YT YouTube2Text

YouTube2Text → Transcripts → Watts, Wafers, and the Future of AI Infra | Gavin Baker

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

by Invest Like The Best · 14,375 words · 2,357 segments · language en · Watch on YouTube

Open in transcriber Summarize with AI Download SRT/VTT

Full transcript

0:00What was happening in AI
0:01was I think the most extraordinary
0:03moment in the history of capitalism, the
0:06history of American business. Anthropic
0:08they added $11 of AR. Three highest
0:11profile SaaS companies founded in the
0:14last 10 12 years
0:16are Palantir,
0:18Snowflake, and Databricks.
0:20And these three companies spent 10
0:22[music] years building their businesses.
0:24Anthropic added their combined
0:26businesses in 1 month.
0:30That's just nothing like that has ever
0:32happened in the history of capitalism.
0:34Forget my career.
0:36Just the flat-out history of capitalism.
0:39The history of business.
0:41>> [music]
0:53>> All right, so this is our sixth time
0:55doing this if you can believe it, which
0:57puts you back into first place. At least
0:59tied for first place with Gurley. And I
1:01think even since last time when we did
1:03this, which was so exciting and
1:06spectacular, I think we're in an even
1:08more interesting time now. Maybe just
1:10start by riffing on how it felt for you
1:12living through March and April of this
1:15year, which which felt to me just like a
1:17completely unique economic, technology,
1:20and market environment. And you're the
1:22biggest student of the history and of
1:24these times, so what did it feel like?
1:25Now I'd say broadly speaking, there are
1:27two kinds of drawdowns. There are
1:29drawdowns where you're wrong, a company
1:32misestimates,
1:33your hypothesis was invalidated, and you
1:36have to take your medicine and you
1:38crystallize that loss.
1:40And then there are drawdowns or periods
1:42of underperformance where you're
1:45underperforming because of companies you
1:46know really really well,
1:48and where you profoundly disagree with
1:50the price action, and you can lean in.
1:53And instead of crystallizing
1:55uh negative performance, you can kind of
1:58build pent-up alpha, pent-up future
2:00performance. And for me, that is what
2:02March felt like. It felt like uh
2:05you know, the NASDAQ was selling off,
2:07and at the same time, what was happening
2:09in AI was I think the most extraordinary
2:12moment in the history of capitalism, the
2:15history of American business. And what I
2:17just mean by that is an Anthropic, they
2:19added $11 billion of ARR.
2:21And what is astonishing to me about this
2:25is that
2:27the SaaS and cloud revolution it created
2:29will call it between 5 and 10 trillion
2:31dollars of value. And I would say
2:33arguably the three highest profile SaaS
2:36companies to have kind of
2:38been founded in the last 10, 12 years
2:42are Palantir,
2:43Snowflake, and Databricks. And these
2:46three companies
2:48have spent and employed thousands of
2:50people, tens of thousands collectively.
2:52They've all spent 10 years building
2:54their businesses. And Anthropic added
2:57their combined businesses in 1 month.
3:01>> [laughter]
3:01>> That's just nothing like that has ever
3:04happened in the history of capitalism.
3:06Forget my career.
3:08Just the flat-out history of capitalism.
3:11The history of business. I mean,
3:13it's wild that that Krishna comes out
3:15this show and shares some stats, 500%
3:18in DR.
3:19>> Yeah, you do the math on that for 3
3:20years.
3:21Insanity. We So, there's just no
3:24precedent for this, and we
3:27you know, tech tech investors you you
3:29hear a lot of discussions about S-curves
3:31and investing in exponentials. I've just
3:33never seen an exponential like this. It
3:35felt even more extreme than Deep Seek,
3:38which was a very similar setup. If we go
3:41back to 25,
3:43there was a huge sell-off at Deep Seek,
3:45which was very strange because the paper
3:47gets published
3:497 days
3:51before Deep Seek Monday. It got
3:52published
3:54I believe on a Monday that was a holiday
3:57in America. And I read it, I thought,
3:59"Hmm, you know, this this feels like it
4:02might not read
4:03>> [laughter]
4:04>> that positively for
4:06uh you know, the AI trade. Yeah, I I
4:09took action. We had Deep Seek Monday
4:11where AI really imploded a week later.
