# Watts, Wafers, and the Future of AI Infra | Gavin Baker

https://www.youtube.com/watch?v=Mmj_G9RlW-I

[00:00] What was happening in AI was I think the most extraordinary moment in the history of capitalism, the history of American business.
[00:07] Anthropic they added 11 billion of ARR.
[00:11] The three highest profile SAS companies founded in the last 10 12 years are Palunteer, Snowflake and Data Bricks.
[00:21] And these three companies spent 10 years building their businesses.
[00:23] Anthropic added their combined businesses in one month.
[00:30] That's just nothing like that has ever happened in the history of capitalism.
[00:35] Forget my career.
[00:35] Just the flatout history of capitalism, the history of business.
[00:53] All right.
[00:53] So, this is our uh sixth time doing this, if you can believe it, which puts you back into first place uh or at least tied for first place with Girly.
[01:01] And I think even since last time when we did this, which was so exciting and spectacular, I think we're in an even more interesting time now.
[01:10] Maybe just start by riffing on how it felt for you living through March and April of this year, which which felt to me just like a completely unique economic, technology, and market environment.
[01:22] and you're the biggest student of of of history and of these times.
[01:24] So what does it feel like?
[01:25] I would say broadly speaking there are two kinds of draw downs.
[01:29] They're drawowns where you're wrong, a company misestimates, your hypothesis was invalidated and you have to take your medicine and you crystallize that loss.
[01:41] And then there are draw downs or periods of underperformance where you you're underperforming because of companies you know really really well and where you profoundly disagree with the price action and you can lean in and instead of crystallizing uh negative performance you can kind of build pent up alpha pent up future performance and for me that is what
[02:03] March felt like.
[02:05] It felt like uh you know the NASDAQ was selling off and at the same time what was happening in AI was I think the most extraordinary moment in the history of capitalism the history of American business and what I just mean by that is that anthropic they added 11 billion of ARR and what is astonishing to me about this is that the SAS and cloud revolution it created we'll call it between 5 and 10 trillion dollars of value and I would say arguably the three highest profile SAS companies to have kind of been founded in the last 10 12 years are Palunteer, Snowflake and Data Bricks.
[02:44] And these three companies have spent employ thousands of people, tens of thousands collectively.
[02:51] They've all spent 10 years building their businesses.
[02:54] And Anthropic added their combined businesses in one month.
[03:00] That's just nothing like that has ever
[03:04] happened in the history of capitalism.
[03:06] Forget my career.
[03:09] Just the flatout history of capitalism, the history of business.
[03:12] I wild.
[03:15] And then Krishna comes on this show and shares some stats.
[03:18] 500% in DR.
[03:19] Yeah.
[03:20] You do the math on that for three years.
[03:22] Insanity.
[03:25] We So there's just no precedent for this.
[03:28] And we, you know, tech tech investors, you hear a lot of discussions about S-curves and investing in exponentials.
[03:32] I've just never seen an exponential like this.
[03:34] It felt even more extreme than Deepseek,
[03:39] which was a very similar setup.
[03:41] If we go back to 25
[03:43] and there's a huge sell-off on Deepseek,
[03:45] which was very strange because the paper gets published 7 days before Deepseek Monday.
[03:52] got published,
[03:54] I believe, on a Monday that was a holiday in America.
[03:56] And I read it, I thought, hm, you know, this this feels like it might not read that positively for um you know, the AI
[04:07] Trade. You I took action.
[04:10] We had DeepSeek Monday where AI really imploded a week later and that was really strange.
[04:18] Because by DeepSec Monday it was super clear that this was going to be the most positive thing that had ever happened to compute demand.
[04:25] Prices in the AWS available availability zones in Asia had already like doubled.
[04:31] You were seeing GPU availability go down.
[04:37] And this was just the first time we saw how much more compute-hungry reasoning models are during inference than non-reasoning models.
[04:46] And so that was a similar setup, but you you had to do some work to see that.
[04:51] I mean, it's not that hard to say, oh wow, stocks are selling off.
[04:55] The price of DRAM is going vertical.
[04:57] The price of GPUs in Asia going vertical.
[05:00] U GPU availability is going down.
[05:02] And then like two or three days later, you know, GPU prices in in America started going up, GPU rental prices.
[05:07] All you had to do
[05:09] In in March was just simply observe what was happening to Anthropic.
[05:14] There there's all these people who seem to regret, you know, not buying during 22, not buying during COVID, not buying during Deep Seek.
[05:22] You had the same valuation setup at the beginning of April and an even clearer AI inflection.
[05:32] And so there have been all these chances to buy into AI and then of course what complicated it was the straight of foremost I became a believer and am a believer that I think maybe one thing that the market was mispricing and I'm I'm no macro expert I do do a lot of professional security investing.
[05:53] And so I do have access to people who are experts and are excited to share their thoughts and opinions with me that the straight of horm being closed is actually relatively awesome for America.
[06:07] Why?
[06:08] Because particularly for the goals of
[06:10] the current administration.
[06:12] So electricity is a very important industrial or manufacturing input.
[06:17] The key input into American electricity prices which feeds into AI is in G1 natural gas on Bloomberg that was down 20%.
[06:26] And natural gas in Asia, Europe, everywhere else doubled or tripled.
[06:35] So our relative manufacturing competitiveness improved overnight and for better or worse that is what the Trump administration seems to care about.
[06:46] They are very focused on America's relative position.
[06:48] And I think a lot of people had memories of the 1970s.
[06:53] And what made the 70s so traumatic was it wasn't just that prices went up, it's that there were actual gas shortages.
[07:01] And then you go through, okay, well the US economy is dramatically less energy intensive than it was.
[07:07] US econ the United States is now the world's largest producer of oil and gas and we've become
[07:12] now the world's largest exporter of oil and gas and then on top of that there's this relative manufacturing advantage.
[07:21] and so that made it I think easier to stay focused on AI fundamentals stay focused on what were historically attractive valuations.
[07:33] I think on a relative basis tech essentially got as cheap as it's been versus the rest of the market has at any point over the last 10 years and just think about that in the context of market efficiency.
[07:45] We have the most extraordinary moment in the history of capitalism that's wildly bullish for AI and you get a chance to buy AI at really attractive valuation.
[07:56] What do you make of the multiples that specifically Anthropic and OpenAI, which in my mind are like the reference assets that are the most pure play takes on this trend really being not that crazy?
[08:10] Like if you just look at the sales multiple and compare it to maybe what
[08:13] data bricks and snowflake and these companies traded at at their peak like how do you make sense of it?
[08:18] I do think OpenAI and Enthropic are pretty different animals from a capital efficiency perspective.
[08:21] And Enthropic clearly is has a dramatically lower cost per token than OpenAI.
[08:26] They just do.
[08:30] And you can just see that in the amount of money that they have burned to get to a roughly similar revenue scale.
[08:37] I think have have they burned maybe 80% less than OpenAI.
[08:41] So as businesses, they clearly have very different structural ROIC's.
[08:46] I think OpenAI is doing a lot.
[08:48] I think Sarah Frier is one of the most exceptional CFOs.
[08:49] I think they're doing a lot of things to try to improve this and they've secured a lot of compute more more than they've secured a lot of compute.
[08:57] That's another big difference.
[08:59] Um it turns out being aggressive really paid but yeah I just anthropic at 900 billion for 50 billion and you know ARR and you know growing a thousand%.
[09:10] Yeah, growing at ridiculous rates.
[09:13] Maybe a true statement is that if Anthropic
[09:15] had all the compute, they'd probably be doing well north of hundred billion dollars today, maybe 150.
[09:25] And I do, you know, they have clearly deprecated the intelligence of Claude.
[09:28] There's an analysis Claude is even on Opus is generating 70% less tokens for the exact same question.
[09:35] And you know, as we talked about last time, token quantity equals quality of answer and quality of thinking at some level.
[09:39] you know and there is an intelligence density per token that also matters you know I think I felt that as as a user so I think they would be doing materially more 100 150 maybe 200 billion so you might be buying it at more like five times unconstrained I'm going to make up a new number urr unconstrained run rate revenue yes
[10:09] why do you think they don't raise $und00 billion at a $3 trillion valuation or something like this.
[10:14] Like if you were
[10:16] The Anthropic CFO, uh Krishna is awesome.
[10:18] We just had him on.
[10:19] Or if you're the open, if you're Sarah, certainly if if the inbound I received following the Krishna episode is any indication, everyone I've ever met is trying to invest in in both these companies.
[10:29] So I think it's wise it the future is uncertain.
[10:36] You are clearly in a very capital intensive game even if you are, you know, Enthropic.
[10:43] Um, I'm sure is at very positive gross margins on inference today.
[10:45] I think probably starts generating cash this year if they are not already generating cash, which I think is probably the case.
[10:55] But still, you probably want to be able to raise more capital, access more compute.
[10:59] The world is uncertain, Ukraine is starting to really, really win.
[11:01] How is Russia going to respond?
[11:04] And, you know, I think there's still a lot of uncertainty in Iran.
[11:07] All this uncertainty, I think, probably amplifies geopolitical uncertainty over Taiwan.
[11:12] So, it's an uncertain world.
[11:14] If if I think about
[11:16] Elon, Elon has always made investors money.
[11:19] He treats it like a sacred covenant.
[11:21] And as a result, because he's made people money for now 20 years, he has a superpower.
[11:27] And that is he can essentially raise as much capital as he wants, whenever he wants.
[11:32] And I think it's wise that these companies are taking I don't know if that's how they think about it, but I do think being focused on making investors money is wise and creates benefits that don't just last for like a year or two.
[11:45] They can last for the next 20 to 30 years.
[11:54] And the way Elon did this was sort of systematically underpricing SpaceX or whatever else.
[11:59] Like what is the actual method?
[12:02] Just never being greedy on valuation, never pushing valuation.
[12:06] Just that simple.
[12:08] You know, my friend Antonio pointed out SpaceX compounded, you know, low 30% per year for whatever that was a decade.
[12:12] And and that was just because Elon was, I
[12:18] think, focused on preserving the superpower and having trying to strike a fair balance between investors and employees.
[12:26] But I I think it's wise.
[12:26] But could Anthropic raise money at probably at least a 100% premium to this rumored latest mark?
[12:38] Of course, most software companies try to maximize your time on their app to juice engagement.
[12:41] RAMP does the exact opposite.
[12:43] RAMP understands that no one wants to spend hours chasing receipts, reviewing expense reports, and checking for policy violations.
[12:49] So, they built their tools to give that time back using AI to automate 85% of expense reviews with 99% accuracy.
[12:56] And since Ramp saves companies 5%, it's no wonder that Shopify runs on RAM, Stripe runs on RAM, and my business does, too.
[13:03] To see what happens when you eliminate the busy work, check out ramp.com/invest.
[13:10] Felix by Rogo is a personal finance agent that turns a single prompt into finished client ready work using your firm's own templates, context, and standards.
[13:17] Send Felix an email like,
[13:19] Take these comments and turn them for me or update my tracker with the context of these emails.
[13:25] Or, Run the ability to pay math on this buyer and Felix sends back finished PowerPoint decks, Excel models, and sourced research.
[13:31] Felix works the way your team already does, delivering work quickly and accurately around the clock.
