因此他們很快就會被 token 定價。

[13:06] using Claude to generate, you know, their software product is not
你知道，他們的軟體產品不一定真的創造了很多價值，因此他們很快就會被 token 定價。

[13:08] their software product is not necessarily actually creating a ton of
他們的軟體產品不一定真的創造了很多價值，因此他們很快就會被 token 定價。

[13:09] necessarily actually creating a ton of value and therefore they're going to get
價值，因此他們很快就會被 token 定價。

[13:11] value and therefore they're going to get priced out of tokens uh soon enough.
價值，因此他們很快就會被 token 定價。

[13:15] priced out of tokens uh soon enough.
很快就买不起代币了。

[13:16] Are you at all surprised that I I had this experience just today where on the flight here I got rate limited out on something I saw 4.7 came out and what I immediately wanted was like to be on 4.7 that second and I was it just I couldn't think about using 4.6 anymore.
你对我今天遇到的这种情况一点也不感到惊讶吗？在这里的航班上，我因为看到了 4.7 的发布而被限制了，而我立刻就想在下一秒就用上 4.7，但我就是无法再考虑使用 4.6 了。

[13:29] or not.
或者不。

[13:31] This 47 is out.
这个 47 已经出来了。

[13:33] I was perfectly happy with 4.6 for the last many weeks.
在过去几周里，我一直对 4.6 非常满意。

[13:35] It's amazing.
太棒了。

[13:37] Are you surprised that people are so insistent on going to the most expensive leading edge thing to the degree they are?
你是否惊讶于人们如此坚持要使用最昂贵的前沿技术，以至于达到了他们所达到的程度？

[13:42] Without a doubt.
毫无疑问。

[13:44] One of my funniest memories in the past month and a half is myself and a buddy of mine, Leopold, being on our knees in front of an anthropic co-founder begging him for access to Methos and then pretending it doesn't exist cuz we knew it existed.
过去一个半月里我最有趣的回忆之一就是我和我的朋友利奥波德，双膝跪在一位 Anthropic 的联合创始人面前，恳求他给我们 Methos 的访问权限，然后假装它不存在，因为我们知道它存在。

[14:00] were like, "Please give us access."
我们就像说：“请给我们访问权限。”

[14:02] And he's like, "I don't know what you're talking about."
而他说：“我不知道你在说什么。”

[14:04] What was your reaction to that rate card or that eval card coming out?
你对那个费率卡或那个评估卡出来有什么反应？

[14:08] It was rumored in the Bay Area.
这在湾区有传言。

[14:09] Everyone, you know, we sort of like knew it was supposed to be really good, but um if you just look at the benchmarks and obviously benchmarks change over
你知道，大家都差不多知道它应该会非常好，但是，嗯，如果你只看基准测试，而且显然基准测试会随着时间而变化

[14:15] and obviously benchmarks change over time, Mythos is potentially the biggest
顯然，基準會隨著時間而改變，Mythos 可能是最大的

[14:18] time, Mythos is potentially the biggest step up in model capabilities in like 2
時間，Mythos 可能是模型功能方面最大的進步，在過去的兩年裡

[14:21] step up in model capabilities in like 2 years. I think that's really really an
兩年裡模型功能方面的進步。我認為這是一個非常非常重要的細節，你知道

[14:23] years. I think that's really really an an important detail that you know it
你知道，它太好了，以至於他們甚至不想發布它，儘管他們已經宣布了價格

[14:25] an important detail that you know it it's so good that they're like don't
對於他們的員工，他們為 Cyber​​ 進行了選擇性發布，而且它的代幣成本是 5 到 10 倍。

[14:27] it's so good that they're like don't want to release it even though they're
他們只是不想發布它，因為他們擔心它對世界的影響，他們正在向我們發布一個糟糕的、更糟糕的開源 47 版本，

[14:28] want to release it even though they're they they already announced the price to
他們明確表示，在模型卡中，我們實際上優先使其在網絡安全方面變得更糟。我不知道你是否讀過。

[14:31] their people that they did a selective release for cyber for and it's like five
無論你是誰，如果你有足夠的資本，你應該獲得一個該死的企業雲或企業 Anthropic 訂閱，你按代幣付費，

[14:33] their people that they did a selective release for cyber for and it's like five or 10x the token cost. They just don't
而不是像這些訂閱一樣，因為這樣你就不會受到 كثيرا的速率限制。然後你必須弄清楚如何利用這些代幣來完成最高價值的任務，並從中獲利，因為最終你可能在一年或兩年後

[14:34] or 10x the token cost. They just don't want to release it um because they're
從現在開始，這項業務實際上只是在套利代幣，對嗎？代幣很棒，但讓我們弄清楚將它們指向哪個方向，然後三四年

[14:36] want to release it um because they're worried about the like impact on the
他們只是不想發布它，因為他們擔心它對世界的影響，他們正在向我們發布一個糟糕的、更糟糕的開源 47 版本，

[14:38] worried about the like impact on the world and they're releasing a shitty
他們擔心它對世界的影響，他們正在向我們發布一個糟糕的、更糟糕的開源 47 版本，

[14:39] world and they're releasing a shitty worse version of open 47 to us and they
他們明確表示，在模型卡中，我們實際上優先使其在網絡安全方面變得更糟。我不知道你是否讀過。

[14:42] worse version of open 47 to us and they explicitly said in the model card hey we
他們明確表示，在模型卡中，我們實際上優先使其在網絡安全方面變得更糟。我不知道你是否讀過。

[14:45] explicitly said in the model card hey we actually preferentially made it worse at
無論你是誰，如果你有足夠的資本，你應該獲得一個該死的企業雲或企業 Anthropic 訂閱，你按代幣付費，

[14:47] actually preferentially made it worse at cyber. I don't know if you read that.
實際上優先使其在網絡安全方面變得更糟。我不知道你是否讀過。

[14:49] cyber. I don't know if you read that. whoever you are, if you have enough
無論你是誰，如果你有足夠的資本，你應該獲得一個該死的企業雲或企業 Anthropic 訂閱，你按代幣付費，

[14:50] whoever you are, if you have enough capital, you should get a freaking
無論你是誰，如果你有足夠的資本，你應該獲得一個該死的企業雲或企業 Anthropic 訂閱，你按代幣付費，

[14:52] capital, you should get a freaking enterprise cloud uh enterprise anthropic
企業雲或企業 Anthropic 訂閱，你按代幣付費，

[14:54] enterprise cloud uh enterprise anthropic subscription where you pay per token,
企業 Anthropic 訂閱，你按代幣付費，而不是像這些訂閱一樣，因為這樣你就不會受到 كثيرا的速率限制。

[14:56] subscription where you pay per token, not with these like subscriptions
你按代幣付費，而不是像這些訂閱一樣，因為這樣你就不會受到 كثيرا的速率限制。

[14:58] not with these like subscriptions because then you won't get rate limited
而不是像這些訂閱一樣，因為這樣你就不會受到 كثيرا的速率限制。

[14:59] because then you won't get rate limited much. And then you must you need to
因為這樣你就不會受到 كثيرا的速率限制。然後你必須弄清楚如何利用這些代幣來完成最高價值的任務，並從中獲利，因為最終你可能在一年或兩年後

[15:01] much. And then you must you need to figure out how to leverage those tokens
然後你必須弄清楚如何利用這些代幣來完成最高價值的任務，並從中獲利，因為最終你可能在一年或兩年後

[15:02] figure out how to leverage those tokens to the highest value task um and make
弄清楚如何利用這些代幣來完成最高價值的任務，並從中獲利，因為最終你可能在一年或兩年後

[15:03] to the highest value task um and make money off of it because ultimately what
完成最高價值的任務，並從中獲利，因為最終你可能在一年或兩年後

[15:05] money off of it because ultimately what you're doing maybe maybe like a year
獲利，因為最終你可能在一年或兩年後

[15:07] you're doing maybe maybe like a year from now or two years from now the
你可能在一年或兩年後

[15:09] from now or two years from now the business is actually just arbitrageing
從現在開始，這項業務實際上只是在套利代幣，對嗎？

[15:10] business is actually just arbitrageing tokens, right? The tokens are amazing,
這項業務實際上只是在套利代幣，對嗎？代幣很棒，但讓我們弄清楚將它們指向哪個方向，然後三四年

[15:11] tokens, right? The tokens are amazing, but let's figure out what direction to
代幣很棒，但讓我們弄清楚將它們指向哪個方向，然後三四年

[15:13] but let's figure out what direction to point them in and then three or four
但讓我們弄清楚將它們指向哪個方向，然後三四年

[15:16] point them in and then three or four years from now the model will know
将它们指向其中，然后从现在起三到四年，模型就会知道

[15:17] years from now the model will know, you know, what to do with the tokens and how
从现在起几年后，模型就会知道，你知道，如何处理令牌以及如何

[15:18] know, what to do with the tokens and how to make the most value.
知道，如何处理令牌以及如何创造最大的价值。

[15:20] to make the most value. You know, you can look at this retroactively.
创造最大的价值。你知道，你可以回顾性地看待这件事。

[15:21] can look at this retroactively. Pick any benchmark.
可以回顾性地看待这件事。选择任何基准。

[15:24] benchmark. The cost to hit a certain capability tier used to cost X and now
基准。达到某个能力等级的成本曾经是 X，而现在

[15:27] capability tier used to cost X and now it cost 1/100th or 1/ 1,000th of that.
能力等级的成本曾经是 X，而现在成本是它的 1/100 或 1/1000。

[15:30] it cost 1/100th or 1/ 1,000th of that. Deepseek, for example, on GPD4 was
成本是它的 1/100 或 1/1000。例如，Deepseek 在 GPD4 上的成本是

[15:33] Deepseek, for example, on GPD4 was 1/600th the cost. And since then, the
Deepseek，例如，在 GPD4 上的成本是 1/600。此后，

[15:36] 1/600th the cost. And since then, the costs have fallen further for GPD4 class
1/600 的成本。此后，GPD4 级模型的成本进一步下降。

[15:38] costs have fallen further for GPD4 class models. Of course, no one gives a crap
GPD4 级模型的成本进一步下降。当然，没有人关心