4:16And
4:17that was really strange because by Deep
4:19Seek Monday, it was super clear
4:22that this was going to be the most
4:23positive thing that had ever happened to
4:24compute demand.
4:26Prices in the AWS available availability
4:29zones in Asia
4:30had already
4:32like doubled. You were seeing GPU
4:35availability go down.
4:37And this was just the first time we saw
4:40how much more compute-hungry reasoning
4:42models are during inference than
4:45non-reasoning models.
4:46And so that was a similar setup.
4:49But you you had to do some work to see
4:51that. I mean, it's not that hard
4:53to say, "Oh, wow, stocks are selling
4:55off. The price of DRAMs going vertical.
4:56The price of GPUs in Asia are going
4:58vertical.
4:59Um GPU availability's going down. And
5:02then like two or three days later, you
5:04know, GPU prices in in in America
5:06started going up, GPU rental prices."
5:08All you had to do in in March was
5:11just simply observe what was happening
5:13to Anthropic. There's all these people
5:15who seem to regret
5:17you know, not buy during '22, not buy
5:20during COVID, not buy during Deep Seek.
5:23You had the same valuation setup
5:26at the beginning of April.
5:28And and even clearer AI inflection.
5:32And so there've been all these chances
5:35to buy into AI. And then of course, what
5:38complicated it was the straight-up FOMO.
5:40I became a believer, an every believer,
5:43that I think maybe one thing that the
5:45market was mispricing. it. I'm no
5:48background expert. I do do a lot of pro
5:51national security investing. So, I do
5:53have access
5:55to people who are experts that are
5:58excited to share their thoughts and
5:59opinions with me.
6:01And that the Strait of Hormuz being
6:03closed is actually relatively
6:05awesome for America. Why?
6:08Because, particularly for the goals of
6:10the current administration.
6:12So, electricity is a very important
6:14industrial or manufacturing input.
6:17The key
6:18input into American electricity prices,
6:20which feeds into AI,
6:22is in G 1. Natural gas went up in
6:25Bloomberg. That was down 20%.
6:27And natural gas in Asia, Europe,
6:30everywhere else doubled or tripled.
6:34So, our relative manufacturing
6:37competitiveness
6:38improved overnight.
6:40And for better or worse, that is what
6:42the Trump administration seems to care
6:45about. They are very focused on
6:47America's relative position.
6:49And I think a lot of people had memories
6:51of the 1970s.
6:53And what made the '70s so dramatic was
6:56it wasn't just that prices went up.
6:58It's that there were actual gas
7:00shortages. And then you go through,
7:01okay, well, the US economy is
7:04dramatically less energy intensive than
7:05it was. The US economy The United States
7:08is now the world's largest producer of
7:10oil and gas. And we've become now the
7:12world's largest exporter of oil and gas.
7:17And then on top of that, there's this
7:19relative manufacturing advantage. And
7:22so, that made it, I think, easier to
7:26stay
7:28focused on AI fundamentals, stay focused
7:31on what were
7:33historically attractive valuations. I
7:35think on a relative basis, tech
7:37essentially got as cheap as it's been
7:39versus the rest of the market has at any
7:41point over the last 10 years. And just
7:44think about that in the context of
7:45market efficiency. We have the most
7:47extraordinary moment in the history of
7:49capitalism
7:50that's wildly bullish for AI and you get
7:53a chance to buy AI
7:56at really attractive valuation. What do
7:58you make of the multiples that
8:01specifically Anthropic and OpenAI, which
8:03in my mind are like the reference assets
8:05that are the most pure play takes on
8:07this trend?
8:08Really being not that crazy. Like if you
8:10just look at the sales multiple and
8:12compare it to maybe what Databricks and
8:14Snowflake and these companies traded at
8:15at their peak. Like how do you make
8:17sense of it? I do think OpenAI and
8:18Anthropic are pretty different animals
8:20from a capital efficiency perspective.
8:22And Anthropic clearly is has a
8:25dramatically lower cost per token than
8:27OpenAI.
8:28They just do. And you can just see that
8:31in the amount of money that they have
8:33burned
8:34to get to a roughly similar revenue
8:36scale. I think they have they have
8:37they've burned maybe 80% less than
8:39OpenAI.