[13:36] Learn more at rogo.ai/felix.
[13:40] OpenAI, Cursor, Enthropic, Perplexity, and Versell all have something in common.
[13:44] They all use WorkOS.
[13:47] And here's why.
[13:49] To achieve enterprise adoption at scale, you have to deliver on core capabilities like SSO, skim, arbback, and audit logs.
[13:53] That's where work OS comes in.
[13:55] Instead of spending months building these mission critical capabilities yourself, you can just use work OS APIs to gain all of them on day zero.
[14:02] That's why so many of the top AI teams you hear about already run on work OS.
[14:07] Work OS is the fastest way to become enterprise ready and stay focused on what matters most, your product.
[14:13] Visit works.com to get started.
[14:15] Let's get to the Watson wafers part of the discussion.
[14:17] Always my favorite thing to talk about with you.
[14:19] Uh the importance
[14:21] of this infrastructure buildout.
[14:24] I feel like every time I feel like it's getting overheated and then the next time I talk to you, it seems like we should have done way more than we did.
[14:31] And you've studied S-curves and the steepness of those S-curves a lot.
[14:33] Uh and you know a lot about history.
[14:35] Talk us through how you're thinking about Watson wafers today as the key to inputs into this whole thing.
[14:42] Yeah, I would say I think capitalism is going to solve the Watts shortage.
[14:49] absent big regulatory political blowback which I think is a real possibility.
[14:53] the head of kind of data center infrain investing at one of the big PE firms.
[14:58] You know, I think Blackstone, Apollo, KKR said it used to be energy and chips were our biggest gating factors.
[15:07] Now it's zoning and approval much more important.
[15:10] And I think a lot of companies are waiting till after the midterms to take action in terms of maybe workforce reductions.
[15:17] Nobody wants to be, you know, piñata during the
[15:22] midterms,
[15:24] but you know, you've seen a lot of companies that make turbines significant announce of plans to significantly increase capacity.
[15:30] There's like two of these machines that can cast these big blades.
[15:34] We haven't made one in 80 years in the West.
[15:37] We don't know how to make them anymore, etc., etc., etc.
[15:42] All of that is true.
[15:42] And I and and by no means am I minimizing, you know, the industrial engineering, you know, magic and artistry that goes into those, but capitalism is very good at solving problems like these over time.
[15:54] There's other sources of energy besides these turbines with a longer time frame.
[15:56] So I think the watts shortage will probably begin to alleviate 27 28 and then I think orbital compute will really solve that.
[16:08] And I do I do want to like reframe orbital compute because I think when people hear data centers in space they which we discussed on our last episode they picture a pentagon sized building in space.
[16:19] They're like well we can't do
[16:23] That. That's not what it is.
[16:28] A blackwell rack weighs you 3,000 lbs.
[16:31] It's 8 ft high.
[16:31] It's 4 feet deep.
[16:31] 3 feet wide.
[16:35] It's racks in space.
[16:38] And SpaceX has showed you an illustration and it's a rack.
[16:41] That's the satellite.
[16:44] Uh but it's probably about the size of a blackwell rack.
[16:46] It has these solar wings that are probably 500 ft long on each side.
[16:48] You keep it in a sunsynchronous orbit.
[16:51] So those solar panels are always in the sun.
[16:54] And then because it's in an exactly sunsynchronous orbit, the radiator which extends behind it for hundreds of feet.
[17:05] This is a common criticism.
[17:05] Yeah.
[17:06] how you going to go.
[17:07] I've spent a lot of time at Starbase over the years and I've talked to a lot of SpaceX engineers and I do think it is the most talented group of engineers on planet Earth and they're very confident they have solved this and they're not
[17:23] always confident like I think probably.
[17:27] you know there's some engineering that needs to happen to turn the starship into a Mars colonial transporter.
[17:30] Will they do that?
[17:32] Absolutely.
[17:34] What are they more focused on?
[17:37] I would say probably, you know, the repair and maintenance.
[17:39] These are the two big, you know, the two big responses, the radiator and and how do you repair the whatever issue goes wrong in the rack.
[17:45] And the answer is like until you have probably an, you know, floating optimuses.
[17:50] You don't.
[17:52] Now, I do think Starship is going to change the space economy in ways we cannot imagine.
[17:56] Particularly if regulation becomes a constraint to data centers, none of it's going to matter.
[17:59] you're going to sell as much orbital compute as you can make.
[18:04] And then obviously you link these racks using lasers traveling through vacuum which are already on every Starlink.
[18:10] And it's just it's just mindblowing to me that SpaceX operates the world's largest satellite fleet which is like 98 or 99% of all satellites in orbit.
[18:17] Every
[18:24] Starlink they're cooling it today.
[18:28] And you know, I think Starlink V3 is going to operate at 20 kW.
[18:33] A Blackwell rack is only 100 kows.
[18:37] And people talk a lot about density.
[18:39] Well, if you're connecting the racks with lasers through vacuum, you know, you can make the rack bigger physically.
[18:45] You're focused on weight, not size.
[18:49] In a data center on Earth where you're trying to connect racks, ideally using copper, minimize lengths, etc., etc.
[18:56] Cabling is a big cost.
[18:59] um you do want that rack to be small because you know copper when you can, optics when you must.
[19:03] But in space, you know, there's all sorts of things that SpaceX can do that I think maybe some of these naysayers are not contemplating, but it's just they operate more satellites than you want.
[19:12] They have a 20 kow satellite today, so maybe you just scale that up to 60 kilowatts to start.
[19:18] They seem very confident they're going to go right to 100 to 120.
[19:23] And they also the same company now also operates the largest
[19:26] data center on earth.
[19:28] They have the world's best hardware engineers and all sorts of people almost all of whom are not smart enough or practical enough to work at SpaceX are these armchair skeptics.
[19:42] You know, I don't want to quote Larry Ellison, but somebody was, you know, being skeptical and Larry and Larry was just like, "Listen, he's out there landing rockets.
[19:51] I don't see anybody else landing rockets.
[19:53] And the reality is is that 10 years later, no other company is consistently capable of landing and fully reusing an orbital rocket.
[20:03] And none of this works makes sense without reusability.
[20:05] That means you have to land it.
[20:06] I would like to redefine orbital compute has racks in space, not giant floating pentagoniz data centers in space, which is just, you know, that's silly.
[20:19] But you can, you know, what makes a data center is you're connecting these racks with lasers.
[20:23] So it'll be racks in space that are connected with lasers into a virtual data center.
[20:26] And and if you think about
[20:28] that state of the world, let's say that
[20:31] all happens and we're really good at
[20:32] getting these things up economically,
[20:34] running matrix multiplication all over
[20:36] space. What does that mean for
[20:37] terrestrial data centers? Someone once
[20:40] said, um, you know, America was going to
[20:44] suck as hard as it can on every energy
[20:47] source it can get. And I just think the
[20:49] same is true of compute.
[20:51] >> It's why I'm probably less worried about
[20:53] like an edge AI barecase than I was.
[20:57] >> We're going to consume as much compute
[21:02] as we can. And inference I think is very
[21:06] sensible for orbital compute. Training
[21:09] will be done on Earth for a long time.
[21:12] So I don't think that this is super
[21:14] bearish for terrestrial data centers. I
[21:16] think those are going to be valuable for
[21:18] my lifetime.
[21:21] But I do think if you are in this
[21:23] ecosystem of power production and
[21:26] cooling and you are massively ramping
[21:31] capacity and you know a lot of these
[21:33] capacity ramps are going to be hitting
[21:36] just as I think you know all of the
[21:38] silly skeptics start to understand that
[21:40] orbital compute is very real like I
[21:42] think it's worth thinking long and hard
[21:44] about that if you're one of those
[21:46] companies and then all sorts of cool
[21:48] stuff is happening in the interim you
[21:49] know we're getting really good at
[21:51] repurposing jet engines. You know,
[21:53] there's that boom aerospace that is
[21:54] doing this.
[21:55] >> So, there's a lot like capitalism is
[21:57] hard at work
[21:59] >> on on watts. On wafers, though, it's
[22:03] just this group of, you know, flinty
[22:08] older humans in Taiwan who are the most
[22:11] important humans in Taiwan. They are the
[22:14] overwhelming fraction of the country's
[22:15] GDP, water usage, electricity usage.
[22:18] They talk about the Silicon Shield. They
[22:21] all view themselves as inheritors of,
[22:25] you know, Morris Chain's sacred legacy.
[22:27] I vividly remember like visiting Science
[22:29] Park more than 20 years ago and, you
[22:34] know, talking to them. Do you think you
[22:35] could catch Intel? And they said, "This
[22:38] is such a beautiful dream, but it's a
[22:40] dream for our grandchildren."
[22:43] >> And they did it. partly because of
[22:45] Intel's self-inflicted wounds, but just
[22:48] they don't they think very differently.
[22:51] You know, one reason, you know, Jensen
[22:53] flies over there so much is he wants
[22:55] them to expand capacity. I do think it's
[22:57] wild that Jensen has never had a
[22:59] contract with Taiwan Semi. They do
[23:01] business on what seems fair in
[23:03] handshakes. Just fascinating. No
[23:05] contract. It's going to be fair over
[23:07] time. We're partners. We're going to be
[23:09] fair to each other. And the truth is,
[23:11] you know, based on every every prior
[23:15] market precedent for a foundational new
[23:17] technology like AI, you've always had a
[23:19] bubble. You know, Carleta Perez wrote
[23:21] this great book about this. And
[23:23] basically, markets are efficient. They
[23:25] correctly understand that this is a
[23:27] foundational technology.
[23:29] There's what Mobison calls a breakdown
[23:31] in diversity.
[23:33] Everyone becomes bullish on this new
[23:35] technology. And I am beginning to worry
[23:38] a little bit about a diversity
[23:39] breakdown. And then you get a bubble.
[23:43] That bubble funds the buildout of this
[23:45] new technology, but supply gets ahead of
[23:48] demand. And you get a crash and it's a
[23:51] particularly severe crash if it's a
[23:53] debtfueled buildout like the year 2000.
[23:56] And one thing really happy about really
[23:59] good about the current buildout is it's
[24:00] still overwhelmingly funded out of
[24:02] operating cash flows which is a a really
[24:05] important fundamental difference versus
[24:06] the year 2000 has is valuation has is
[24:09] the fact that every GPU is running at
[24:11] 100% utilization when 99% of fiber was
[24:14] unutilized. So there's all these
[24:15] fundamental differences, but we do have
[24:17] to history doesn't repeat, but it rhymes
[24:19] and and as investor, we have to be very
[24:21] cognizant of it
[24:23] and recognize that based on the last two
[24:26] or 30 hundred years, you know, forget
[24:28] the internet bubble. We had a railroad
[24:29] bubble, a canal bubble, we should expect
[24:31] a bubble. And that's terrifying. Like
[24:35] nobody wants a bubble. A bubble is
[24:37] terrible. reason it's terrible is if
[24:39] you're valuation sensitive, you like
[24:41] massively underperform. You get fired by
[24:44] probably all your clients. George
[24:46] Vanderhiden, who um is is is no longer
[24:49] with us, great port uh fidelity
[24:51] portfolio manager, he fought the bubble
[24:54] in 99 and he retired in two in early
[24:58] 2000 because I think he just couldn't
[24:59] couldn't take it.