[15:41] models. Of course, no one gives a crap about GP4 class models. They want the
GP4 级模型。他们想要的是前沿模型，因为前沿模型能让他们

[15:43] about GP4 class models. They want the frontier because the frontier lets them
GP4 级模型。他们想要的是前沿模型，因为前沿模型能让他们

[15:44] frontier because the frontier lets them create the economically valuable things.
前沿模型，因为前沿模型能让他们创造经济上有价值的东西。

[15:46] create the economically valuable things. But GP4 class models can still be used
创造经济上有价值的东西。但是 GP4 级模型仍然可以用于

[15:48] But GP4 class models can still be used in like stuff and so people are using
但是 GP4 级模型仍然可以用于类似的东西，所以人们正在使用

[15:50] in like stuff and so people are using them in some like tiny use cases. It's
在一些类似的小用例中。这只是因为成本下降得太快了。

[15:52] them in some like tiny use cases. It's just the cost have fallen so fast. It's
在一些类似的小用例中。这只是因为成本下降得太快了。这

[15:54] just the cost have fallen so fast. It's it's not really what's driving the
只是因为成本下降得太快了。这并不是真正驱动

[15:55] it's not really what's driving the demand. What's driving the demand is is
这并不是真正驱动需求。驱动需求的是

[15:57] demand. What's driving the demand is is all these new use cases. Yeah. Current
需求。驱动需求的是所有这些新的用例。是的。目前的

[16:00] all these new use cases. Yeah. Current 4.6 opus or 4.7 opus tier models a year
所有这些新的用例。是的。目前的 4.6 opus 或 4.7 opus 级模型，一年后

[16:04] from now my spend for the same exact quality of the model would probably be
从现在起，我为模型相同质量的支出可能会是

[16:07] from now my spend for the same exact quality of the model would probably be like 70k.
从现在起，我为模型相同质量的支出可能会是 7 万。

[16:10] quality of the model would probably be like 70k. I bet you it'll be 100 times
模型质量的支出可能会是 7 万。我敢打赌它会便宜 100 倍。

[16:13] like 70k. I bet you it'll be 100 times cheaper. irrelevant because I'm going to
7 万。我敢打赌它会便宜 100 倍。无关紧要，因为我将要

[16:15] cheaper. irrelevant because I'm going to be using a way way way better model
便宜。无关紧要，因为我将要使用一个好得多得多的模型

[16:17] be using a way way way better model which can do way way better things.
使用一个好得多的模型，它可以做许多更好的事情。

[16:18] which can do way way better things.
它可以做许多更好的事情。

[16:20] Enthropic mythos is more expensive as a model but it spends a lot less tokens to
Enthropic mythos作为一个模型更昂贵，但它花费的令牌少得多，用于

[16:22] model but it spends a lot less tokens to do the thing and therefore it is
模型，但它花费的令牌少得多，用于完成任务，因此它

[16:24] do the thing and therefore it is actually cheaper in most tasks than 46
完成任务，因此它在大多数任务中实际上比 46

[16:26] actually cheaper in most tasks than 46 opus because it's just way more
实际上比 46 opus 更便宜，因为它效率高得多

[16:28] opus because it's just way more efficient even though each individual
opus，因为它的效率高得多，即使每个单独的

[16:29] efficient even though each individual token is smarter.
效率高，即使每个单独的令牌都更智能。

[16:30] token is smarter.
令牌更智能。

[16:32] When I last saw you Methos had just come out maybe the day before or something or
我上次见到你时，Methos 可能在前一天或类似的时候刚刚发布，或者

[16:34] out maybe the day before or something or the the card had just come out and you
刚刚发布，或者卡刚刚发布，你

[16:36] the the card had just come out and you said something like uh it actually made
刚刚发布，你说的话，好像它实际上让你

[16:38] said something like uh it actually made you feel like a little scared it was so
感觉有点害怕，它太好了。

[16:40] you feel like a little scared it was so good. What did you mean by that?
你感觉有点害怕，它太好了。你是什么意思？

[16:41] good. What did you mean by that?
太好了。你是什么意思？

[16:46] Anthropic's whole like goal in 2025 was and and even a lot of 2024 they're like
Anthropic 在 2025 年的整个目标是，甚至在 2024 年的很多时候，他们都说

[16:48] and and even a lot of 2024 they're like hey by the end of 2025 we need an L4
甚至在 2024 年的很多时候，他们都说嘿，到 2025 年底，我们需要一个 L4

[16:51] hey by the end of 2025 we need an L4 software engineer uh in our model and
嘿，到 2025 年底，我们需要一个 L4 软件工程师在我们的模型中，并且

[16:54] software engineer uh in our model and and they by and large achieved that with
软件工程师在我们的模型中，并且他们基本上已经实现了，用

[16:55] and they by and large achieved that with 46 Opus. What they didn't say is that
并且他们基本上已经用 46 Opus 实现了。他们没有说的是

[16:57] 46 Opus. What they didn't say is that you know and if you look at Mythos and
46 Opus。他们没有说的是，你知道，如果你看看 Methos，并且

[16:59] you know and if you look at Mythos and if you compare like the benchmarks it's
你知道，如果你看看 Methos，并且如果你比较一下基准测试，它就像

[17:01] if you compare like the benchmarks it's like an L6 engineer. So L4 is like
如果你比较一下基准测试，它就像一个 L6 工程师。所以 L4 就像

[17:04] like an L6 engineer. So L4 is like pretty new. L6 is like quite well
一个 L6 工程师。所以 L4 就像一个新手。L6 就像经验丰富。

[17:06] pretty new. L6 is like quite well experienced. I think Anthropic said that
新手。L6 就像经验丰富。我认为 Anthropic 说

[17:08] experienced. I think Anthropic said that the model internally was available in
经验丰富。我认为 Anthropic 说模型在内部于

[17:10] the model internally was available in February. So in two months they've gone
二月可用。所以在两个月内，他们已经从

[17:13] February. So in two months they've gone from L4 engineer to L6 engineer. Uh
二月。所以在两个月内，他们已经从 L4 工程师变成了 L6 工程师。嗯

[17:16] from L4 engineer to L6 engineer. Uh what's next? Um you know when when you
从 L4 工程师变成了 L6 工程师。嗯，接下来是什么？嗯，你知道，当当

[17:19] what's next?
接下来是什么？

[17:21] Um you know when when you think about the model progress it's only accelerated.
嗯，你知道，当你想到模型的进展时，它只是加速了。

[17:23] Enthropic release cadence has compressed.
Anthropic 的发布节奏已经压缩。

[17:25] Open's release cadence has compressed.
Open 的发布节奏已经压缩。

[17:27] Why? Because these models generally to make a better model you need a few things right.
为什么？因为这些模型通常要制造一个更好的模型，你需要一些正确的东西。

[17:28] You need amazing compute.
你需要惊人的计算能力。

[17:30] Compute is very expensive and it has a time scale that we you know we track and it's like you know it's growing but like you know it's it's sort of set in stone for the next you know short short term.
计算非常昂贵，它有一个时间尺度，你知道我们跟踪它，就像你知道它在增长，但你知道它在接下来的，你知道的短期内，基本上是确定的。

[17:38] it's like kind of set in stone what you've already signed.
这就像你已经签署的东西一样是确定的。

[17:39] Um there will be delays and shifts and some somehow you can find a little more but it's generally pretty set in stone.
嗯，会有延迟和变化，不知何故你可以找到更多一点，但它通常是相当确定的。

[17:44] There's amazing researchers that people are paying tens of millions of dollars for.
有很棒的研究人员，人们为此支付数千万美元。

[17:47] And then lastly there's implementation.
最后是实现。

[17:49] Historically has been very difficult.
历史上一直非常困难。

[17:51] If I have an idea now I have to implement it.
如果我现在有一个想法，我必须实现它。

[17:52] Implementing is hard.
实现是困难的。

[17:54] Now ideas are there.
现在想法已经有了。

[17:57] Implementation is very easy.
实现非常容易。

[18:00] It's expensive but it's very easy.
它很昂贵，但它非常容易。

[18:02] So how do you how does one decide what ideas to implement?
那么你如何决定实现哪些想法呢？

[18:05] And it turns out if your implementation is just so much easier now you can just implement more ideas and move on the treadmill faster and faster and faster.
结果是，如果你的实现现在变得如此容易，你就可以实现更多的想法，并且在跑步机上越跑越快。

[18:10] Whether that is AI model research and so now your model release cadence is shrunk to down to 2 months from where it was 6 months before
无论是人工智能模型研究，所以现在你的模型发布节奏已经从之前的 6 个月缩短到 2 个月了。

[18:19] months from where it was 6 months before or hey I want to I want to take every power plant in the US and every transmission line and model it and run regressions and see the micro supply and demand.
幾個月前的情況，或者嘿，我想我要把美國的每一個發電廠和每一條輸電線路都建立模型並運行迴歸分析，看看微觀的供需。

[18:26] I can also do that. The idea is cheap. You know which idea makes sense? which idea is worth the capital that you have to spend on the tokens because the implementation is there.
我也可以做到。這個想法很便宜。你知道哪個想法有意義嗎？哪個想法值得你花費代幣的資本，因為實施就在那裡。

[18:35] It's it's that's the I think the key learning and if implementation costs continue to tank which they are um we don't even have mythos yet.
這是，這是，我認為這是關鍵的學習，如果實施成本繼續下跌，而它們確實如此，嗯，我們甚至還沒有神話。

[18:46] It's only been you know a handful of hours since Opus 47 launched but you know my team is pretty excited about it internally.
你知道，自 Opus 47 推出以來才幾個小時，但你知道我的團隊在內部對此感到非常興奮。

[18:50] What now comes to the world uh it's a complete reordering of how like economies work.
現在傳到世界上，呃，這將徹底重塑經濟的運作方式。

[18:57] What used to matter a lot was execution was very very difficult and ideas were cheap.
過去很重要的東西是執行非常非常困難，而想法很便宜。

[19:02] Now ideas are cheap and plentiful but execution is very easy.
現在想法便宜且豐富，但執行非常容易。

[19:07] So really only the good ideas are worth are the ones that can justify the spend on super cheap implementation.
所以真正有價值的想法是那些能夠證明花費在超級廉價實施上的價值的想法。