8:41So as businesses, they clearly have very
8:44different structural ROICs. I think
8:46OpenAI is doing a lot I think Sarah
8:48Friar is one of the most exceptional
8:49CFOs. I think they're doing a lot of
8:51things to try to improve this. And
8:53they've secured a lot of compute more
8:55more than others
8:55>> They secured a lot of compute. That's
8:57another big difference. Um it turns out
8:59being aggressive really paid.
9:01But yeah, I just Anthropic at 900
9:04billion for 50 billion and
9:06you know, ARR and you know, I I But
9:09growing a thousand Yeah, growing at
9:11ridiculous rates. Maybe a true statement
9:14is that if Anthropic had all the
9:15compute, they'd probably be doing well
9:19north of a hundred billion dollars
9:20today.
9:22Maybe 150.
9:25And I do you know, they have clearly
9:26deprecated the intelligence of Claude.
9:28There's analysis Claude is even on Opus
9:31is generating 70% less tokens for the
9:33exact same question. And you know, as we
9:35talked about last time, token quantity
9:37equals quality of answer and quality of
9:39thinking at some level. You know, and
9:41there is an intelligence density per
9:43token that also matters. You know, I
9:45think I felt that as as a user. So, I
9:47think they would be doing materially
9:49more. 100, 150 maybe 200 billion. So,
9:53you might be buying it at more like five
9:57times
9:59unconstrained
10:01I'm going to make up a new number.
10:03URR unconstrained run rate revenue.
10:06>> [laughter]
10:06>> Yes.
10:08Why do you think they don't raise a 100
10:11billion dollars at a 3 trillion dollar
10:13valuation or something like this? Like
10:15if if you were the Anthropic CFO, uh
10:18Krishna's awesome, we just had him on.
10:19Or if you're the if you're Sarah,
10:20certainly if if the inbound I received
10:22following the Krishna episode is any
10:24indication, everyone I've ever met is
10:26trying to invest in in both these
10:28companies. So,
10:30I think it's wise.
10:32It the future is uncertain.
10:36You are clearly in a very capital
10:38intensive game, even if you are you
10:41know, Anthropic um I'm sure is at very
10:44positive gross margins on inference
10:46today. I can Anthropic probably starts
10:48generating cash this year if they are
10:50not already generating cash, which I
10:52think is probably the case.
10:55But still, you probably want to be able
10:57to raise more capital, access more
10:58compute. The world is uncertain. Ukraine
11:01is starting to really really win. How is
11:03Russia going to respond?
11:06You know, I think there's still a lot of
11:07uncertainty in Iran. All this
11:09uncertainty I think probably amplifies
11:11geopolitical uncertainty over Taiwan.
11:14So, it's an uncertain world. If if I
11:16think about Elon, Elon has always made
11:18investors money.
11:20He treats it like a sacred covenant. And
11:22as a result, because he's made people
11:24money for now 20 years, he has a
11:27superpower. And that is he can
11:29essentially raise as much capital
11:32as he wants whenever he wants.
11:34And I think it's wise that these
11:36companies are taking I don't know if
11:37that's how they think about it,
11:39but I do think being focused on making
11:42investors money
11:44is wise
11:46and creates benefits that don't just
11:49last for like a year or two.
11:51They can last for the next 20 to 30
11:53years.
11:54And the way Elon did this was sort of
11:56systematically underpricing SpaceX or
11:59whatever else. What is the actual
12:00method?
12:02Just never being greedy on valuation.
12:04Never pushing valuation.
12:06Just that simple. You know, my friend
12:08Antonio pointed out SpaceX compounded
12:10it, you know, low 30% per year for
12:13whatever that was, a decade.
12:15And and that was just cuz Elon was I
12:17think focused on
12:19preserving the superpower and having to
12:21trying to strike a fair balance between
12:23investors and employees.
12:25But I I think it's wise. But could
12:28Anthropic raise money
12:31at probably at least a 100% premium
12:35to this rumored latest mark? Of course.
12:38Most software companies try to maximize
12:40your time on their app to juice
12:41engagement. Ramp does the exact [music]
12:43opposite. Ramp understands that no one
12:45wants to spend hours chasing receipts,
12:47reviewing expense reports, and checking
12:49for policy violations. So, they built
12:51their tools to give that time back using
12:53AI to automate 85% of expense reviews
12:56with 99% accuracy. And since Ramp saves
12:59companies 5%, [music]
13:00it's no wonder that Shopify runs on
13:01Ramp, Stripe runs on Ramp, and my
13:03business does, too. To see what happens
13:05when you eliminate the busy work, check
13:06out ramp.com/invest.