[25:00] >> He knew it was wrong and you know his
[25:03] his clients were deeply skeptical.
[25:05] George, you're out of step. you know, he
[25:07] had he had white hair. He's truly great
[25:09] man. I I only overlapped with him
[25:11] briefly, but he was a very important
[25:13] mentor and friend to my good friend and
[25:16] mentor Jennifer Urick. So, I have a lot
[25:18] of Vanderhiden DNA through her. He was
[25:21] the same person who said being early is
[25:23] the same thing as being wrong. George
[25:24] retires because he can't take the
[25:27] underperformance and he can't take
[25:29] clients saying what's wrong with you?
[25:31] You don't get it. and he has like 40% of
[25:34] his funded tobacco, 40% did homebuilders
[25:38] and literally he underperfor he probably
[25:40] outperformed the NASDAQ
[25:43] by like 20 or 30x over the next three
[25:47] years. Okay. And I have been optimistic
[25:50] that this fundamental shortage of wafers
[25:53] which really today is controlled by
[25:56] Taiwan Semi will prevent one. If Taiwan
[25:58] Semi did what Jensen wanted, I think
[26:00] Nvidia could sell two trillion dollars
[26:02] of GPUs in 26 in 26 or 27, maybe two.5
[26:07] trillion, maybe three trillion, but
[26:09] there is a limit where consumers would
[26:11] consume so much that you probably would
[26:14] be in an overbuild. And so Taiwan Smi,
[26:17] if we don't get a bubble, like we need
[26:18] to throw a party for them because they
[26:20] will have single-handedly prevented a
[26:22] bubble. Okay, you are starting to see
[26:25] companies go to Intel
[26:28] and Samsung.
[26:29] >> Let's just assume TSM stays super supply
[26:31] constraint versus you know the latent
[26:33] demand like what what happens?
[26:35] >> Well, one of you know the history
[26:38] markets is I don't know who but one of
[26:40] Intel and Samsung they're not going to
[26:42] stay disciplined. They will break and
[26:44] then at some level that will force
[26:47] everyone else to break.
[26:50] So like I think a lot of this may come
[26:52] down to the degree to which Taiwan SIM
[26:55] can maintain a lead over Intel and
[26:59] Samsung. You got to remember it's
[27:00] whatever it is it's 9 12 15 months.
[27:02] >> Sort of like the leading node edge. You
[27:04] mean
[27:04] >> exactly you know the pace at which they
[27:07] expand capacity. Like if I were to watch
[27:10] one thing to understand whe there's a
[27:11] bubble it's Taiwan Simmy's capacity
[27:13] decisions. And I think there's a
[27:15] Goldilocks zone where they expand enough
[27:21] they make it hard for Intel or Samsung
[27:24] to really truly emerge as like a um at
[27:29] scale second source with something you
[27:31] know well north of 30% market share. And
[27:35] yet they also keep this fundamental
[27:38] constraint on wafers
[27:41] that you know helps us avoid a bubble.
[27:44] And then obviously I think the terapab
[27:47] um is going to play into this too.
[27:48] >> Say more about that for people that
[27:51] >> the turfab it's a SpaceX I believe
[27:53] Tesla's involved as well um joint
[27:56] venture to build the world's largest fab
[27:59] here in America and I'm I think they're
[28:03] going to be successful. on they have a
[28:05] partnership with Intel which is very
[28:06] important um because they're getting
[28:08] access to 50 years of institutional
[28:11] knowledge that's just you know a nine
[28:13] months a few quarters 12 months 3 to
[28:16] five quarters behind the front that's an
[28:17] advantage it's also an advantage that I
[28:21] believe that terafab is going to get
[28:23] attention from the a teams at all the
[28:25] semicap equipment companies like one big
[28:27] reason Taiwan semicought up is ASML and
[28:31] KA tenor and lamb research and applied
[28:33] material materials. They wanted them to
[28:35] catch up. They didn't they don't like
[28:37] having a monopsiny and so the A teams
[28:40] were in Taiwan working. Intel made some
[28:42] mistakes and presto. And so the A teams
[28:46] will will be here because of Elon's
[28:48] reputation in in hardware engineering.
[28:51] And then just to a degree that I think
[28:54] is u maybe hard for people to imagine in
[28:58] America um where you know politics has
[29:00] replaced religion because Elon had his
[29:02] fora into politics that makes it hard
[29:04] for some people in America to see him
[29:07] clearly which is sad because I do think
[29:10] you know he's probably doing more for
[29:11] America than any other American. You
[29:14] know he's single-handedly bringing
[29:16] manufacturing back to America. He's
[29:18] revived Dince Tech. SpaceX is in some
[29:21] ways the most important defense
[29:22] contractor in America. You know, what
[29:24] he's doing with Starlink is amazing for
[29:26] the world. He's creating all these blue
[29:29] collar manufacturing jobs, which is like
[29:30] a goal, I think, of a lot of liberals
[29:32] and good for America. He's done more
[29:34] than any living human to decarbonize the
[29:37] world. And if you are upset about data
[29:39] sitters on Earth for environmental
[29:41] reasons, well, here you go. You know, so
[29:45] it's it's sad, but he is a living deity.
[29:49] in China, Taiwan, South Korea, and
[29:53] Japan.
[29:55] And having watched him for a long time,
[29:59] what he's going to do is they're going
[30:00] to recruit the best people because the
[30:04] best engineers want to work for Elon,
[30:08] especially in hardware engineering. He's
[30:10] going to recruit incredible engineers.
[30:12] And then they'll be next to next to
[30:14] Turfab, they'll be a Taiwan town. Oh,
[30:17] these are your favorite restaurants. I'm
[30:19] going to move them and their whole staff
[30:21] from Taiwan to Texas and we're going to
[30:24] make everything the way they like it.
[30:26] And then we'll have Japan Town. Same
[30:28] thing. Then we're going to have Korea
[30:29] Town. We're going to have all these
[30:30] things exactly but dialed to recruit the
[30:35] best engineers. And that's just not the
[30:39] way that the people who run Intel at
[30:42] Seung think. So he's going to have the
[30:44] best talent. He's going to have the A
[30:46] teams at the wafer fab equipment
[30:48] companies. He's he has intel which is
[30:51] important. It's so good for all of any
[30:54] administration's political goals. And I
[30:56] think it's different enough that it will
[30:58] not alienate Taiwan SMI.
[31:00] >> And these have long lead times, right?
[31:02] So like Terrafab is going to be pumping
[31:04] out Nvidia G or whatever GPUs, whatever
[31:06] chips like quite quite a long time from
[31:08] now.
[31:09] >> Elon tends to do things differently.
[31:10] Everybody else has taken three years to
[31:12] build a data center. He built one in 122
[31:14] days. You know, Samsung had to give him
[31:17] an office in their fab in Texas because
[31:20] he was so unhappy about like the pace at
[31:22] which they're expanding a building.
[31:24] We'll see. Are you surprised by you
[31:27] mentioned Deep Seek earlier? The simple
[31:29] reaction to that was okay, these models
[31:31] are just going to get 95% as effective
[31:34] for some tiny fraction of the cost to
[31:36] still Chinese open source models. Like
[31:38] we'll be able to use these for most of
[31:39] what we want to do. Fast forwarded a
[31:41] little bit of time, you know, two years
[31:43] from now, there's no reason I have to
[31:45] spend a million dollars a year in my
[31:46] small little firm on on tokens or
[31:48] something. But then the actual reality
[31:50] seems quite different than this. And I'm
[31:52] curious why there's that dissonance in
[31:54] your mind.
[31:54] >> I do think it's the fascinating the
[31:56] returns to the frontier, all the
[31:58] economic returns to AI at the model
[32:01] layer, not all of them, but an
[32:04] overwhelming amount of them have been at
[32:05] the frontier, which is surprising to me.
[32:09] And I think it's been surprising to a
[32:11] lot of people and I think this is one of
[32:15] the most important questions to be
[32:18] answered and you need to have a
[32:19] hypothesis on it as an investor. Are
[32:21] frontier tokens going to continue
[32:24] capturing the overwhelming majority of
[32:27] economic value created at the model
[32:29] layer? And it is surprising like I just
[32:31] I remember when Gemini 3.1 Pro came out
[32:35] and it was it was mind-blowing to me. It
[32:37] was so good. And today it's intolerable.
[32:42] >> Intolerable.
[32:43] And you know there's probably a little
[32:45] bit of a dynamic where companies
[32:46] prototype with Frontiers then when they
[32:48] put something into production you're
[32:50] hearing a lot of people do use Vertex or
[32:52] you know open source. But still it is it
[32:56] is a fact today that the overwhelming
[32:58] majority of these economic turns come
[33:00] from Frontier tokens. And that's
[33:02] surprising and whether or not it
[33:04] continues I think is a very interesting
[33:07] question. And I'm much more open-minded
[33:09] to that having had the experience I've
[33:11] had with Gemini 3.1
[33:14] and then Opus. Um, and then I do use Gro
[33:18] 4.3. It is on the paro frontier. like
[33:21] the companies that are on the paro
[33:22] frontier are and this is by the way a
[33:24] big change in a a consequence of what we
[33:26] talked about last time. Google losing
[33:28] their percost token leadership as a
[33:31] result of making very conservative
[33:32] design decisions with TPU V8 to try and
[33:35] take it away partially from Broadcom and
[33:37] Nvidia um continuing to make aggressive
[33:40] choices. Uh but Google dominated the
[33:43] prao frontier. The prao frontier being
[33:45] intelligence versus cost. And I think
[33:47] this is the most important thing to look
[33:48] at to analyze AI labs. Google dominated
[33:51] that nine months ago. They at every
[33:53] point on the paro frontier. OpenAI, XAI
[33:58] and Anthropic were inside of them. Now
[34:01] the Paro frontier is dominated by
[34:03] Enthropic, OpenAI. And then Grock 4.3 is
[34:06] on the paro frontier. It's clearly like
[34:09] the, you know, the best lowest cost 500
[34:11] billion parameter model. And then Gemini
[34:14] 3.1 is like hanging on to the paro
[34:17] frontier. And if I were to bet or bet
[34:19] that they're subsidizing that out of
[34:20] pride, I would just say one a violation
[34:23] of Richard Sutton's bitter lesson is for
[34:25] sure the biggest risk to this trade
[34:27] >> to all of AI. Now the closer someone is
[34:30] to AI, the more skeptical they are this
[34:32] will occur. One thing I think
[34:34] contributed to weakness in March was,
[34:36] you know, a much more stupid version of
[34:39] DeepSeek, which was this thing called
[34:40] Turboquant. and Turbo Quan is some
[34:42] Google memory optimization that was
[34:44] written up in a paper a year ago. And
[34:46] then during the middle of an agreement
[34:49] while Google was negotiating with
[34:51] Micron, Samsung and Highex to sign, you
[34:53] know, some LTA that would lock in really
[34:55] high prices for a long time. They
[34:57] released this. You know, what people do
[34:59] is always more important than they say.