[19:12] So are you actually scared or are you just is it just does it just introduce an uncertainty that's hard to grapple with?
所以你真的害怕嗎，還是你只是，它只是引入了一種難以應對的不確定性？

[19:18] Uncertainty is there. Um but I do I do
不確定性是存在的。嗯，但我確實，我確實

[19:22] Uncertainty is there.
存在不確定性。

[19:22] Um but I do I do think that causes some fear in terms of think that causes some fear in terms of how does society reform itself?
嗯，但我確實認為這會引起一些恐懼，關於社會如何改革自身？

[19:29] How does one one exist in a world where actually any you exist in a world where actually any you know your ability to implement something know your ability to implement something is not actually that important.
一個人如何在一個實際上任何你，你存在於一個實際上任何你知道你實施某種能力的世界上，你的實施某種能力的實際上並不那麼重要？

[19:37] Your ability to choose the correct idea for AI to implement and then your ability to sell that idea or sell what the AI has implemented is what matters.
你選擇正確想法讓 AI 實施的能力，然後你推銷該想法或推銷 AI 已實施內容的能力，這才是重要的。

[19:45] Your ability to garner capital towards that is what matters.
你為此籌集資金的能力才是重要的。

[19:49] And going back to the point of like it's very important to have the newest model always.
回到重點，擁有最新的模型總是至關重要的。

[19:52] Who's going to have access to the newest model?
誰能獲得最新的模型？

[19:53] Anthropics project.
Anthropic 的項目。

[19:55] I know it's not called earwig, but I troll anthropic people by calling it earwig.
我知道它不叫 earwig，但我通過稱它為 earwig 來嘲弄 Anthropic 的人。

[19:59] Um, glasswig anthropic earwig, you know, where they only release mythos to certain companies for cyber.
嗯，glasswig anthropic earwig，你知道，他們只向某些公司發布 mythos 以用於網絡安全。

[20:05] That's just going to be something that continues.
這只會是會持續下去的事情。

[20:07] Models will have less broad and less broad deployment.
模型將具有較少廣泛的部署。

[20:09] I know I know Open AI and Enthropic and all these people are like, we want to have great AI for everyone.
我知道我知道 OpenAI 和 Anthropic 以及所有這些人都說，我們希望為每個人提供偉大的 AI。

[20:15] AI is very [ __ ] expensive.
AI 非常 [ __ ] 昂貴。

[20:18] Who's going to pay for the trillion dollars of infrastructure?
誰將支付數萬億美元的基礎設施費用？

[20:19] People who have money and can can build useful
有錢並且能夠建立有用的人

[20:22] have money and can can build useful things with AI. And then you don't want

[20:24] things with AI. And then you don't want people to distill your models. So you

[20:25] people to distill your models. So you don't release them broadly. Uh you

[20:27] don't release them broadly. Uh you release them to a fewer and fewer set of

[20:29] release them to a fewer and fewer set of customers. Those customers are also now

[20:31] customers. Those customers are also now wrestling over the tokens unless

[20:33] wrestling over the tokens unless anthropic jacks them. You know, they

[20:34] anthropic jacks them. You know, they could double their pricing on Opus and I

[20:36] could double their pricing on Opus and I would continue to pay and I bet most

[20:37] would continue to pay and I bet most users would continue to pay.

[20:38] users would continue to pay. >> I bet that wouldn't solve their

[20:40] >> I bet that wouldn't solve their humongous capacity problem that they

[20:42] humongous capacity problem that they have. So then the question becomes where

[20:44] have. So then the question becomes where does this cycle end where you know token

[20:47] does this cycle end where you know token usage and therefore the benefits of

[20:49] usage and therefore the benefits of those tokens the additional value

[20:51] those tokens the additional value generated on top of those tokens

[20:52] generated on top of those tokens aggregates among fewer and fewer and

[20:54] aggregates among fewer and fewer and fewer companies. I don't have mythos.

[20:56] fewer companies. I don't have mythos. You know who has mythos? Top freaking

[20:58] You know who has mythos? Top freaking banks. Um now they're only using it for

[21:00] banks. Um now they're only using it for cyber security. But at some point I can

[21:02] cyber security. But at some point I can envision a world where hey maybe I

[21:04] envision a world where hey maybe I because I have an enterprise enthropic

[21:05] because I have an enterprise enthropic contract and because enthropic people

[21:07] contract and because enthropic people kind of like me they're willing to give

[21:09] kind of like me they're willing to give us like slightly earlier access or

[21:11] us like slightly earlier access or slightly higher rate limits or something

[21:13] slightly higher rate limits or something for a model. I hope that's what happens.

[21:15] for a model. I hope that's what happens. And then my competitor whoever that is

[21:18] And then my competitor whoever that is doesn't have that and I'm able to

[21:19] doesn't have that and I'm able to [ __ ] crush them. There are people who

[21:21] [ __ ] crush them. There are people who are like Ken Griffin of Citadel is like

[21:23] are like Ken Griffin of Citadel is like super well-connected and super rich and

[21:25] super well-connected and super rich and he's like he he just signs, you know,

[21:27] he's like he he just signs, you know, who knows? He goes and signs a deal with

[21:28] who knows? He goes and signs a deal with Open Arenthropic that's like, "Yeah, I'm

[21:30] Open Arenthropic that's like, "Yeah, I'm going to get access to your models. Um,

[21:32] going to get access to your models. Um, and I'll buy the first $10 billion worth

[21:35] and I'll buy the first $10 billion worth of tokens each year. So, whenever you

[21:36] of tokens each year. So, whenever you release the model, you know, I'll spend

[21:38] release the model, you know, I'll spend the first 10 billion tokens and then

[21:39] the first 10 billion tokens and then everyone else can get the model after

[21:40] everyone else can get the model after that." And it's like, okay, well, now

[21:42] that." And it's like, okay, well, now what does that do? Well, now he's going

[21:43] what does that do? Well, now he's going to crush everyone in the market. That's

[21:44] to crush everyone in the market. That's just an example. Could be cyber like

[21:46] just an example. Could be cyber like Anthropic is worried about, oh, now I

[21:47] Anthropic is worried about, oh, now I can hack people. could be information

[21:49] can hack people. could be information services business like myself where I

[21:50] services business like myself where I crush someone else. I think you know it

[21:52] crush someone else. I think you know it it's it's such a broad base. We don't

[21:54] it's it's such a broad base. We don't know what these models can do. Anthropic

[21:55] know what these models can do. Anthropic doesn't know what these models can do.

[21:56] doesn't know what these models can do. No one knows what these models can do.

[21:57] No one knows what these models can do. It's up to the end user to figure out

[21:59] It's up to the end user to figure out where they can leverage the tokens to

[22:00] where they can leverage the tokens to see what they can build and imagine

[22:02] see what they can build and imagine which is tremendously productive and

[22:04] which is tremendously productive and uplifting for humanity. But then what

[22:06] uplifting for humanity. But then what happens to the concentration of

[22:07] happens to the concentration of resources and usage of it?

[22:08] resources and usage of it? >> Presumably right now robotics or robots

[22:11] >> Presumably right now robotics or robots consume relatively zero tokens versus

[22:14] consume relatively zero tokens versus everything else. Do you see what's your

[22:16] everything else. Do you see what's your view of that? If that's like a second

[22:18] view of that? If that's like a second demand curve that could start to

[22:19] demand curve that could start to ratchet, there's a new startup every

[22:21] ratchet, there's a new startup every single day, you know, within a mile of

[22:23] single day, you know, within a mile of here trying to build something

[22:24] here trying to build something interesting in robotics.

[22:25] interesting in robotics. >> So there's this concept of software only

[22:27] >> So there's this concept of software only singularity, which is that the world

[22:29] singularity, which is that the world has, you know, AI singularity, but only

[22:31] has, you know, AI singularity, but only in software. And now what about the rest

[22:33] in software. And now what about the rest of the world? Vast majority of the world

[22:35] of the world? Vast majority of the world is physical. You can see the world

[22:38] is physical. You can see the world orient around hardware, not software.

[22:40] orient around hardware, not software. That's actually why I think software

[22:42] That's actually why I think software only singularity is like just a blip and

[22:44] only singularity is like just a blip and not like a you know we we do get

[22:45] not like a you know we we do get everything else because once software is

[22:47] everything else because once software is super easy what makes robots really hard

[22:49] super easy what makes robots really hard it's like programming microcontrollers

[22:51] it's like programming microcontrollers and actuators and controlling all this

[22:52] and actuators and controlling all this stuff is very difficult right now the

[22:55] stuff is very difficult right now the interesting thing about models AI models

[22:57] interesting thing about models AI models is they're actually really inefficient

[22:59] is they're actually really inefficient in learning it's just we're able to give

[23:01] in learning it's just we're able to give them so much data that they're able to

[23:03] them so much data that they're able to learn and surpass us in certain ways

[23:05] learn and surpass us in certain ways robots currently the robot models um

[23:07] robots currently the robot models um VA's uh vision language action models

[23:10] VA's uh vision language action models which is very popular right now is

[23:12] which is very popular right now is probably not going to be the thing that

[23:14] probably not going to be the thing that ultimately scales beyond. They are

[23:16] ultimately scales beyond. They are inefficient in data um and we can't

[23:18] inefficient in data um and we can't scale the data for them fast enough.

[23:20] scale the data for them fast enough. There is going to be some way to large

[23:22] There is going to be some way to large scale pre-train robot models where just

[23:24] scale pre-train robot models where just like humans see all this data throughout

[23:26] like humans see all this data throughout their lives. And what's interesting is

[23:28] their lives. And what's interesting is humans the reason why we're so good is

[23:30] humans the reason why we're so good is we're sample efficient. One example, two

[23:32] we're sample efficient. One example, two example, we're good. And so applying

[23:33] example, we're good. And so applying that to robotics. So once you once you

[23:35] that to robotics. So once you once you have this software only singularity

[23:37] have this software only singularity implementation is super cheap. anyone

[23:38] implementation is super cheap. anyone can start to build these mo people can

[23:40] can start to build these mo people can start to build models that now robots

[23:43] start to build models that now robots are actually useful and so I think in

[23:45] are actually useful and so I think in the next six to 18 months we'll start

[23:46] the next six to 18 months we'll start seeing real breakthroughs in robotics

[23:49] seeing real breakthroughs in robotics that enable few shot learning i.e.

[23:52] that enable few shot learning i.e. there's a pre-trained robot model and

[23:54] there's a pre-trained robot model and now there's a robot that you have hired

[23:56] now there's a robot that you have hired or bought or whatever. You show it a few

[23:58] or bought or whatever. You show it a few examples and it's able to do it. You

[23:59] examples and it's able to do it. You tell it to stack these two things or you

[24:01] tell it to stack these two things or you tell it, hey, this can can actually like

[24:03] tell it, hey, this can can actually like balance perfectly, you know, and and it

[24:05] balance perfectly, you know, and and it starts doing these things.

[24:06] starts doing these things. >> Nicely done.

[24:07] >> Nicely done. >> One shot.

[24:09] >> One shot. >> No, trust me, I've spilled many of

[24:11] >> No, trust me, I've spilled many of times.

[24:12] times. >> So, I think I think robots will get fot

[24:15] >> So, I think I think robots will get fot learning right now. Now, you know,

[24:16] learning right now. Now, you know, there's a lot of companies doing robots

[24:18] there's a lot of companies doing robots for like, you know, advertisement or

[24:19] for like, you know, advertisement or robots for like simple stuff like that,

[24:21] robots for like simple stuff like that, but it'll be like, oh, folding clothes,

[24:23] but it'll be like, oh, folding clothes, but it's going to get really niche like

[24:24] but it's going to get really niche like robots just for cleaning chalkboards.

[24:26] robots just for cleaning chalkboards. Um, and it's a rental service or, you

[24:28] Um, and it's a rental service or, you know, it'll be it'll be a model package

[24:30] know, it'll be it'll be a model package that you download onto your standard

[24:31] that you download onto your standard robot that then does that, right? And

[24:33] robot that then does that, right? And and you pay for that. And anyways, there

[24:35] and you pay for that. And anyways, there will be a huge explosion in physical

[24:36] will be a huge explosion in physical good acceleration and and deflationary

[24:39] good acceleration and and deflationary effects there. But and and so that's

[24:41] effects there. But and and so that's that's ultimately going to keep token

[24:43] that's ultimately going to keep token demand going crazy. I I don't think

[24:45] demand going crazy. I I don't think token demand slows down personally.

[24:46] token demand slows down personally. >> Did you learn anything else about the

[24:48] >> Did you learn anything else about the world based on Mythos's results and how

[24:51] world based on Mythos's results and how it was built? My way of asking like the

[24:53] it was built? My way of asking like the you know if you break down the the

[24:54] you know if you break down the the components of the scaling laws like the

[24:56] components of the scaling laws like the >> So Methos is a materially larger model

[24:58] >> So Methos is a materially larger model than prior models and so yes it is a

[25:01] than prior models and so yes it is a much larger model. Now whether or not

[25:03] much larger model. Now whether or not it's it's what chip it's trained on is

[25:05] it's it's what chip it's trained on is not really relevant. It's the scale and

[25:07] not really relevant. It's the scale and obviously you know to a 100,000 black

[25:09] obviously you know to a 100,000 black wells is equivalent to hundreds of

[25:11] wells is equivalent to hundreds of thousands of prior generation chips.

[25:12] thousands of prior generation chips. TPUs and tranium have their different

[25:14] TPUs and tranium have their different release cadence. So it's not exactly

[25:15] release cadence. So it's not exactly like mirrored one to one. Um but

[25:17] like mirrored one to one. Um but ultimately yes mythos is a significantly

[25:19] ultimately yes mythos is a significantly larger model. It's proof that the

[25:20] larger model. It's proof that the scaling laws still work. Um everything

[25:22] scaling laws still work. Um everything about it shows the trend line continues

[25:24] about it shows the trend line continues of models. More compute into model makes

[25:26] of models. More compute into model makes model better. And along the whole way

[25:28] model better. And along the whole way it's not just more compute into model

[25:29] it's not just more compute into model makes model better. along the whole way

[25:31] makes model better. along the whole way we're also getting these compute

[25:32] we're also getting these compute efficiency wins which are you know as as

[25:35] efficiency wins which are you know as as all this research compute that the labs

[25:37] all this research compute that the labs are spending is actually turning into if

[25:39] are spending is actually turning into if I want x capability tier model every 6

[25:42] I want x capability tier model every 6 months that cost or every two months

[25:43] months that cost or every two months that cost is dramatically decreasing but

[25:45] that cost is dramatically decreasing but then if I scale it up massively I get a

[25:47] then if I scale it up massively I get a humongous capability jump as well and so

[25:50] humongous capability jump as well and so yes it's it's proof that this is still

[25:52] yes it's it's proof that this is still happening Google and anthropic are not

[25:53] happening Google and anthropic are not heavy heavy users of GPUs on the

[25:55] heavy heavy users of GPUs on the training side but openai they'll they'll

[25:58] training side but openai they'll they'll start having their new class of models I

[26:00] start having their new class of models I think they're taking a more sensible

[26:01] think they're taking a more sensible principled approach to scaling uh in

[26:04] principled approach to scaling uh in small steps. Enthropic really went for a

[26:06] small steps. Enthropic really went for a huge jump. We'll see better and better

[26:07] huge jump. We'll see better and better models throughout the year and the

[26:08] models throughout the year and the release cadence is only going to get

[26:10] release cadence is only going to get faster.

[26:10] faster. >> We've gone a long way in the

[26:11] >> We've gone a long way in the conversation with saying almost nothing

[26:13] conversation with saying almost nothing about OpenAI which would have been so

[26:14] about OpenAI which would have been so strange.

[26:15] strange. >> So, so this is this is the interesting

[26:16] >> So, so this is this is the interesting thing. Everyone's like, okay, so

[26:18] thing. Everyone's like, okay, so Anthropics just won, right? You know,

[26:19] Anthropics just won, right? You know, they had Methos in February. They never

[26:21] they had Methos in February. They never even released it cuz they didn't feel

[26:22] even released it cuz they didn't feel the need to. They're already sold out.

[26:23] the need to. They're already sold out. Their revenue is already adding $10

[26:25] Their revenue is already adding $10 billion a month. Um and then you've got

[26:27] billion a month. Um and then you've got Opus 47 today all before open eyes you

[26:31] Opus 47 today all before open eyes you know um alleged Spud release which you

[26:34] know um alleged Spud release which you know media such as the information and

[26:36] know media such as the information and others have have posted about. So

[26:38] others have have posted about. So clearly Anthropic is in the lead right

[26:40] clearly Anthropic is in the lead right and OpenAI is cooked. What's interesting

[26:42] and OpenAI is cooked. What's interesting is because Anthropic has such bounds on

[26:45] is because Anthropic has such bounds on compute and they can only grow it so

[26:48] compute and they can only grow it so fast and sort of to the point of you

[26:49] fast and sort of to the point of you know you know Daria Daria used to gloat

[26:51] know you know Daria Daria used to gloat about how OpenAI was being too

[26:54] about how OpenAI was being too aggressive on compute and Anthropic was

[26:56] aggressive on compute and Anthropic was more sensible in their scaling and now

[26:58] more sensible in their scaling and now Enthropic is like [ __ ] we should have I

[27:00] Enthropic is like [ __ ] we should have I wish we had a lot more compute. OpenAI

[27:01] wish we had a lot more compute. OpenAI is able to pay the bills perfectly fine.

[27:04] is able to pay the bills perfectly fine. In fact, they've raised a ton of money

[27:05] In fact, they've raised a ton of money to get incremental compute in addition

[27:08] to get incremental compute in addition to the irresponsible levels of compute

[27:10] to the irresponsible levels of compute that they were buying from Oracle and

[27:11] that they were buying from Oracle and Core and SoftBank and all these people

[27:13] Core and SoftBank and all these people and Microsoft uh you know such as

[27:15] and Microsoft uh you know such as Tranium. Now they're getting tranium as

[27:16] Tranium. Now they're getting tranium as well from Amazon. Um so so they've done

[27:19] well from Amazon. Um so so they've done this like insane thing on compute and

[27:21] this like insane thing on compute and they need know they also know they need

[27:22] they need know they also know they need more. But what's interesting is if you

[27:24] more. But what's interesting is if you were to say Opus 46, you know, let's

[27:27] were to say Opus 46, you know, let's ignore models getting better over time.

[27:29] ignore models getting better over time. Let's just take diffusion of this

[27:31] Let's just take diffusion of this technology. You and I may get jump on

[27:33] technology. You and I may get jump on the model immediately day one, but other

[27:35] the model immediately day one, but other businesses take time and it takes time

[27:37] businesses take time and it takes time for people to learn and the spark of oh

[27:40] for people to learn and the spark of oh [ __ ] claude psychosis moment doesn't hit

[27:42] [ __ ] claude psychosis moment doesn't hit everyone at the same time. And so by the

[27:45] everyone at the same time. And so by the end of the year, let's say a 46 opus

[27:46] end of the year, let's say a 46 opus tier model the economy would spend

[27:48] tier model the economy would spend $und00 billion on. I don't think that's

[27:50] $und00 billion on. I don't think that's unreasonable. It's spending $40 billion

[27:51] unreasonable. It's spending $40 billion right now.

[27:52] right now. >> That's like a linear extrapolation.

[27:54] >> That's like a linear extrapolation. >> It's a linear extrapolation, not a not

[27:55] >> It's a linear extrapolation, not a not an exponential. To get the exponential,

[27:57] an exponential. To get the exponential, you need the better models. Enthropic

[27:59] you need the better models. Enthropic won't have enough compute to do that.

[28:00] won't have enough compute to do that. And so and and presumably OpenAI and

[28:03] And so and and presumably OpenAI and Google will hit that tier soon enough.

[28:05] Google will hit that tier soon enough. Whoever hits that tier next, sure,

[28:07] Whoever hits that tier next, sure, Enthropic may get to charge 70 plus%

[28:09] Enthropic may get to charge 70 plus% gross margins, but if OpenAI hits it

[28:11] gross margins, but if OpenAI hits it next, they charge 50% gross margins.