[35:00] And they just kind of publicize it on X
[35:03] and it goes viral like, "Oh my god, DRM
[35:06] is cooked. Here's this DRAM
[35:08] optimization." I was unable to find a
[35:10] single AI engineer on planet earth who
[35:13] believed that turbo quant would have any
[35:15] impact on DRAM demand but nonetheless a
[35:18] violation of Richard Sutton's bitter
[35:20] lesson you know more compute will always
[35:21] outperform human algorithmic ingenuity
[35:23] more computing data or chin beyond
[35:25] chinchilla optimal I guess what what
[35:27] people increasingly do today that's a
[35:30] real risk man and I think the people who
[35:33] are building these models are skeptical
[35:35] of that risk the reason I am a little
[35:38] less skep skeptical is I think we are
[35:40] very close to ASI and who knows if the
[35:43] bitter lesson holds for 400 IQ models
[35:47] just you know or maybe we get a
[35:49] temporary
[35:51] period where these you know if you get
[35:52] to ASI the first thing it wants is
[35:55] probably to be smarter and have more
[35:57] resources. How does it do that? It makes
[35:58] itself more efficient. I think that is
[36:02] an actual risk that humans the bitter
[36:06] lesson literally I believe includes
[36:08] humans in it. So we're about to find out
[36:11] whether the bitter lesson we'll find out
[36:12] if it applies to 300 IQ ahis then 400
[36:16] then 500 and 600 and at some point we
[36:20] may have like a temporary violation of
[36:22] the bitter lesson based upon AI and ASI.
[36:27] So I'm curious how you think about some
[36:29] other parts of the innovation around the
[36:32] model continual learning and memory
[36:34] being two that see people seem to be
[36:36] most focused on as things that might
[36:37] create yet another you know new paradigm
[36:39] that we would enter. What do you think
[36:40] about the role of those two things?
[36:42] Yeah. Well, I think we've done a lot
[36:43] with memory through these harnesses. And
[36:46] it turns out that harness engineering is
[36:49] not as important as the model, but it
[36:52] really matters. And these harnesses in
[36:54] these models are increasingly being
[36:56] co-developed. One of the big things a
[36:58] harness does, which you just think of as
[36:59] like a a runtime that the model operates
[37:03] in, knows where the pool tools are. It
[37:06] like creates context, memory, state, um,
[37:11] you know, has very specific,
[37:13] you know, prompts or instructions and
[37:16] just
[37:16] >> makes a huge difference. Even simple
[37:18] versions,
[37:19] >> it makes an incredible difference. I
[37:20] think the last time I was on here or one
[37:22] of the other times I just said like,
[37:23] "Hey, as an investor, it's very
[37:26] important that you pay for the $250 a
[37:29] month version to get like your own
[37:31] intuitive sense." that's no longer
[37:33] possible to understand what frontier AI
[37:35] is capable of today even for like a
[37:39] non-coding use case you need to have
[37:40] cloud code or codeex and you need to be
[37:43] on an enterprise plan and the reason for
[37:45] this is and this is another I think this
[37:49] is another dynamic that's enabled by
[37:51] Google losing their cost leadership is
[37:54] these AI models just shifted to
[37:56] usagebased pricing and if you're on that
[37:59] $250 or$300 or $280 month plan or
[38:01] whatever it is you are getting severely
[38:04] rate limited. You are getting a
[38:06] labbotomized version of the AI because
[38:09] like we talked about Claude now produces
[38:11] 70% less tokens. You want the tokens
[38:13] that Claude and its harness really think
[38:16] it needs to produce to get you a good
[38:17] answer, you need to be on a usage based
[38:19] plan. And by the way, this is so bullish
[38:23] for AI. I was a telecom analyst in ' 05
[38:25] to07 and cellular had been a great
[38:28] growth industry really for the last 10
[38:30] years and the reason was you had a
[38:32] combination of fixed pricing you had 900
[38:35] minutes for whatever it was and then
[38:37] usage based pricing over that and when
[38:40] did cellular stop being a great growth
[38:42] industry when everybody just went to all
[38:44] you can eat. And and by the way long
[38:47] distance was the same thing. AI is just
[38:48] shifting from all you can eat to pay by
[38:51] the drink. And it turns out people
[38:52] really like to talk to their friends
[38:54] long distance. They really like to talk
[38:56] to their friends on the phone. And
[38:57] people really like to use AI and
[39:00] particularly now that one person can
[39:02] have a 100 agents working. So I think
[39:03] the shift to usage based pricing is
[39:07] probably why you will see OpenAI and
[39:10] Anthropic exceed well over $200 billion
[39:13] in ARR this year. because not only is
[39:15] more compute going to become online, but
[39:17] they're going to be able to push
[39:19] frontier token pricing with these usage
[39:21] enterprise models, but it's it's sad.
[39:24] It's sad for the world and because it
[39:26] just means if you can't afford that,
[39:27] you're not at the frontier. But yeah,
[39:30] continual learning, man, I mean, if we
[39:32] solve that,
[39:32] >> how do you conceptualize that? Like
[39:34] >> there's so many mysteries about the
[39:35] human mind, like we're such sample
[39:38] efficient learners relative to AI. Like
[39:42] I forget what it is, but like an AI
[39:44] needs orders of magnitude.
[39:45] >> Yeah. Many orders of magnitude. Now we
[39:47] have a crude variant of continual
[39:49] learning today when something is
[39:51] verifiable and that's just, you know,
[39:53] reinforcement learning during
[39:54] mid-training. But yeah, continual
[39:56] learning is a model that dynamically
[39:58] adjusts its weights or adjusts in some
[40:02] way in real time. Like as a human,
[40:05] >> that's what you do.
[40:06] >> Yeah. Like if I the first time I touch
[40:08] or you know put my hand in a fire, I've
[40:11] learned I never put it in there before.
[40:13] That model today needs to put its hand
[40:15] in the fire a million times and then
[40:18] have, you know, the designers
[40:20] effectively put a fire in the next
[40:23] training run or an RL gym for it to
[40:26] learn. I think it has to be dynamically
[40:29] updating the weights, but I think people
[40:31] are working on really smart techniques
[40:33] beyond this. But if we get that then we
[40:37] have a really fast takeoff and people
[40:40] seem
[40:42] confident that continual learning is
[40:45] kind of just around the corner. And I do
[40:47] think this is like the third big
[40:50] question. Bitter le violation as a
[40:52] result of ASI or less likely human
[40:55] ingenuity. Will Frontier tokens still
[40:57] command the premium they do? And will we
[41:00] get continual learning? And if so, when?
[41:02] As your business scales up, everything
[41:04] gets more complex, especially your
[41:05] compliance and security needs. With so
[41:07] many tools offering band-aids and
[41:09] patches, it's unfortunately far too easy
[41:11] for something to slip through the
[41:12] cracks. Fortunately, Vanta is a powerful
[41:14] tool designed to simplify and automate
[41:16] your security work and deliver a single
[41:18] source of truth for compliance and risk.
[41:20] There's a reason that Ramp, Cursor, and
[41:22] Snowflake all use Vanta. It frees them
[41:24] to focus on building amazing
[41:26] differentiated products knowing that
[41:27] compliance and security are under
[41:29] control. Invest like the best listeners.
[41:31] Get a special offer of $1,000 off Vanta
[41:33] when you go to vanta.com/invest.
[41:36] I know firsthand how complex the tech
[41:38] stack is for asset management firms. And
[41:40] seemingly every new tool and data source
[41:42] makes the problem even worse, adding
[41:44] more complexity, more headcount, and
[41:46] more risk. Ridgeline offers a better way
[41:48] forward. One unified platform that
[41:49] automates away that complexity across
[41:51] portfolio accounting, reconciliation,
[41:53] reporting, trading, compliance, and
[41:55] more. All at scale. Ridgeline is
[41:58] revolutionizing investment management.
[41:59] helping ambitious firms scale faster,
[42:01] operate smarter, and stay ahead of the
[42:02] curve. See what Ridgeline can unlock for
[42:04] your firm. Schedule a demo at
[42:06] ridgeline.ai.
[42:07] What is the role of new chip companies
[42:10] in all of this? Like we talked a lot
[42:11] about Nvidia and you know their their
[42:13] sort of relationship with TSMC and Intel
[42:15] and all these sorts of things. There's a
[42:17] thousand flowers blooming. I think
[42:19] literally probably a thousand flowers
[42:20] blooming trying to create a new chip to
[42:23] address some part of this bottleneck.
[42:26] I'm curious how you process this space,
[42:28] this opportunity, what role it will
[42:29] play, what role they'll play.
[42:31] >> So, I think this is good and healthy for
[42:32] the world. It's good for Jensen too. Um,
[42:35] you know, because a different
[42:36] administration might take a different
[42:39] view. Competition, I think, is good for
[42:41] everyone. In in tank design, they talk
[42:42] about the iron triangle. The iron
[42:44] triangle of tank design is that all
[42:46] designers of a tank, they have to make
[42:48] trade-offs between attack, defense, and
[42:50] mobility. And you know, for obvious
[42:51] reasons, the more defense you have,
[42:53] which is just armor, the heavier the
[42:55] tank is, the less mobile it is. So you
[42:58] have to live in this triangle and make
[43:00] tradeoffs. Okay? Like the marava in
[43:03] Israel, it's optimized for defense.
[43:05] Russian tanks and like the Leopard are
[43:08] generally more optimized for mobility.
[43:10] Chip design is the same. And you there
[43:13] there are these fundamental constraints
[43:16] imposed by the laws of physics has
[43:19] embedded in the Taiwan semi design rules
[43:22] that you need to live within and you
[43:26] have TPU tranium and AMD which are all
[43:30] um you know essentially trying to be a
[43:34] better GPU and today I think probably
[43:37] Tranium is doing the best. Nobody's a
[43:39] better GPU, but Trrenium is is I think
[43:42] their, you know, they're they're tugging
[43:44] on Superman's cape
[43:46] >> and and this hadn't started yet. The
[43:48] Tranium 3 needs to ramp into production
[43:50] because it has a switch scale up
[43:52] network, which you really need to
[43:53] economically inference models. You know,
[43:56] a lot of companies have a Taurus
[43:58] architecture. Um that that's where
[44:00] Google was and AMD. We'll see. The
[44:03] MI450, we don't know yet. We'll see. We
[44:06] probably know more about Trinidium 3
[44:07] than the MI450, but that's a hard game
[44:10] to play. So you have to do something
[44:14] different and you have to do something
[44:16] different that is also hard to do. So I
[44:20] think the best path for these startups
[44:22] like my rule of thumb is 1% market share
[44:24] is going to be worth 100 billion. 100
[44:26] billion is a pretty good venture
[44:27] outcome. I think what Jensen would say
[44:29] is like, okay, if something somebody
[44:30] does something different and it gets to
[44:33] one or two or 3% share, we'll make that
[44:36] chip and that's that's coming for
[44:39] everyone. But if you're trying to make a
[44:41] better GPU, good luck. If you were doing
[44:43] something different, it also needs to be
[44:47] hard to do. And you can make different
[44:49] trade-offs. you know, the disagregation
[44:51] of prefill and inference really have
[44:53] opened the aperture um for making these
[44:56] different trade-offs because you can
[44:58] make very aggressive trade-offs for
[44:59] decode, aggressive trade-offs for
[45:01] prefill.
[45:02] >> Prefill being taking in the context,
[45:03] decode being, you know, write the
[45:05] output.
[45:06] >> Yeah. I have a great colleague named
[45:07] Andrew Fox who said, "Picture, you know,
[45:09] a British naval ship from the 18th
[45:11] century. Prefill is loading the cannon,
[45:13] decode is firing it." And what prefill
[45:15] literally is is just the model
[45:17] understanding the question, the prompt
[45:19] and then kind of keeping track of its
[45:20] own dec.
[45:22] And that is fundamentally a memory
[45:24] capacity bound problem. Decode is a
[45:27] process of generating new tokens and
[45:28] that is memory bandwidth constraint. And
[45:31] so if you're a chip designer, this gives
[45:33] you a richer canvas to to paint on. But
[45:36] even so, it needs to be hard because if
[45:39] you make different trade-offs in that
[45:40] iron triangle to optimize for memory
[45:43] capacity, and they're not hard
[45:44] trade-offs to make, well then Nvidia is
[45:47] going to make those same trade-offs.
[45:49] They get better prices from Taiwan Semi
[45:52] than you're ever going to get. Um, and
[45:54] good luck. Good luck. And they have the
[45:56] advantage of working with every model
[45:58] company and optimizing their designs.
[46:00] And by the way, another very funny thing
[46:02] is if you're a VC
[46:04] and you're investing in semiconductor
[46:06] company that is telling you they are
[46:08] going to have an advantage because of a
[46:10] Taiwan semi process that they have
[46:12] special access to. I promise you that
[46:15] Jensen saw that process when it was a
[46:19] twinkle in Taiwan Simmy's eyes and it
[46:22] they know more about it than this little
[46:25] company with 200 people can imagine.
[46:28] Taiwan, CMI, everybody in the supply
[46:30] chain is showing Jensen everything the
[46:32] same way they're showing Amazon
[46:34] everything, AMD everything, TPU
[46:37] everything. And that's another reason
[46:38] don't go try to make a better GPU. So
[46:41] you can do something different. You can
[46:42] paint in the pre-filled canvas. You can
[46:44] paint in the decode canvas, but you also
[46:46] have to do something hard because if it
[46:48] gets to scale, you're going to have
[46:51] those four companies has very fast
[46:53] followers. My firm was a was a um
[46:56] venture investor in Cerebras. What
[46:59] Cerebrris has done is something hard and
[47:00] fundamentally different way for scale
[47:02] computing. And it it comes with a set of
[47:05] trade-offs, but that architectural
[47:08] decision they made was hard and lets
[47:11] them do something that no one else can
[47:13] do. And we'll find out how big that is.
[47:17] And you know, they're working on really
[47:18] cool things like um one of the problems
[47:21] Cerebras has. Once you start needing to
[47:22] glue a lot of chips together and scale
[47:24] up networks or scale out networks, you
[47:27] need a lot of IO and IO is bound by
[47:31] what's called the shoreline, the sides
[47:32] of the chip. And so Cerebris has an
[47:35] overwhelming ratio of onchip computed
[47:37] memory relative to shoreline IO. Well,
[47:40] they're really smart people. They did
[47:41] something really hard. They're trying to
[47:43] see if they can put an optical wafer
[47:45] right on top of that and then that
[47:46] solves that problem. Um, I'm sure
[47:48] they're looking at hybrid bonding of
[47:50] DRAM, you know, to get around these
[47:52] alleged limitations that are not true. A
[47:54] cerebrus machine can theoretically run
[47:56] any size model. There are sizes of
[47:58] models where they're much better than
[47:59] other sizes. So, Cerebras, what I think
[48:02] is interesting is they did something
[48:03] different that's hard to do, really hard
[48:05] to do wafer scale computing. So, I do
[48:08] think there's a role for these and you
[48:10] know, I would just encourage them all
[48:12] make a different trade-off
[48:14] and try and do something hard. because
[48:18] everybody's going to get funded after
[48:20] the Cerebrus IPO. It's not going to be a
[48:22] problem. But it took it took Cerebras
[48:25] three generations of chips to get it
[48:28] right. And it's really hard. Like Andrew
[48:31] Feldman, the CEO, you can just see
[48:36] >> how hard it was
[48:39] and that whole team did
[48:42] >> to get where they are today. And they
[48:44] need to have the grit to do that, the
[48:45] resilience. This first chip is a
[48:47] failure. It happens. Can you come back
[48:48] and make a second chip? But the one last
[48:50] thing on this topic, this is going to be
[48:52] amazing for the useful lives of GPUs and
[48:55] may single-handedly save private credit.
[48:57] >> Say more about that. What what do you
[48:58] mean by the private credit? Well, just,
[49:00] you know, private credit, they're in
[49:01] pain from these SAS loans and however
[49:03] much they're marked down, they probably
[49:04] need to be marked down more because if
[49:06] the public companies are struggling to
[49:07] adapt, how's like a debtleaden company
[49:10] going going to adapt um and invest in
[49:13] what is a very different margin
[49:15] structure business? But there's a lot of
[49:17] private credit and GPUs too. They were
[49:19] underwriting that to I think three or
[49:21] four years. And the disagregation of
[49:24] inference means that I think these GPUs
[49:28] are going to have 10 or 15 year lives.
[49:30] The AI skeptics are like, oh, these
[49:32] companies are all cooking their books.
[49:33] You know, the useful life of GU GPU is
[49:35] only a year or two. The useful life of a
[49:37] CPU is only four years because the rapid
[49:39] technological change. No. What rapid
[49:42] technological change has done with the
[49:44] disagregation of pre-fill and inference
[49:46] is mean that you you know you can put a
[49:48] cerebra system or grock LPUs that Nvidia
[49:51] acquired effectively in front of a
[49:53] hopper or even an ampier use that hopper
[49:56] and ampear for prefill and extend the
[49:58] useful life of that GPU
[50:00] until it melts. Now they do melts they
[50:02] do melt so they have a time but you know
[50:04] maybe you you don't have to run them as
[50:06] fast. This is going to be really good
[50:08] for the whole private credit industry.
[50:10] It's going to help finance the AI
[50:11] buildout because if you can start to
[50:13] finance GPUs at more like you know 5% or
[50:17] 6% instead of I think Corw's lowest
[50:19] financing was like low sevens that
[50:21] actually mathematically changes the cost
[50:23] to finance this buildout. We had this
[50:25] technological innovation that it's going
[50:27] to lower the cost of financing extend
[50:29] the useful life of compute on Earth. And
[50:31] then I do think the one last thing
[50:32] that's interesting about that is um my
[50:35] friend Jamon from Kotu just did a
[50:37] podcast and Cotu had a deck and they
[50:39] talked about hey you know the sellers of
[50:42] shortage are doing so much better than
[50:43] the buyers of shortage. Buyers of
[50:44] shortage being you know the the
[50:46] hyperscalers
[50:48] but if you own a giant installed base of
[50:52] what is currently in shortage that's
[50:55] also a very very good place to be. And
[50:57] we're hearing, you know, CPUs are way
[50:59] more important than they were in an
[51:00] agentic world. They do all these things
[51:02] around orchestration, tool calls, etc.,
[51:03] etc., etc. The biggest CPU fleets in the
[51:06] world sit at the hyperscalers. So, I
[51:08] think some of these hyperscalers may
[51:09] have,
[51:10] >> you know, may may catch up a little bit
[51:12] to the sellers of shortage.
[51:13] >> I want to talk about this idea of
[51:15] different and hard applied outside of
[51:17] the infrastructure piece of this. So,
[51:20] now you're starting to interact with new
[51:21] founders, um, existing CEOs and founders
[51:24] that have to adjust to this new world.
[51:26] What are you seeing like the most AI
[51:28] native founders that aren't building
[51:30] chips or infrastructure or models, but
[51:32] just people using this technology to
[51:34] build other stuff? How do they feel the
[51:36] most different to you if if you've
[51:38] observed differences?
[51:39] >> Well, one, I do think this is just for
[51:40] chip design. To me, it's always been a
[51:42] fundamental question for venture. So,
[51:44] there are different ideas that are
[51:47] obvious to everyone on planet Earth as
[51:48] soon as they hear it. And if that's
[51:50] where you are in venture, if it's not
[51:52] hard to do, if it becomes obvious to the
[51:55] world before you have built um scale,
[51:59] scale is the ultimate advantage, you're
[52:00] in trouble. And the great thing Amazon
[52:03] had was um you know, I think it was
[52:06] obvious to a lot of people, but it
[52:08] wasn't obvious to the retail CEOs. And
[52:10] Amazon, they were very smart.
[52:13] Any e-commerce company that VCs invested
[52:16] in, they would destroy. They'd be like,
[52:18] "Oh, that's so cute. We're gonna we're
[52:20] gonna take our margins of that to
[52:22] negative 10,000%."
[52:24] And that's what like like the guys at
[52:26] Wayfair, they did something hard and
[52:27] Amazon tried to kill them and they
[52:28] failed. Those are like tough
[52:30] operationally
[52:31] like really competent CEOs. For me in
[52:34] venture, I always look, is this going to
[52:36] be obvious to the world before this
[52:39] company could build scale
[52:42] or is this both not obvious, different,
[52:45] and really hard to do? I think a lot of
[52:48] founders are really struggling with this
[52:51] >> in AI like I think people are
[52:56] becoming worried you know today in that
[52:59] in Jensen's five layer cake of AI
[53:02] and the profits they're acrewing to
[53:04] energy they're crewing to data centers
[53:06] they're crewing to chips they're
[53:07] acrewing to models they're not really
[53:09] acrewing to the applications cursor and
[53:12] cognition you know got to a scale you
[53:15] know they focused on coding
[53:17] you know 18 months ago the people were
[53:19] focusing on coding. OpenAI was doing
[53:20] everything under the sun. The people
[53:22] focused on coding were cursor cognition
[53:24] and um anthropic and it was really right
[53:27] to focus on code. Um I'm MSAD the
[53:30] founder of Replet tweeted something that
[53:32] I thought was so smart just it was
[53:34] something like you know bitter lesson
[53:36] adjacent is the fact that coding might
[53:39] be the shortest path to ASI and useful
[53:41] AI because if you're really good at
[53:43] coding you can write yourself code to do
[53:45] anything. So I think it was really smart
[53:46] of those companies to focus intensely on
[53:48] coding and I think they all probably got
[53:51] to a scale where they they have a place.
[53:54] I think cognition is doing something
[53:55] really really different but I think a
[53:57] lot of founders are really struggling
[53:59] man they're really struggling
[54:03] and you know I think they're trying to
[54:04] get confidence that in nichier areas
[54:08] >> that they can get to them and get like a
[54:12] you know a data moat
[54:14] >> before the model companies get to that
[54:16] niche or that it's a small enough niche
[54:18] that the model companies won't do it
[54:20] themselves but it can still produce a
[54:22] venture outcome. Is this related to what
[54:23] you would call like the token path? I
[54:24] know you've used that phrase with me
[54:26] before.