[28:14] next, they charge 50% gross margins. They still get all of this incremental

[28:15] They still get all of this incremental demand. And probably they also won't

[28:17] demand. And probably they also won't have enough compute to serve all the

[28:18] have enough compute to serve all the users. And so, sure, maybe Mythos is a

[28:22] users. And so, sure, maybe Mythos is a model where if the world had enough

[28:23] model where if the world had enough compute, it'd be $500 billion of revenue

[28:26] compute, it'd be $500 billion of revenue or something crazy. There is such demand

[28:28] or something crazy. There is such demand for these tokens and such limitations on

[28:30] for these tokens and such limitations on compute, you know, and we see this with

[28:32] compute, you know, and we see this with H100 prices skyrocketing and the useful

[28:34] H100 prices skyrocketing and the useful life of these GPUs continue to extend.

[28:36] life of these GPUs continue to extend. It's pretty clear even the tier 2 lab is

[28:38] It's pretty clear even the tier 2 lab is going to be sold out of tokens, let

[28:39] going to be sold out of tokens, let alone the tier one lab. The tier one lab

[28:41] alone the tier one lab. The tier one lab will have better margins, but the tier

[28:43] will have better margins, but the tier two lab will be sold out and probably

[28:45] two lab will be sold out and probably the tier three lab will also be close to

[28:46] the tier three lab will also be close to sold out. Economic value that the best

[28:48] sold out. Economic value that the best model can deliver is growing faster than

[28:50] model can deliver is growing faster than our ability to actually serve those

[28:52] our ability to actually serve those tokens to people via the infrastructure.

[28:54] tokens to people via the infrastructure. And so this gap will continue to grow

[28:55] And so this gap will continue to grow and the model labs will continue to have

[28:57] and the model labs will continue to have expanding margins until people in the

[28:59] expanding margins until people in the hardware supply chain infrastructure

[29:00] hardware supply chain infrastructure supply chain are like wait no why don't

[29:01] supply chain are like wait no why don't I just jack up my margins. So suffice to

[29:03] I just jack up my margins. So suffice to say I think the assessment today or your

[29:05] say I think the assessment today or your assessment of the demand side is

[29:07] assessment of the demand side is completely explosive in your own

[29:08] completely explosive in your own particular example here at semi analysis

[29:10] particular example here at semi analysis but just more broadly that as people

[29:12] but just more broadly that as people fall in you call it AI psychosis as

[29:14] fall in you call it AI psychosis as people fall into this experience of what

[29:16] people fall into this experience of what they can do the implementation

[29:18] they can do the implementation difficulty going completely away I I've

[29:20] difficulty going completely away I I've certainly felt that you know my own

[29:22] certainly felt that you know my own token spend is just through the absolute

[29:23] token spend is just through the absolute roof just in the matter of weeks so that

[29:26] roof just in the matter of weeks so that that feels like a pretty good assessment

[29:27] that feels like a pretty good assessment anything we're missing on the demand

[29:28] anything we're missing on the demand side

[29:29] side >> if you don't use more tokens you'll

[29:30] >> if you don't use more tokens you'll never escape the permanent underclass

[29:32] never escape the permanent underclass just expand on that.

[29:33] just expand on that. >> So either either you use more tokens and

[29:35] >> So either either you use more tokens and you generate economic value outsized

[29:37] you generate economic value outsized economic value for the use of those

[29:39] economic value for the use of those tokens. Um a lot of people are doing it

[29:40] tokens. Um a lot of people are doing it the boring lazy way. Oh, I guess I'll

[29:42] the boring lazy way. Oh, I guess I'll just work one hour a day instead of

[29:43] just work one hour a day instead of eight hours a day and I'll have AI do

[29:45] eight hours a day and I'll have AI do most of my job. That's the boring way.

[29:47] most of my job. That's the boring way. The cool way is I'll still work eight

[29:49] The cool way is I'll still work eight hours a day and I'll I'll do 8x the work

[29:51] hours a day and I'll I'll do 8x the work and maybe I'll make 5x the money. Um

[29:54] and maybe I'll make 5x the money. Um maybe not you can't do this with a job

[29:55] maybe not you can't do this with a job obviously. There's people who have

[29:57] obviously. There's people who have multiple jobs. Um there's people who

[29:58] multiple jobs. Um there's people who like start companies and start selling

[30:00] like start companies and start selling stuff. get that economic value on on

[30:02] stuff. get that economic value on on this AI before everyone is using it and

[30:04] this AI before everyone is using it and it's table stakes. Uh because it's still

[30:06] it's table stakes. Uh because it's still not table stakes if you don't use more

[30:08] not table stakes if you don't use more tokens and generate the value from them

[30:10] tokens and generate the value from them and capture that value. These there's

[30:11] and capture that value. These there's three different problems here. Using

[30:12] three different problems here. Using more tokens, generating value from those

[30:14] more tokens, generating value from those tokens and capturing value from those

[30:16] tokens and capturing value from those tok uh from the value that you created

[30:17] tok uh from the value that you created from the tokens. Uh if you don't do

[30:19] from the tokens. Uh if you don't do these three things, you'll never escape

[30:20] these three things, you'll never escape the permanent underclass i.e. as models

[30:23] the permanent underclass i.e. as models continue to skyrocket in capability and

[30:25] continue to skyrocket in capability and the concentration of resources

[30:26] the concentration of resources potentially happens.

[30:28] potentially happens. >> Okay, let's talk about supply. what is

[30:29] >> Okay, let's talk about supply. what is going on like how would you describe the

[30:31] going on like how would you describe the frontier of what's changing or what is

[30:33] frontier of what's changing or what is changing at the frontier of supplying

[30:35] changing at the frontier of supplying the the entire stack that's required to

[30:37] the the entire stack that's required to serve all these tokens as the demand

[30:39] serve all these tokens as the demand curve explodes

[30:40] curve explodes >> as demand skyrockets prices are going up

[30:42] >> as demand skyrockets prices are going up for everything on the supply side um

[30:45] for everything on the supply side um whether it be the NGPUs

[30:47] whether it be the NGPUs uh their prices are going up in addition

[30:50] uh their prices are going up in addition their useful life is extending

[30:51] their useful life is extending >> H100 prices look like this

[30:53] >> H100 prices look like this >> yeah exactly there's people who have

[30:54] >> yeah exactly there's people who have argued GPU's full lives are less than 5

[30:56] argued GPU's full lives are less than 5 years complete nonsense

[30:58] years complete nonsense Um there are clusters now resigning

[31:01] Um there are clusters now resigning three or foury old hopper clusters

[31:02] three or foury old hopper clusters resigning for 3 or four more years. Um

[31:05] resigning for 3 or four more years. Um there's A100 clusters that are resigning

[31:07] there's A100 clusters that are resigning for another couple years. So the useful

[31:08] for another couple years. So the useful life is clearly not 5 years. It's maybe

[31:10] life is clearly not 5 years. It's maybe even seven or eight years. Um arguably

[31:12] even seven or eight years. Um arguably we we don't know yet. We'll see. We'll

[31:14] we we don't know yet. We'll see. We'll see when Hopper gets there, but it it's

[31:16] see when Hopper gets there, but it it's clearly not 5 years. So the useful life

[31:17] clearly not 5 years. So the useful life is extending and the prices are going up

[31:19] is extending and the prices are going up on that renewal. So in effect the gross

[31:22] on that renewal. So in effect the gross margin was not 35% on a cluster, it's

[31:25] margin was not 35% on a cluster, it's beyond that. Um so margins are expanding

[31:27] beyond that. Um so margins are expanding in the in the cloud layer. Margins are

[31:30] in the in the cloud layer. Margins are um extremely healthy on the hardware

[31:33] um extremely healthy on the hardware layer with you know Nvidia still

[31:34] layer with you know Nvidia still charging 75 or whatever percent gross

[31:36] charging 75 or whatever percent gross margin as we move down the stack. Memory

[31:38] margin as we move down the stack. Memory obviously margins have skyrocketed

[31:40] obviously margins have skyrocketed there. Places like optics and logic

[31:44] there. Places like optics and logic there are large prepayments um and

[31:46] there are large prepayments um and margins are growing slowly um more so

[31:49] margins are growing slowly um more so the companies that are making chips like

[31:50] the companies that are making chips like Nvidia are paying huge prepayments. So

[31:53] Nvidia are paying huge prepayments. So in effect the cast of capital or timing

[31:55] in effect the cast of capital or timing of cash flow return on invested capital

[31:57] of cash flow return on invested capital is going up even if the gross margin

[31:58] is going up even if the gross margin isn't. And you see this across the whole

[32:00] isn't. And you see this across the whole supply chain. You see ASML is completely

[32:02] supply chain. You see ASML is completely sold out and they need Carl Zeiss to

[32:04] sold out and they need Carl Zeiss to expand faster. Everywhere along the

[32:06] expand faster. Everywhere along the chain

[32:07] chain everyone's either sold out and margins

[32:09] everyone's either sold out and margins are going up or they're getting

[32:10] are going up or they're getting prepayments increases the return on

[32:12] prepayments increases the return on invested capital because the invested

[32:13] invested capital because the invested capital is lower. And so this is like a

[32:15] capital is lower. And so this is like a consistent trend across any part. It's

[32:17] consistent trend across any part. It's it's even like you know a PCB to make a

[32:19] it's even like you know a PCB to make a PCB requires copper foil and that copper

[32:22] PCB requires copper foil and that copper foil is sold out and people are making

[32:23] foil is sold out and people are making prepayments for it. It's like anything

[32:25] prepayments for it. It's like anything and everything that like has a pulse and

[32:28] and everything that like has a pulse and is like sold out. People are like

[32:29] is like sold out. People are like jumping to get more incremental supply

[32:31] jumping to get more incremental supply and fighting over the supply for the

[32:33] and fighting over the supply for the years after.

[32:34] years after. >> As your business scales up, everything

[32:35] >> As your business scales up, everything gets more complex, especially your

[32:37] gets more complex, especially your compliance and security needs. With so

[32:39] compliance and security needs. With so many tools offering band-aids and

[32:40] many tools offering band-aids and patches, it's unfortunately far too easy

[32:42] patches, it's unfortunately far too easy for something to slip through the

[32:43] for something to slip through the cracks. Fortunately, Vanta is a powerful

[32:45] cracks. Fortunately, Vanta is a powerful tool designed to simplify and automate

[32:47] tool designed to simplify and automate your security work and deliver a single

[32:49] your security work and deliver a single source of truth for compliance and risk.