[54:26] >> Yeah, I he comes from a guy um at
[54:28] alttimeter, Jamon Ball. He just said if
[54:30] you're a software company or an AI
[54:32] company of any kind, you have to be in
[54:33] the token path. So, data bricks that's
[54:35] in the token path. Comparable companies
[54:37] are in the token path. If you're not in
[54:39] the token path
[54:41] and you're not in some really niche
[54:45] thing, life may be hard. And even for
[54:49] these vertical niches, I think if you
[54:51] talk to the people at the model
[54:54] companies, they're even skeptical of
[54:56] some of these because all of the data
[54:59] that's, you know, being generated in
[55:01] these niches come from humans. But then
[55:03] you're betting that you're able to use
[55:05] that proprietary data in this narrow
[55:07] vertical to train a model that's lower
[55:10] cost than the frontier labs can ever get
[55:12] to. And maybe that's a good bet, but I
[55:14] just think you have to be very very
[55:16] careful. Now on the other hand, if the
[55:19] returns to these frontier tokens
[55:21] relative to other tokens come down,
[55:24] there's going to be an explosion in
[55:26] value creation at the application layer.
[55:29] And I think another really important
[55:31] point is
[55:34] I have a belief that whenever he wants,
[55:39] Jensen can probably get pretty close to
[55:41] the frontier
[55:43] >> with his own model.
[55:44] >> With his own model, they're doing some
[55:45] really cool things. Neimatronics
[55:47] >> commoditize your compliment as
[55:49] >> say I don't think he wants to do that
[55:52] that is what open AI and you know
[55:56] anthropic are kind of trying to do to
[55:58] him unsuccessfully
[56:01] but so it's just like he's a very
[56:02] logical thinker this is the logical
[56:04] counter move
[56:06] >> and I think you will see that like
[56:08] opensource frontier which today consists
[56:12] of you know Chinese models with stolen
[56:15] American tokens you know Somebody told
[56:16] me that like Deep Seek
[56:19] uh the latest one or maybe the original
[56:21] one was only 150,000 reasoning traces.
[56:23] There's many ways to launder this if
[56:25] you're a Chinese company. You know, you
[56:28] can hit all these different APIs. You
[56:31] can make it hard. Now, the American labs
[56:32] are working really hard on
[56:34] anti-distillation technology. But I I
[56:36] just think Chinese open source, they're
[56:38] doing really impressive things in a very
[56:40] resource constrained way. But there's a
[56:42] lot of distillation. And this is why I
[56:45] think in addition to there not being
[56:46] enough compute to serve Mythos
[56:49] just they did not want it to be
[56:52] distilled. they wanted to use Mythos,
[56:55] you know, distill it themselves, use it
[56:57] to RL their next model, whatever it is.
[57:00] And then I think what they and
[57:02] eventually I think if OpenAI gets to,
[57:04] you know, economics feel good about
[57:06] anyone on the frontier will do is just
[57:07] say, you know, there's going to be some
[57:09] very interesting game theory because
[57:11] it's it is it's a new kind of prisoner's
[57:13] dilemma. You know, we talked about the
[57:14] old prisoners dilemma being just around
[57:17] like, hey, you you're in a prisoner's
[57:18] dilemma where you have to spend. The new
[57:20] prisoners dilemma is going to be if you
[57:22] were at the frontier, do you release
[57:24] that model via API or not?
[57:26] >> And if everyone at the frontier agrees
[57:30] not to do that, then Chinese open source
[57:33] is quickly
[57:34] >> if one person defects, they're going to
[57:37] have the best model. They're going to
[57:39] have a lot of revenue and cash flow and
[57:41] then of course resources equal
[57:42] intelligence. So they'll start to pull
[57:44] ahead and then that will lead to, you
[57:47] know, everybody else releasing it. So
[57:48] it's a new game theory. It's kind of the
[57:50] same game theory that you have with
[57:51] Taiwan semi Samsung and Intel. The
[57:53] reality is like if if a company like
[57:55] Nvidia were or AMD were to ever really
[57:58] really use one of these other
[58:01] foundaries, that foundry would get
[58:02] better really quickly. So I do think
[58:06] Jensen is going to keep open source a
[58:09] certain time frame behind the frontier.
[58:12] I think that's going to be a very
[58:14] interesting thing to watch. And then by
[58:16] the way, open source gets monetized.
[58:17] There's this misnomer that open source
[58:19] is free. Open source tokens, they cost
[58:21] energy. They, you know, they cost energy
[58:23] to produce. You need to make up on GPUs
[58:25] and the open source model companies
[58:26] almost always get a revenue share.
[58:28] >> How are you preparing a trades for the
[58:31] world of Mythos 3, Mythos 4.
[58:34] >> We're just trying to overinvest in cyber
[58:36] security. You know, something I've like,
[58:37] you know, said in multiple forums and I
[58:39] really believe is you everybody needs to
[58:42] have a safe word. Everybody needs to go
[58:46] leave your digital devices behind.
[58:48] Literally go to the ocean and have a
[58:50] family safe word or a company safe word.
[58:52] And it can't be one that can be like
[58:54] socially engineered. And this is just to
[58:56] avoid like cyber crime where like what
[58:58] looks like your son or your daughter or
[59:00] your your grandparents or your parents
[59:02] or whatever facetimes you. It's an
[59:05] utterly accurate
[59:08] simulation of them. they know everything
[59:10] and can extrapolate based on what
[59:12] they've said, what they're likely to
[59:13] say, and says, you know, wire me a
[59:16] million bucks.
[59:17] >> That's defensive. What about what will
[59:18] you still be able to do that it won't be
[59:19] able to do, I guess,
[59:21] >> on the analytical side.
[59:22] >> So, it's a good question. I did just
[59:23] have I I just watched The Last Samurai
[59:25] and I asked um people at my firm to
[59:27] watch it. And The Last Samurai, if you
[59:29] haven't seen it, I highly recommend
[59:31] watching it. It's actually a movie
[59:32] that's aged really well. Tom Cruz movie
[59:34] from 20 years ago. You know, the conceit
[59:36] is Tom Cruz is this like bitter, washed
[59:38] up Civil War veteran who's actually a
[59:40] very good soldier. He's bitter and
[59:42] washed up because he feels like he
[59:44] participated in negative actions against
[59:46] the Native Americans. He's hired by
[59:48] Japan to train just during the Miji
[59:50] restoration. And he's hired by the
[59:52] modern elements of the Japanese
[59:54] government to train like an army of
[59:56] peasants
[59:57] >> how to fight the samurai. There's a
[59:59] first battle. Of course, the samurai win
[01:00:01] even though they don't have guns. He
[01:00:03] fights valiantly. So the samurai decide
[01:00:05] not to kill him, take him to their
[01:00:06] village. He becomes a samurai. It feels
[01:00:08] like the civil war to him. So he fights
[01:00:10] on the side of the samurai.
[01:00:12] And at the end, he's massacred by a
[01:00:14] peasant with a machine gun. And like the
[01:00:16] machine gun is here and if we do not all
[01:00:21] become masters of the machine gun, we're
[01:00:23] going to get mastered. So I am trying to
[01:00:25] become a master of the machine gun. And
[01:00:27] then, you know, I'm optimistic. There's
[01:00:30] a long period of time where just like if
[01:00:33] you were a 50year-old samurai veteran of
[01:00:37] many wars, I fought many wars, master
[01:00:39] dwarf. Um, you will have advantages
[01:00:42] using the machine gun. And I'm
[01:00:43] optimistic as a lifelong student of
[01:00:46] investing. I'm going to be able to
[01:00:47] master the machine gun, this new
[01:00:49] technology, um, integrate it into my own
[01:00:52] process, integrate it into our firm's
[01:00:54] process in ways that, you know, let me
[01:00:57] contribute value as a human being for a
[01:01:00] long time. But, you know, like everyone,
[01:01:01] like, you know, I have agents running
[01:01:03] all the time now.
[01:01:04] >> What's your most useful agent? The most
[01:01:05] useful agent honestly is as and I think
[01:01:08] I told you this and I don't want to hurt
[01:01:10] your business, but my single most useful
[01:01:12] agent is a really good summary of the
[01:01:16] points that would be interesting to me
[01:01:18] from podcasts. There's like six hours a
[01:01:21] day of stuff that I feel like it's in my
[01:01:23] job description to watch, you know,
[01:01:25] every time every time somebody from
[01:01:27] OpenAI, XAI,
[01:01:31] Google,
[01:01:32] you know, Cursor,
[01:01:34] Fireworks, Bin, I say nothing of like
[01:01:37] Jensen, Elon, Daario. Um, I feel
[01:01:42] compelled to watch and I just don't have
[01:01:44] that much time. And there's some real
[01:01:47] needles and hay stacks. There's a set of
[01:01:48] things I always like to see like I'm
[01:01:50] very sensitive to management
[01:01:51] compensation. What are they incented to
[01:01:53] do? They do they just have stupid RSUs
[01:01:56] or do they have PSUs? And if they have
[01:01:58] PSUs, what are those PSUs incent them to
[01:02:00] do? I think systems that do a very good
[01:02:02] first pass at that and you know that
[01:02:05] saves people a lot of time. It frees
[01:02:07] them up for more creative work than like
[01:02:10] you know going through the proxy pulling
[01:02:12] the PSU thing looking at how it's
[01:02:15] changed versus all the proxies because
[01:02:17] there's signal in that and that's very
[01:02:19] labor intensive and that's so good for
[01:02:20] an AI and there's obviously all sorts of
[01:02:22] same things within investing. This is
[01:02:24] the most exciting thrilling time to be
[01:02:26] an investor
[01:02:28] >> and there is and it is I am a little I'm
[01:02:30] getting a little bit worried
[01:02:32] >> the diversity breakdown thing. Yeah, I'm
[01:02:34] getting
[01:02:35] >> Say just like a little bit more about
[01:02:36] like the kinds of people that are
[01:02:37] >> I don't know anyone like me who's not
[01:02:39] really bullish on DRM.
[01:02:42] >> No one.
[01:02:42] >> No one. There's all these interesting
[01:02:44] things happening with AI right now. So,
[01:02:46] one is cross-sectionally the valuations
[01:02:48] do not make sense. They just flat out do
[01:02:51] not make sense. They cannot all be true.
[01:02:54] You have semicap equipment companies
[01:02:56] trading at 40 times next quarter's
[01:02:58] annualized earnings and DRAM companies
[01:03:00] trading at mid-s single digit. at the
[01:03:02] peak of the last cycle that was like
[01:03:04] five verse 12. At one point it was like
[01:03:06] three verse 45. Those can't both be
[01:03:09] true. And yes, semiconductor capex
[01:03:12] business models have improved more than
[01:03:14] the memory business models. We don't
[01:03:16] know how much HBM is going to improve
[01:03:18] memory business models yet. Yes, they
[01:03:20] have some element of recurring revenue
[01:03:22] with parts and maintenance, but it's not
[01:03:25] worth a,000% multiple gap. I think it's
[01:03:27] hard to square like the valuation of
[01:03:29] something like Nvidia which is still you
[01:03:31] know in in in early April was
[01:03:33] essentially as cheap as it gets relative
[01:03:35] to the market like in the last 10 or 12
[01:03:37] years or whatever it is and very cheap
[01:03:39] absolute it's very hard to square that
[01:03:41] valuation with something like GE
[01:03:44] Vernova's valuation
[01:03:46] >> because it builds in like an
[01:03:48] unfathomable amount of share loss for
[01:03:51] Nvidia. So valuations cross-sectionally
[01:03:53] are really different because we are in
[01:03:56] shortages.