[32:51] source of truth for compliance and risk. There's a reason that Ramp, Cursor, and

[32:53] There's a reason that Ramp, Cursor, and Snowflake all use Vanta. It frees them

[32:55] Snowflake all use Vanta. It frees them to focus on building amazing

[32:56] to focus on building amazing differentiated products, knowing that

[32:58] differentiated products, knowing that compliance and security are under

[32:59] compliance and security are under control. Learn more at vanta.com/invest.

[33:03] control. Learn more at vanta.com/invest. I know firsthand how complex the tech

[33:05] I know firsthand how complex the tech stack is for asset management firms. And

[33:07] stack is for asset management firms. And seemingly every new tool and data source

[33:09] seemingly every new tool and data source makes the problem even worse. Adding

[33:11] makes the problem even worse. Adding more complexity, more headcount, and

[33:12] more complexity, more headcount, and more risk. Ridgeline offers a better way

[33:15] more risk. Ridgeline offers a better way forward. One unified platform that

[33:16] forward. One unified platform that automates away that that automates away

[33:18] automates away that that automates away that complexity across portfolio

[33:20] that complexity across portfolio accounting, reconciliation, reporting,

[33:22] accounting, reconciliation, reporting, trading, compliance, and more. All at

[33:24] trading, compliance, and more. All at scale. Ridgeline is revolutionizing

[33:26] scale. Ridgeline is revolutionizing investment management, helping ambitious

[33:28] investment management, helping ambitious firms scale faster, operate smarter, and

[33:30] firms scale faster, operate smarter, and stay ahead of the curve. See what

[33:31] stay ahead of the curve. See what Ridgeline can unlock for your firm.

[33:33] Ridgeline can unlock for your firm. Schedule a demo at ridgeline.ai.

[33:36] Schedule a demo at ridgeline.ai. What do you think are the most important

[33:37] What do you think are the most important bottlenecks? Like typically in economic

[33:39] bottlenecks? Like typically in economic history when there's this kind of

[33:41] history when there's this kind of demand, supply reorients and rises very

[33:44] demand, supply reorients and rises very very quickly to meet the demand. It

[33:47] very quickly to meet the demand. It seems like it's almost impossible for

[33:48] seems like it's almost impossible for supply right now in this moment to keep

[33:50] supply right now in this moment to keep up. You know, famous last words, every

[33:51] up. You know, famous last words, every every shortage is followed by a glut

[33:53] every shortage is followed by a glut historically. But what are the most

[33:55] historically. But what are the most interesting bottlenecks to you on across

[33:57] interesting bottlenecks to you on across the supply side?

[33:58] the supply side? >> Supply chains are usually very fast to

[34:00] >> Supply chains are usually very fast to react. Um, one unique thing is that our

[34:03] react. Um, one unique thing is that our supply chains now are more complex than

[34:05] supply chains now are more complex than ever. and the things we're building are

[34:06] ever. and the things we're building are more complex than ever and therefore the

[34:07] more complex than ever and therefore the lead times are longer. Um, and it's not

[34:10] lead times are longer. Um, and it's not like we haven't seen 18-monthl long lead

[34:12] like we haven't seen 18-monthl long lead times in other industries. It's just

[34:15] times in other industries. It's just building incremental supply didn't take

[34:17] building incremental supply didn't take years. Um, and this is the case with

[34:19] years. Um, and this is the case with memory, right? Memory can only grow

[34:22] memory, right? Memory can only grow capacity, you know, low double digit

[34:24] capacity, you know, low double digit percentages a year, right? 20s 30% a

[34:27] percentages a year, right? 20s 30% a year. Um, even less for NAND, a little

[34:29] year. Um, even less for NAND, a little bit higher for DRM. Even though the

[34:30] bit higher for DRM. Even though the demand signal was very strong at the end

[34:31] demand signal was very strong at the end of 2025, the memory companies

[34:33] of 2025, the memory companies immediately sort of started reacting.

[34:35] immediately sort of started reacting. None of that incremental capacity really

[34:37] None of that incremental capacity really gets here until the second that they've

[34:38] gets here until the second that they've decided to do in addition to the typical

[34:40] decided to do in addition to the typical 20 to 30%. You know, they can stretch a

[34:43] 20 to 30%. You know, they can stretch a little bit, but really the true

[34:44] little bit, but really the true incremental supply doesn't come till 28,

[34:46] incremental supply doesn't come till 28, which is a very unique thing. Even if

[34:48] which is a very unique thing. Even if they wanted to build as fast as

[34:49] they wanted to build as fast as possible, it doesn't come till 28 uh

[34:51] possible, it doesn't come till 28 uh early late 27 at best. And so the result

[34:54] early late 27 at best. And so the result is memory prices have, you know, gone

[34:57] is memory prices have, you know, gone through the roof. And guess what?

[34:58] through the roof. And guess what? they're going to double and triple

[34:59] they're going to double and triple again. Um, at least on DRAM especially,

[35:02] again. Um, at least on DRAM especially, people are like, "Oh, the memory storage

[35:03] people are like, "Oh, the memory storage is overplayed. Everyone gets it." And

[35:04] is overplayed. Everyone gets it." And it's like, "No, no, no. You don't get

[35:05] it's like, "No, no, no. You don't get it." DM will double or triple from here

[35:08] it." DM will double or triple from here still because that's that's how much

[35:11] still because that's that's how much capacity is required and they have to

[35:13] capacity is required and they have to steal capacity from somewhere else. And

[35:15] steal capacity from somewhere else. And the only way to steal capacity from

[35:16] the only way to steal capacity from somewhere else in a in a capitalist

[35:18] somewhere else in a in a capitalist economy is demand destruction via higher

[35:20] economy is demand destruction via higher pricing. We're not like rationing stuff

[35:21] pricing. We're not like rationing stuff here. And so ultimately, that's what's

[35:23] here. And so ultimately, that's what's going to happen. And so margins continue

[35:24] going to happen. And so margins continue to go up. Um, I think Logic also has

[35:28] to go up. Um, I think Logic also has humongous uh capacity problems. TSMC

[35:30] humongous uh capacity problems. TSMC just had their earnings. Uh, they keep

[35:32] just had their earnings. Uh, they keep upping capex. Ultimately, you know, it

[35:34] upping capex. Ultimately, you know, it takes them quite some time to build

[35:35] takes them quite some time to build fabs. Um, they're trying to do

[35:36] fabs. Um, they're trying to do everything they can to squeeze every

[35:38] everything they can to squeeze every little output out of every fab that they

[35:40] little output out of every fab that they have. But ultimately, they're not

[35:42] have. But ultimately, they're not raising prices fast because they're good

[35:43] raising prices fast because they're good people. It seems like, um, you know,

[35:45] people. It seems like, um, you know, singledigit price increases instead of,

[35:47] singledigit price increases instead of, you know, tripledigit price increases

[35:49] you know, tripledigit price increases like the memory guys have had. And so

[35:50] like the memory guys have had. And so you ultimately have like this like

[35:52] you ultimately have like this like market where yeah TSMC is a great

[35:53] market where yeah TSMC is a great company but are they are they actually

[35:55] company but are they are they actually going to extract all the value? I

[35:56] going to extract all the value? I mentioned things like copper foil, glass

[35:58] mentioned things like copper foil, glass fibers for PCBs, lasers. These are

[36:01] fibers for PCBs, lasers. These are things that are like well understood and

[36:02] things that are like well understood and niche supply chains but they're very

[36:04] niche supply chains but they're very very tight. Um and ultimately upstream

[36:07] very tight. Um and ultimately upstream the semiconductor wafer fabrication

[36:09] the semiconductor wafer fabrication equipment supply chain is one that like

[36:10] equipment supply chain is one that like I still think is it's gone up a lot but

[36:12] I still think is it's gone up a lot but it's still very underappreciated. TSMC

[36:15] it's still very underappreciated. TSMC capex this year they say 56. Uh we've

[36:17] capex this year they say 56. Uh we've had 57.4 4 billion since January. Um,

[36:20] had 57.4 4 billion since January. Um, and we may up it slightly more just

[36:22] and we may up it slightly more just because we see some some ways that they

[36:24] because we see some some ways that they can get incremental capex. But what

[36:26] can get incremental capex. But what people aren't focusing on is what does

[36:27] people aren't focusing on is what does that mean next year and what does that

[36:28] that mean next year and what does that mean the year after? And it turns out 3

[36:31] mean the year after? And it turns out 3 years from now TSMC is going to spend

[36:32] years from now TSMC is going to spend hundred billion on capex. U maybe two

[36:35] hundred billion on capex. U maybe two years from now, right? Maybe 28.

[36:36] years from now, right? Maybe 28. Sincerely, they may spend $und00 billion

[36:38] Sincerely, they may spend $und00 billion on capex in 2028. And people like just

[36:41] on capex in 2028. And people like just can't fathom that. But what does that

[36:42] can't fathom that. But what does that mean for their downstream supply chains?

[36:44] mean for their downstream supply chains? um you know companies like Lamb Research

[36:46] um you know companies like Lamb Research or Applied Materials or ASML or their

[36:48] or Applied Materials or ASML or their further downstream supply chains like

[36:50] further downstream supply chains like MKSI and and all these other companies

[36:52] MKSI and and all these other companies the tail whip just gets whipped harder

[36:54] the tail whip just gets whipped harder and harder and harder and ultimately

[36:57] and harder and harder and ultimately that's a shortage if you know TSMC wants

[36:58] that's a shortage if you know TSMC wants to spend $100 billion in 2028 which is a

[37:01] to spend $100 billion in 2028 which is a real possibility I think people would

[37:03] real possibility I think people would think that's insane but that's a real

[37:04] think that's insane but that's a real real possibility

[37:05] real possibility >> what about other parts of the chip

[37:06] >> what about other parts of the chip ecosystem where GPUs have been

[37:08] ecosystem where GPUs have been completely dominant what about like CPUs

[37:10] completely dominant what about like CPUs or AS6 or things that start to pop out

[37:12] or AS6 or things that start to pop out as both opportunities and bottlenecks

[37:14] as both opportunities and bottlenecks beyond just like Nvidia's GPU dominance.

[37:17] beyond just like Nvidia's GPU dominance. >> Yeah, I mean AS6 are obviously taking

[37:18] >> Yeah, I mean AS6 are obviously taking off, but I'll sort of pivot away from AI

[37:20] off, but I'll sort of pivot away from AI chips to talk about these other things.

[37:22] chips to talk about these other things. There's a project we did on FPGAAS and

[37:24] There's a project we did on FPGAAS and there turns out there's 120 FPGAs per

[37:26] there turns out there's 120 FPGAs per per um next generation rack um AI rack

[37:30] per um next generation rack um AI rack and then like what about all the FPGA

[37:32] and then like what about all the FPGA names CPU wise all these reinforcement

[37:34] names CPU wise all these reinforcement learning environments plus all the slop

[37:36] learning environments plus all the slop code you and I are generating that is

[37:37] code you and I are generating that is now running on some you know Versell

[37:39] now running on some you know Versell instance or whatever it is um or some

[37:42] instance or whatever it is um or some AWS instant or some bucket that we've

[37:43] AWS instant or some bucket that we've spun up all of that requires CPU and so

[37:46] spun up all of that requires CPU and so CPUs are completely sold out and demand

[37:48] CPUs are completely sold out and demand is skyrocketing there

[37:49] is skyrocketing there >> help people understand the role that CPU

[37:51] >> help people understand the role that CPU plays and everything.

[37:52] plays and everything. >> Yeah. So there's two there's two main

[37:53] >> Yeah. So there's two there's two main reasons why you need tons of CPUs. One

[37:55] reasons why you need tons of CPUs. One is when you're doing reinforcement

[37:56] is when you're doing reinforcement learning um the CPU is very critical to

[38:00] learning um the CPU is very critical to that. So so before you would throw all

[38:02] that. So so before you would throw all the internet's data into the model,

[38:03] the internet's data into the model, train it, spit it spits and it it spits

[38:05] train it, spit it spits and it it spits some stuff out. Now you train all the

[38:07] some stuff out. Now you train all the world's internet you put all the

[38:08] world's internet you put all the internet data into the model. Then you

[38:10] internet data into the model. Then you put it in this environment. This

[38:11] put it in this environment. This environment is like hey model try this

[38:13] environment is like hey model try this out and it tries stuff out. It tries a

[38:15] out and it tries stuff out. It tries a bunch of different things and in the end

[38:17] bunch of different things and in the end there is an environment which scores

[38:20] there is an environment which scores whether or not what it tried out is

[38:22] whether or not what it tried out is successful and it grades it. And these

[38:23] successful and it grades it. And these environments can be anything. It can be,

[38:25] environments can be anything. It can be, hey, check if the text was outputed in

[38:27] hey, check if the text was outputed in the right way, structured outputs. It

[38:29] the right way, structured outputs. It can be very simple stuff. It can be very

[38:30] can be very simple stuff. It can be very complex stuff. Um, and people are

[38:32] complex stuff. Um, and people are starting to get into very complex

[38:33] starting to get into very complex things, right? Like, hey, I want you to

[38:36] things, right? Like, hey, I want you to open this file, change it, edit it,

[38:38] open this file, change it, edit it, update it, submit it to this website. I

[38:40] update it, submit it to this website. I want you to open up this physics

[38:41] want you to open up this physics simulation from Seammens and edit this

[38:43] simulation from Seammens and edit this CAD model. So the environments can get

[38:45] CAD model. So the environments can get more and more complex and those

[38:46] more and more complex and those environments run on CPUs. They don't run

[38:48] environments run on CPUs. They don't run on GPUs. They don't run on AS6. The AS6

[38:50] on GPUs. They don't run on AS6. The AS6 run the model that takes the input data

[38:53] run the model that takes the input data from the environment, runs it through

[38:55] from the environment, runs it through the model. The model creates outputs of

[38:57] the model. The model creates outputs of various different trajectories, right?

[38:59] various different trajectories, right? Ways that it think it could solve it um

[39:01] Ways that it think it could solve it um in different instances. those

[39:04] in different instances. those trajectories are graded slashscored and

[39:06] trajectories are graded slashscored and the ones that are successful you train

[39:07] the ones that are successful you train on and you update and you reiterate and

[39:10] on and you update and you reiterate and you iterate iterate iterate and so CPUs

[39:12] you iterate iterate iterate and so CPUs are very useful for that one and then

[39:14] are very useful for that one and then once you have these great models and

[39:16] once you have these great models and you're deploying them those models are

[39:17] you're deploying them those models are generating code they're generating

[39:19] generating code they're generating useful output that useful output it

[39:21] useful output that useful output it doesn't go from a GPU straight to the

[39:23] doesn't go from a GPU straight to the human brain um it goes from a GPU or an

[39:26] human brain um it goes from a GPU or an ASIC through to you know a deployed app

[39:29] ASIC through to you know a deployed app that you're deploying somewhere that

[39:30] that you're deploying somewhere that actually just runs on CPUs so that's

[39:32] actually just runs on CPUs so that's another area where there's a lot of

[39:33] another area where there's a lot of demand and and things are sold out um in

[39:36] demand and and things are sold out um in a large large way.

[39:37] a large large way. >> As you continue to assess and try to be

[39:39] >> As you continue to assess and try to be the world's best informed person on both

[39:41] the world's best informed person on both the trajectory of supply and demand,

[39:43] the trajectory of supply and demand, what are things that you wish you knew

[39:45] what are things that you wish you knew to make that understanding that you

[39:47] to make that understanding that you don't know?

[39:47] don't know? >> I think the hardest area for us um and

[39:52] >> I think the hardest area for us um and for everyone is understanding

[39:54] for everyone is understanding tokconomics, economics of tokens. Um, I

[39:56] tokconomics, economics of tokens. Um, I think we have a really tremendously like

[39:58] think we have a really tremendously like good insight into how much it costs to

[40:00] good insight into how much it costs to run infrastructure, what the cost of

[40:02] run infrastructure, what the cost of tokens are, what the cost of models are,

[40:04] tokens are, what the cost of models are, what the margins of these labs are, but

[40:06] what the margins of these labs are, but the usage and adoption is what's really

[40:08] the usage and adoption is what's really difficult to model, you know,

[40:10] difficult to model, you know, continuously, right? We we have these

[40:12] continuously, right? We we have these like we had like crazy in January, we

[40:13] like we had like crazy in January, we had crazy estimates for February,

[40:14] had crazy estimates for February, anthropic smashed them. How do we

[40:16] anthropic smashed them. How do we calibrate this model? What are the data

[40:17] calibrate this model? What are the data sources for this? February, uh, we had

[40:20] sources for this? February, uh, we had crazy assumptions for March and then

[40:21] crazy assumptions for March and then they smashed them. And everyone sees the

[40:23] they smashed them. And everyone sees the number of 10 billion and they're like

[40:24] number of 10 billion and they're like what the how do they add 10 billion in

[40:26] what the how do they add 10 billion in revenue? Who is using all these tokens?

[40:28] revenue? Who is using all these tokens? Why are they using them? What are they

[40:29] Why are they using them? What are they building with them? And then more

[40:30] building with them? And then more importantly with what they're building

[40:32] importantly with what they're building with these tokens, how is that actually

[40:34] with these tokens, how is that actually diffusing into the economy? And what

[40:35] diffusing into the economy? And what value is that generating? Because it's

[40:37] value is that generating? Because it's not really something that you can

[40:38] not really something that you can capture in any any GDP statistic, right?

[40:41] capture in any any GDP statistic, right? all of the value of the tokens that I

[40:43] all of the value of the tokens that I use get transformed into better

[40:44] use get transformed into better information which I then sell at a

[40:47] information which I then sell at a discount to what people used to sell

[40:48] discount to what people used to sell information for relatively because um

[40:51] information for relatively because um and therefore that information is now

[40:53] and therefore that information is now making its way throughout the economy

[40:55] making its way throughout the economy and and people are making better

[40:56] and and people are making better investment decisions or better

[40:59] investment decisions or better competitive decisions if they're a semi

[41:00] competitive decisions if they're a semi company or data center company or

[41:02] company or data center company or hyperscaler and now how how much what

[41:04] hyperscaler and now how how much what what is the value of this and what has

[41:05] what is the value of this and what has that what has that done to the economy

[41:07] that what has that done to the economy it's clearly by every subjective metric

[41:10] it's clearly by every subjective metric amazing Amazing. But where is the

[41:12] amazing Amazing. But where is the phantom GDP? What is the phantom GDP?

[41:15] phantom GDP? What is the phantom GDP? How do we track the real economic?

[41:17] How do we track the real economic? Because because the GDP metrics are not,

[41:20] Because because the GDP metrics are not, you know, accurate if you were to say

[41:22] you know, accurate if you were to say what is the GDP that Dylan Patel is

[41:23] what is the GDP that Dylan Patel is making. It's tiny compared to what the

[41:25] making. It's tiny compared to what the value that I think is being created. And

[41:27] value that I think is being created. And so ultimately, what is the value being

[41:29] so ultimately, what is the value being created by these tokens? Not on a basis

[41:31] created by these tokens? Not on a basis of, you know, just simple, you know,

[41:33] of, you know, just simple, you know, what is the knock-on effect, right? What

[41:35] what is the knock-on effect, right? What is the knock-on effect of all the things

[41:36] is the knock-on effect of all the things that these things are doing? I think

[41:37] that these things are doing? I think that's the real uh question and

[41:39] that's the real uh question and challenge uh that's hard to measure. I

[41:42] challenge uh that's hard to measure. I think we've got a tremendous, you know,

[41:44] think we've got a tremendous, you know, reading on the supply side of things. I

[41:46] reading on the supply side of things. I think we've got a tremendous reading on

[41:47] think we've got a tremendous reading on even a lot of the demand side signals,

[41:49] even a lot of the demand side signals, but it's it's what is the value these

[41:50] but it's it's what is the value these tokens are generating. That's hard to

[41:52] tokens are generating. That's hard to quantify and measure.