[01:03:58] The lowest quality companies are doing
[01:04:00] the best. So if you're an oil and gas
[01:04:03] investor or you know a mighty investor,
[01:04:05] natural resources investor and you're,
[01:04:08] you know, you're well versed in thinking
[01:04:09] of costs, this is very intuitive to you.
[01:04:11] In a real bull market for a commodity,
[01:04:14] the commodity suppliers with the highest
[01:04:16] costs go up the most because it's the
[01:04:19] most beneficial to them. They go from on
[01:04:21] the verge of bankruptcy to gushing cash.
[01:04:23] And this is, I think, one reason
[01:04:25] commodity investing is really, really
[01:04:26] hard because quality outperforms during
[01:04:29] the cycles, but you get all of the
[01:04:31] outperformance during the downturns when
[01:04:33] the high-cost guys that moon during the
[01:04:36] shortages and the commodity bull
[01:04:37] markets, you know, go bankrupt or
[01:04:39] whatever. You're seeing that happen in
[01:04:40] every industry. the lowest quality
[01:04:43] players in, you know, these different
[01:04:45] industries that are hated and detested
[01:04:49] by the hyperscalers and the buyers
[01:04:51] because they have high costs, they're
[01:04:53] unreliable, the parts fail at a high
[01:04:55] rate, etc., etc. They're sold out and
[01:04:57] raising prices. Um, and then that
[01:04:59] activity gets the interest of like these
[01:05:02] retail accounts on X and these stocks
[01:05:05] get bid to the moon. whereas some of the
[01:05:08] higher quality expressions
[01:05:10] have like actually really underperformed
[01:05:13] and you know as an investor it's it's
[01:05:15] hard because you know within a like
[01:05:20] shadow of a doubt that that thing that's
[01:05:23] moved you know 10x in 3 months or 6
[01:05:26] months is going to go right back down
[01:05:30] subject to what they do with all the
[01:05:31] cash. But like these low quality
[01:05:33] companies really do smart stuff with
[01:05:34] cash. And so it worries me a little bit
[01:05:36] that people who were very skeptical a
[01:05:38] year ago are no longer skeptical. But
[01:05:41] then I just contrast that with like the
[01:05:43] valuations of these like highquality
[01:05:46] companies which are just not extended
[01:05:49] and it makes me feel better. But it does
[01:05:51] kind of feel like, you know, I always
[01:05:52] thought it was funny in 24 and 25 that
[01:05:54] anyone asked about an AI bubble or
[01:05:56] talked about it because it's like you
[01:05:58] have this nuclear bubble and this
[01:05:59] quantum bubble right here, right in
[01:06:01] front of you. What are we talking about?
[01:06:03] This is so real. Some of that nuclear
[01:06:06] quantum silliness is maybe spread into
[01:06:09] more speculative, lower quality, smaller
[01:06:12] cap names where if you have a big
[01:06:15] presence on X or Reddit, it's easy to
[01:06:18] move them. And that frightens me a
[01:06:20] little bit, but I just wish there were
[01:06:22] more AI bears. Like I wish there were
[01:06:24] more memory bears. You know, one reason
[01:06:26] I'm um you know, Astera is a stock I've
[01:06:29] been close to a long time. There's a lot
[01:06:31] of bears on that. I love that. Great.
[01:06:34] You know, I first invested in the series
[01:06:36] C. Good luck thinking you're going to
[01:06:37] price that, you know, differentially for
[01:06:40] me. You know, good luck thinking that's
[01:06:41] a copper loser. And then there's also
[01:06:44] you can feel the baskets in the market
[01:06:46] and the leverage baskets and what
[01:06:48] baskets you're in is really important.
[01:06:50] You know, copper, optical, DRAM, NAND.
[01:06:54] Um, and a very interesting thing that's
[01:06:55] happened this year, um, is in 24 and 25
[01:06:58] the AI trade traded together. So like
[01:07:02] you could be long GPU compute, scale up
[01:07:05] networking, and optical scale across and
[01:07:08] like short power. that trade worked from
[01:07:12] like a riskmanagement sense because you
[01:07:13] know I'm very factor aware that all blew
[01:07:16] out in January of this year it's like
[01:07:19] you know scale up networking would go
[01:07:21] crazy while scale out was going down or
[01:07:23] DRM massively underperforming NAND and
[01:07:27] HDDs which had not h happened so these
[01:07:30] cross-sectional correlations within AI
[01:07:34] really fell apart and you had to get
[01:07:36] very fine grained you couldn't hedge
[01:07:39] your memory
[01:07:40] anymore with like some semicap equipment
[01:07:43] or nan everything cross-sectionally
[01:07:48] really changed and in a very interesting
[01:07:50] way in January and I think maybe one
[01:07:53] reason for that was you know the AI got
[01:07:55] to a quality where it was all of a
[01:07:58] sudden really easy for a bunch of people
[01:08:00] to get really smart on these different
[01:08:02] subsectors start trading them and then
[01:08:04] they get put into baskets and those
[01:08:07] baskets
[01:08:07] >> yeah creating price efficiency Yeah,
[01:08:09] exactly. And then it's like if you like
[01:08:12] I think some of the biggest
[01:08:13] opportunities outside of these higher
[01:08:15] quality names that I think can compound
[01:08:16] for a long time and they're safe unlike
[01:08:19] these lowquality names which are
[01:08:20] terrifying is in names that are
[01:08:22] miscatategorized
[01:08:24] like Astera was in a lot of copper loser
[01:08:26] baskets. Astera their biggest product is
[01:08:29] going to be a switch. You use both
[01:08:32] copper and optics to connect switches to
[01:08:35] accelerators.
[01:08:37] And so definitionally, if you're a
[01:08:40] switch company or an accelerator
[01:08:42] company, you cannot be a copper loser
[01:08:45] because you're going to be on the other
[01:08:46] side of that connection. I
[01:08:47] >> I wonder if you could riff just for like
[01:08:49] a sentence or two on each of the major
[01:08:51] companies. I feel like I always forget
[01:08:52] to ask you like Google, Microsoft,
[01:08:54] Amazon, you know, the the major players
[01:08:55] that are public that all the
[01:08:57] conversation is centered around these
[01:08:59] exciting new companies.
[01:09:00] >> Yeah. So Google um it was incredible
[01:09:02] last year because they had that TPU
[01:09:04] advantage which is now gone. The reason
[01:09:05] I think they're still in a great
[01:09:07] position is just they have the most
[01:09:09] compute of everyone. We talked about the
[01:09:11] value of installed bases being higher as
[01:09:13] a result of shortages.
[01:09:15] >> They have the biggest installed base of
[01:09:16] compute. Yeah,
[01:09:18] >> I am a little surprised
[01:09:22] by
[01:09:24] their inability and Google IO is this um
[01:09:28] is this week
[01:09:30] >> and um like if they don't release
[01:09:34] something that even slightly leapfrogs
[01:09:39] open AI
[01:09:40] andor um clawed
[01:09:43] like that that's interesting and it's
[01:09:46] not a disaster. faster for Google. It's
[01:09:48] just interesting and it just means this
[01:09:50] Nvidia effect we discussed is even more
[01:09:52] powerful than maybe I'd imagined. But
[01:09:53] I'm very curious to see what the paro
[01:09:56] frontier looks like literally in 5 days
[01:09:58] after Google's announced its new stuff.
[01:10:01] This is a big card for them. But Google,
[01:10:03] you know, between um the amount of data
[01:10:05] they have and the YouTube data is
[01:10:07] actually really genuinely valuable. It's
[01:10:09] actually it is valuable in a world of
[01:10:12] robotics. The amount of compute they
[01:10:14] have and you know the search business
[01:10:16] they have. Google's never not going to
[01:10:18] be in a good position. And then you see
[01:10:20] that with GCP going crazy. You got to
[01:10:22] give Zuckerberg immense credit. Um what
[01:10:25] he's done in terms of making Meta an AI
[01:10:27] first company internally and I do think
[01:10:30] he is the only one of those true
[01:10:32] internet giants to have done that. And I
[01:10:36] give him a lot of credit for that. I
[01:10:37] give him a lot of credit for um paying
[01:10:40] up when he did for you know all those
[01:10:43] you know those billion dollar contracts
[01:10:45] that talent
[01:10:46] >> and Muse I think was a really big upside
[01:10:49] surprise um you know was the first model
[01:10:53] from MSL and it's not on the paro
[01:10:56] frontier with you know XAI Google's one
[01:10:59] entrant and then openAI and claude but
[01:11:02] it's pretty close that was very
[01:11:04] impressive to me so I think meta is in a
[01:11:07] better position. Still not as strong of
[01:11:09] an absolute position as Google, but like
[01:11:11] they're better position and rates of
[01:11:13] change matter more than level as you
[01:11:15] know in markets particularly over short
[01:11:17] like three-year time frames over like
[01:11:19] long time frames level of competitive
[01:11:21] advantages tends to dominate but even
[01:11:23] within that you know the changes changes
[01:11:25] are really matter. Amazon I think is in
[01:11:28] a really strong position because of
[01:11:29] Trrenium. you're going to see like real
[01:11:32] P&L efficiencies from robotics over the
[01:11:34] next 18 months in their retail business.
[01:11:36] I actually think Nova their internal
[01:11:38] models are not where Muse is, but
[01:11:41] they're better than they get credit for.
[01:11:43] Microsoft, I think Satya is a really
[01:11:45] brilliant man, but you know, in in
[01:11:48] investor conversations,
[01:11:50] people just don't talk about him the way
[01:11:52] that they did. I I like Satya. I admire
[01:11:55] him. I think he's an exceptional CEO
[01:11:59] and I give him a lot of credit for the
[01:12:01] decisions he's made, but you know, he
[01:12:04] did go from we're going to make Google
[01:12:05] dance to being the product manager of
[01:12:07] Copilot in like three years. I I would
[01:12:10] love to know during the coup attempt
[01:12:12] against OpenAI, does Satcha regret his
[01:12:15] decisions?
[01:12:17] Does Satia wish that he had supported
[01:12:20] Ilia and instead of Sam and that kind of
[01:12:23] Ilia and Meera were really running
[01:12:27] OpenAI today? In his heart of hearts, I
[01:12:30] would love to know because I think the
[01:12:33] Microsoft OpenAI partnership might look
[01:12:35] very different in that world. I think
[01:12:38] that's a very interesting question that
[01:12:40] we'll never know the answer to.
[01:12:43] But I give him a lot of credit like he
[01:12:45] is what he is doing now he's taking risk
[01:12:50] so they could earn you know this goes to
[01:12:52] the decisions you have to make in that
[01:12:54] cone of uncertainty are not only how
[01:12:56] much you spend but what you're going to
[01:12:58] spend it on I think Microsoft flinched
[01:13:03] for like a moment in early 25 you know
[01:13:06] they have this algorithm we spend this
[01:13:08] much capex dollars we get this return
[01:13:10] that algorithm was kind of off and if
[01:13:13] you flinch you lose position
[01:13:15] >> you lose all these allocations and it's
[01:13:17] difficult to get it back. So they
[01:13:19] flinched and now the decision Satya is
[01:13:21] making which the market has punished him
[01:13:23] for but I think is the right decision is
[01:13:26] we're going to use our compute rather
[01:13:28] than making I mean who knows how fast
[01:13:30] Azure could be growing if they're
[01:13:32] willing to just sell GPUs to OpenAI.
[01:13:35] We're going to use our compute
[01:13:37] internally to make our own products
[01:13:39] better. You know, one reason C-pilot is
[01:13:41] so bad or has been so bad is just not
[01:13:43] enough compute available. They're fixing
[01:13:44] that. He's the product manager of
[01:13:47] Copilot. I do think he's a great CEO and
[01:13:51] they're trying to use their compute to
[01:13:52] train their own models. I don't I am a
[01:13:55] little skeptical that they have the
[01:13:56] right team to succeed there but you know
[01:13:59] they can certainly like just like Meta
[01:14:01] they can afford to hire maybe maybe a
[01:14:04] different team but I think he's making
[01:14:07] good decisions that are risky decisions
[01:14:11] to position Microsoft from for this
[01:14:13] world where frontier models are are no
[01:14:17] longer API accessible
[01:14:19] >> and I think it's a really courageous
[01:14:20] decision that I give him a lot of credit
[01:14:22] for and he is foregoing I Microsoft
[01:14:24] probably be an $800 stock today if they
[01:14:27] were using their GPUs to serve OpenAI
[01:14:30] solely OpenAI and anthropics capacity
[01:14:32] instead of using them for their own
[01:14:34] products. So I give him a lot of credit
[01:14:36] for making a great decision. What's
[01:14:38] really interesting is the degree to
[01:14:41] which these companies are outward facing
[01:14:44] in their decisions. The two companies
[01:14:46] who are the most deeply engaged with
[01:14:48] startups are Amazon and Nvidia by a
[01:14:51] mile. Then there's a really intense
[01:14:55] engagement with Google, their next most
[01:14:57] intense. Broadcom is engaged in a
[01:15:00] different way. They're just, you know,
[01:15:02] everybody's favorite AS6 supplier. Like
[01:15:05] it's, you know, if you're a startup,
[01:15:06] it's considered like a level up if you
[01:15:08] get to work with Broadcom for your
[01:15:09] second gen chip. And it's considered
[01:15:11] mana from heaven if Broadcom works with
[01:15:13] you for their first gen chip. And then
[01:15:15] you see essentially
[01:15:17] zero engagement with startups from AMD,
[01:15:22] Microsoft, and Meta. And I just Yeah, I
[01:15:24] mean when I say zero, it's a little. And
[01:15:27] I just wonder about that decision
[01:15:30] because some of the best teams
[01:15:35] are no longer at big public companies.
[01:15:37] They're at these smaller startups.
[01:15:40] And I think it's going to end up being a
[01:15:42] pretty big advantage for Nvidia, AMD,
[01:15:44] Google right behind them to have this
[01:15:47] engagement
[01:15:49] that you just don't see from these other
[01:15:53] um hyperscalers.
[01:15:54] >> As we wrap up, I'm curious for you to
[01:15:55] riff on any other like out there
[01:15:57] knock-on effects that you've started to
[01:15:59] think about for this giant trend. We've
[01:16:01] talked about the specific companies in a
[01:16:02] lot of detail that this most impacts. We
[01:16:05] talked a little bit about the
[01:16:05] application layer and what would have to
[01:16:07] happen for there to be more value
[01:16:08] occurring to that layer of the stack.
[01:16:10] I'm curious like any other just fun
[01:16:12] knock-on things that you've been
[01:16:13] thinking about as this world changes so
[01:16:15] quickly.
[01:16:16] >> Yeah. And it is wild. I mean at the
[01:16:17] application layer, forget value
[01:16:18] acrewing, just value has been destroyed.
[01:16:20] >> AI has net destroyed. Even if you count
[01:16:22] cursor cognition, the most successful AI
[01:16:25] natives, value has been trillions of
[01:16:28] dollars of value has been destroyed by
[01:16:30] AI at the application layer. And just in
[01:16:32] this context, I do think it's a little
[01:16:34] it's something we need to be aware of.
[01:16:36] The companies that are doing the best
[01:16:38] today that are seeing kind of their
[01:16:41] values increase the most that are
[01:16:43] creating economic value are the
[01:16:45] companies with the highest ratio,
[01:16:48] highest effective ratio of utilized GPUs
[01:16:51] per human.
[01:16:52] >> And you know, maybe this just means that
[01:16:54] every human's going to get a lot of
[01:16:55] GPUs, but I think that's an interesting
[01:16:57] fact that we kind of need to be
[01:16:59] cognizant of. I will just say and maybe
[01:17:01] this is a little dark. I am more more
[01:17:04] and more worried about personal safety
[01:17:06] and I worry about this a lot more for
[01:17:08] people who are you know have a much
[01:17:11] bigger public presence and are much more
[01:17:12] associated with AI but I really worry
[01:17:15] about personal safety. I hope nothing
[01:17:17] tragic happens, but like there is this
[01:17:19] upsurge in political violence here in
[01:17:21] America and as AI increasingly becomes
[01:17:24] political, I worry that's going to get
[01:17:26] directed at more and more AI political
[01:17:28] leaders. You know, just whatever we can
[01:17:30] agree, you know, whatever whatever I may
[01:17:32] think or may not think of open AI like I
[01:17:35] think it is terrible that someone threw
[01:17:36] mal malatto cocktails at Sam Alman's
[01:17:39] house. I am worried that we are headed
[01:17:42] into a higher variance,
[01:17:46] higher beta,
[01:17:48] higher risk world because of AI. And
[01:17:51] that's for me as an individual and then
[01:17:53] you know for people who are big players
[01:17:55] on the chess board. Think about what it
[01:17:57] means geopolitically like we're watching
[01:18:00] the Ukrainians are really starting to
[01:18:01] win. And the reason they're winning I I
[01:18:04] think is not really because they have
[01:18:05] better drones. I think they do have
[01:18:07] better drones. That's part of it. I
[01:18:08] think the reason Ukraine is really
[01:18:09] winning is they have the best
[01:18:11] battlefield AI outside of probably
[01:18:14] America and Israel and has China has our
[01:18:19] adversaries begin to process that
[01:18:22] like how do they respond? Like if the
[01:18:25] United States because of its edge in AI
[01:18:28] um it's great if you're America but it
[01:18:32] is destabilizing for the rest of the
[01:18:34] world. Something I think a lot about is
[01:18:36] creating a charity to just like educate
[01:18:37] the world on how awesome the west has
[01:18:39] been. Slavery was endemic to essentially
[01:18:41] almost every civilization and slavery
[01:18:43] was really ended by the British Empire.
[01:18:45] Tell that story. Um but America after
[01:18:49] 1945
[01:18:51] we had the nuclear bomb. No one else had
[01:18:53] it. We could have controlled the world
[01:18:56] forever. Instead, we rebuilt Germany and
[01:18:59] Japan and now we're America's most
[01:19:03] reliable allies. Israel, South Korea,
[01:19:05] Japan. That's a testament to like the
[01:19:07] American spirit in our country. We
[01:19:08] didn't take over the world. You know,
[01:19:10] there were these fears, you know, that
[01:19:11] were documented at the time that the
[01:19:13] American generals and, you know,
[01:19:15] MacArthur was a little bit of an
[01:19:16] American emperor in Japan,
[01:19:19] but um we're just going to take over the
[01:19:21] world. And they could have and they
[01:19:23] didn't. They came home, we demilitarized
[01:19:26] and then you had this, you know, this
[01:19:28] period of of great global stability
[01:19:30] between, you know, it was scary. They
[01:19:31] were turn America. Yeah. You had the Pax
[01:19:33] Americana.
[01:19:34] >> So maybe it's not destabilizing. Maybe
[01:19:36] it leads to the another Pax Americana
[01:19:40] >> informed by our AI dominance. And I'm so
[01:19:43] optimistic that AI is going to be
[01:19:45] amazing for the world. There's someone
[01:19:47] like me whose daughter was diagnosed
[01:19:49] with a very rare mutation. there's no
[01:19:52] cure. He was able to assemble a lot of
[01:19:54] resources. He was able to get a lot of
[01:19:56] compute from the labs. Um we were made
[01:19:59] aware of what was happening, spun up an
[01:20:01] immense amount of agents, came up using
[01:20:05] AI with a drug on the market that can
[01:20:07] actually impact his daughter's disease
[01:20:09] and then has spun up a company to cure
[01:20:12] it.
[01:20:13] And like her life is already
[01:20:16] immeasurably different because of AI. So
[01:20:18] I'm like an AI I'm like an AI optimist
[01:20:21] maximalist but I also just acknowledge
[01:20:23] it's like an event horizon. It for sure
[01:20:26] I think is going to be a discontinuity.
[01:20:28] We need to navigate has societ as
[01:20:30] society. I think the lites are going to
[01:20:32] be wrong but we need to be like really
[01:20:34] thoughtful in how we address their
[01:20:36] concerns. We need to make sure that it's
[01:20:38] good for everyone. Like it is a little
[01:20:40] dystopian that now the best AI is only
[01:20:42] available to people with a lot of money.
[01:20:44] Like we need to solve that. We need to
[01:20:47] approach this with humility, recognize
[01:20:48] there's a lot of uncertainty, and be
[01:20:50] thoughtful.
[01:20:50] >> When I do this with you, I tell people
[01:20:52] afterwards, I'm like, "May you find
[01:20:53] something that you love as much as Gavin
[01:20:55] loves markets and companies and
[01:20:57] capitalism and history uh on display
[01:21:00] today as always." Gavin, thanks so much
[01:21:01] for your time.
[01:21:02] >> Thank you. Thanks, Patrick.
[01:21:06] >> You know how small advantages compound
[01:21:08] over time? That's true in investing and
[01:21:10] just as true in how you run your
[01:21:11] company. Your spending system is your
[01:21:13] capital allocation strategy. Ramp makes
[01:21:15] it smarter by default. Better data,
[01:21:17] better decisions, better economics over
[01:21:19] time. See how at ramp.com/invest.
[01:21:22] As your business grows, Vant scales with
[01:21:23] you, automating compliance and giving
[01:21:25] you a single source of truth for
[01:21:27] security and risk. Learn more at
[01:21:29] vanta.com/invest.
[01:21:30] Every investment firm is unique and
[01:21:32] generic AI doesn't understand your
[01:21:34] process. Rogo does. It's an AI platform
[01:21:36] built specifically for Wall Street,
[01:21:37] connected to your data, understanding
[01:21:39] your process, and producing real
[01:21:40] outputs. Check them out at
[01:21:42] rogo.ai/invest.
[01:21:44] The best AI and software companies from
[01:21:45] OpenAI to Cursor to Perplexity use work
[01:21:48] OS to become enterprise ready overnight,
[01:21:50] not in months. Visit works.com to skip
[01:21:53] the unglamorous infrastructure work and
[01:21:54] focus on your product. Ridgeline is
[01:21:56] redefining asset management technology
[01:21:58] as a true partner, not just a software
[01:22:00] vendor. They've helped firms 5x and
[01:22:02] scale, enabling faster growth, smarter
[01:22:04] operations, and a competitive edge.
[01:22:06] Visit ridgelineapps.com
[01:22:08] to see what they can unlock for your
[01:22:09] firm.