[41:54] quantify and measure. >> I hope we get a chance to do this like

[41:55] >> I hope we get a chance to do this like every 3 months because this changes so

[41:56] every 3 months because this changes so quickly. What do you think is going to

[41:58] quickly. What do you think is going to happen next? Like when I when I come

[42:00] happen next? Like when I when I come back 3 months from now and we're in San

[42:01] back 3 months from now and we're in San Francisco together again, what do you

[42:02] Francisco together again, what do you expect?

[42:03] expect? >> Large scale protests.

[42:05] >> Large scale protests. >> Really?

[42:05] >> Really? >> Yeah. Yeah, I think there will be a

[42:06] >> Yeah. Yeah, I think there will be a large scale protest against anthropic

[42:09] large scale protest against anthropic >> and open AI.

[42:09] >> and open AI. >> Expand on that a little more.

[42:10] >> Expand on that a little more. >> Um, people hate AI. Um, AI is less

[42:14] >> Um, people hate AI. Um, AI is less popular than ICE, less popular than

[42:17] popular than ICE, less popular than politicians. Confused how Pew surveyed

[42:19] politicians. Confused how Pew surveyed this, but apparently AI is less popular

[42:21] this, but apparently AI is less popular than politicians. You know, with

[42:22] than politicians. You know, with Enthropic adding so much revenue, that's

[42:25] Enthropic adding so much revenue, that's going to start causing business changes

[42:26] going to start causing business changes downstream. People are going to get more

[42:28] downstream. People are going to get more and more scared of AI. they'll start

[42:30] and more scared of AI. they'll start blaming more and more of their own

[42:31] blaming more and more of their own problems and things that are, you know,

[42:34] problems and things that are, you know, global, you know, have been deep-seated

[42:36] global, you know, have been deep-seated problems for a long time. Those will

[42:37] problems for a long time. Those will bubble up and be blamed on AI. Um,

[42:41] bubble up and be blamed on AI. Um, probably some politician or some social

[42:43] probably some politician or some social media people will start to be able to

[42:44] media people will start to be able to take uh influencer will be able to start

[42:46] take uh influencer will be able to start taking and weaponizing AI against

[42:49] taking and weaponizing AI against people. You look at the comments of news

[42:51] people. You look at the comments of news articles where Sam Alman had a Molotov

[42:53] articles where Sam Alman had a Molotov cocktail thrown in his house twice in

[42:55] cocktail thrown in his house twice in like two weeks. They're like, people are

[42:57] like two weeks. They're like, people are cheering it on. Uh, and this is just the

[43:00] cheering it on. Uh, and this is just the beginning. So, I think I think we'll see

[43:01] beginning. So, I think I think we'll see large scale protests against AI in three

[43:03] large scale protests against AI in three months.

[43:04] months. >> What is the counterwe to that? Like, how

[43:06] >> What is the counterwe to that? Like, how should the AI industry head that off?

[43:08] should the AI industry head that off? >> First of all, Sam Alman and Dario have

[43:10] >> First of all, Sam Alman and Dario have to stop getting on interviews. They're

[43:11] to stop getting on interviews. They're so uncarismatic.

[43:13] so uncarismatic. I don't know what they're doing. Every

[43:15] I don't know what they're doing. Every interview they do is like, wow, normal

[43:17] interview they do is like, wow, normal people are going to hate you even more.

[43:19] people are going to hate you even more. Like, Sam being on Tucker Carlson

[43:21] Like, Sam being on Tucker Carlson probably made all Republicans hate

[43:22] probably made all Republicans hate OpenAI. And same with Dario. They just

[43:24] OpenAI. And same with Dario. They just have no charisma. I think that's first.

[43:26] have no charisma. I think that's first. Two, they need to start showing

[43:28] Two, they need to start showing uplifting things that can be done with

[43:30] uplifting things that can be done with AI. Um, three, they need to stop talking

[43:32] AI. Um, three, they need to stop talking about how the capabilities are going to

[43:34] about how the capabilities are going to change the whole world constantly

[43:35] change the whole world constantly because then people are going to get

[43:36] because then people are going to get fear of that capability because they

[43:38] fear of that capability because they have no connection.

[43:39] have no connection. >> They don't know how to use it. Yeah.

[43:40] >> They don't know how to use it. Yeah. >> There's no connection to it either. It's

[43:41] >> There's no connection to it either. It's like the average person doesn't know an

[43:43] like the average person doesn't know an anthropic employee. The average person

[43:45] anthropic employee. The average person doesn't know an open eye employee.

[43:46] doesn't know an open eye employee. average person doesn't know who these

[43:48] average person doesn't know who these people are, what their goals are, and

[43:49] people are, what their goals are, and they just view them as like this like

[43:51] they just view them as like this like sneaky cobball of like 5,000 people at

[43:54] sneaky cobball of like 5,000 people at this company that are going to change

[43:54] this company that are going to change the world and automate all the jobs and

[43:56] the world and automate all the jobs and and destroy society. That's what they

[43:58] and destroy society. That's what they view it as. And and as people who are

[44:00] view it as. And and as people who are funding the building of all these data

[44:02] funding the building of all these data centers and and power plants that are

[44:04] centers and and power plants that are going to pollute the world, right? They

[44:05] going to pollute the world, right? They don't quite understand what's happening.

[44:06] don't quite understand what's happening. You know, they have to stop talking

[44:07] You know, they have to stop talking about the future thing that's going to

[44:09] about the future thing that's going to happen and only talk about present, how

[44:10] happen and only talk about present, how uplifting AI is. I think it's a huge

[44:13] uplifting AI is. I think it's a huge reorg and rebranding that needs to be

[44:14] reorg and rebranding that needs to be done.

[44:15] done. >> I love doing this with you. Thanks for

[44:16] >> I love doing this with you. Thanks for your time.

[44:16] your time. >> Awesome. Thanks.

[44:22] >> Your finance team isn't losing money on

[44:24] >> Your finance team isn't losing money on big mistakes. It's leaking through a

[44:25] big mistakes. It's leaking through a thousand tiny decisions nobody's

[44:27] thousand tiny decisions nobody's watching. Ramp puts guard rails on

[44:29] watching. Ramp puts guard rails on spending before it happens. Real-time

[44:31] spending before it happens. Real-time limits, automatic rules, zero

[44:32] limits, automatic rules, zero firefighting. Try it at ramp.com/invest.

[44:37] firefighting. Try it at ramp.com/invest. As your business grows, Vant scales with

[44:39] As your business grows, Vant scales with you, automating compliance, and giving

[44:40] you, automating compliance, and giving you a single source of truth for

[44:42] you a single source of truth for security and risk. Learn more at

[44:44] security and risk. Learn more at vanta.com/invest.

[44:47] Ridgeline is redefining asset management

[44:49] Ridgeline is redefining asset management technology as a true partner, not just a

[44:51] technology as a true partner, not just a software vendor. They've helped firms 5x

[44:53] software vendor. They've helped firms 5x and scale, enabling faster growth,

[44:55] and scale, enabling faster growth, smarter operations, and a competitive

[44:57] smarter operations, and a competitive edge. Visit ridgelineapps.com

[44:59] edge. Visit ridgelineapps.com to see what they can unlock for your

[45:01] to see what they can unlock for your firm.

[45:02] firm. Every investment firm is unique, and

[45:04] Every investment firm is unique, and generic AI doesn't understand your

[45:05] generic AI doesn't understand your process. Rogo does. It's an AI platform

[45:08] process. Rogo does. It's an AI platform built specifically for Wall Street,

[45:09] built specifically for Wall Street, connected to your data, understanding

[45:11] connected to your data, understanding your process, and producing real

[45:12] your process, and producing real outputs. Check them out at

[45:14] outputs. Check them out at rogo.ai/invest.

[45:17] The best AI and software companies from

[45:19] The best AI and software companies from OpenAI to Cursor to Perplexity use work

[45:21] OpenAI to Cursor to Perplexity use work OS to become enterprise ready overnight,

[45:23] OS to become enterprise ready overnight, not in months. Visit works.com to skip

[45:26] not in months. Visit works.com to skip the unglamorous infrastructure work and

[45:28] the unglamorous infrastructure work and focus on your product.

The Supply and Demand of AI Tokens | Dylan Patel Interview

https://www.youtube.com/watch?v=LF3aUIM57uw

Summary

TL;DR — The AI token market is experiencing explosive demand, significantly outpacing supply, leading to skyrocketing costs and driving innovation as companies race to leverage the latest models. This intense demand is reshaping industries and forcing businesses to adapt rapidly to remain competitive.

Key points

• The cost of AI tokens has surged dramatically, with companies like Semi Analysis seeing their annual spend jump from tens of thousands to millions of dollars, now representing a significant portion of their operational budget.

• AI adoption is democratizing complex tasks, enabling individuals with no prior coding experience to perform sophisticated analyses and create applications that previously required large teams and extensive time.

• The rapid advancement of AI models, exemplified by Anthropic's rapid release cycles and the emergence of highly capable models like Mythos, is creating a continuous demand for the most cutting-edge technology.

• Supply chain constraints, particularly for GPUs and memory, are exacerbating the token shortage, leading to increased prices and extended lead times, which in turn drives up margins for hardware providers.

• The economic impact of AI is profound but difficult to quantify, creating

摘要 / Summary (zh-TW)

簡而言之 — 人工智慧代幣市場正經歷爆炸性的需求，遠遠超過供應，導致成本飆升，並推動創新，因為公司競相利用最新的模型。這種強烈的需求正在重塑各行各業，並迫使企業迅速適應以保持競爭力。

重點

• 人工智慧代幣的成本急劇上漲，像 Semi Analysis 這樣的公司年支出從數萬美元飆升至數百萬美元，現在已佔其營運預算的很大一部分。

• 人工智慧的普及正在使複雜任務民主化，讓沒有程式設計經驗的個人能夠執行複雜的分析並創建以前需要龐大團隊和大量時間的應用程式。

• 人工智慧模型的快速發展，例如 Anthropic 的快速發布週期以及 Mythos 等高度先進模型的出現，正在創造對最新技術的持續需求。

• 供應鏈限制，特別是 GPU 和記憶體方面，加劇了代幣短缺，導致價格上漲和交貨時間延長，進而提高了硬體供應商的利潤。

• 人工智慧的經濟影響是巨大的，但難以量化，創造了「