13:10Felix by Rogo is a personal finance
13:12agent [music] that turns a single prompt
13:13into finished client-ready work using
13:15your firm's own templates, contexts, and
13:17standards. Send Felix an email like,
13:19"Take these comments and turn them for
13:21me." or "Update my tracker with the
13:22context of these emails." or "Run on
13:25ability to pay [music] math on this
13:26buyer. And Felix sends back finished
13:28PowerPoint decks, Excel models, and
13:30sourced research.
13:31Felix works the way your team already
13:33does, [music] delivering work quickly
13:34and accurately around the clock. Learn
13:36more at rogo.ai/felix.
13:38[music]
13:40OpenAI, Cursor, Anthropic, Perplexity,
13:43and Vercel all have something in common.
13:44They all use WorkOS. And here's why. To
13:47achieve enterprise adoption at scale,
13:49you have to deliver on core capabilities
13:51like SSO, SCIM, RBAC, [music] and audit
13:54logs. That's where WorkOS comes in.
13:56Instead of spending months building
13:57these mission-critical capabilities
13:58yourself, you can just use WorkOS APIs
14:01to gain all of them on day zero. That's
14:03why so many of the top AI teams you hear
14:05about already run on WorkOS. WorkOS is
14:08the fastest way to become enterprise
14:10ready and stay focused on what matters
14:11most, your product. Visit workos.com to
14:14get started.
14:15>> Let's get to the Watson wafers part of
14:16the discussion. [laughter] Always my
14:18favorite thing to talk about with you.
14:20Uh
14:21the importance of this infrastructure
14:23build-out. I feel like every time I feel
14:25like it's getting overheated and then
14:26the next time I talk to you, it seems
14:28like we should have done way more than
14:29we did. And you've studied S-curves and
14:32the steepness of those S-curves a lot.
14:34Uh and you know a lot about history.
14:36Talk us through how you're thinking
14:37about Watson wafers today as the key to
14:41inputs into this whole thing. Yeah, I
14:43would say I think capitalism is going to
14:45solve the watts
14:47shortage
14:49absent big regulatory or political
14:52blowback, which I think is a real
14:53possibility. The head of kind of data
14:55center infra investing at one of the big
14:57PE firms, you know, think Blackstone,
14:59Apollo, KKR said it used to be
15:03energy and chips
15:04were our biggest gating factors. Now
15:07it's zoning and approval.
15:09Much more important. And I think a lot
15:11of companies are waiting till after the
15:13midterms
15:15to take action in terms of maybe
15:17workforce reductions.
15:19Nobody wants to be
15:21you know, a piñata during the midterms.
15:23But, you know, you've seen a lot of
15:26companies that make turbines significant
15:28announce a plans to significantly
15:30increase capacity. There's like two of
15:32these machines that can cast these big
15:34blades.
15:35We haven't made one in 80 years in the
15:37West. We don't know how to make them
15:39anymore, etc. etc. etc. All of that is
15:42true and I
15:43and by no means am I
15:45minimizing, you know, the industrial
15:47engineering, you know, magic and
15:49artistry that goes into those, but
15:50capitalism is very good at solving
15:52problems like these over time. There's
15:54other sources of energy besides these
15:56turbines with a longer time frame. So, I
15:58think the watts shortage
16:01will probably begin to alleviate 27, 28.
16:07And then I think orbital compute will
16:09really solve that. And I do
16:11I do want to like reframe orbital
16:13compute because [snorts] I think when
16:15people hear data centers in space,
16:17they, you know, which we discussed in
16:18our last episode, they picture a
16:20Pentagon-sized building in space.
16:22They're like, "Well, we can't do that."
16:24That's not what it is.
16:26A Blackwell rack weighs 3,000 lb. It's 8
16:30ft high. It's 4 ft deep, 3 ft wide.
16:35It's racks in space. It's SpaceX has
16:38showed you an illustration.
16:40And it's a rack. That's the satellite.
16:43Um it's probably about the size of a
16:45Blackwell rack. It has these solar wings
16:47that are probably 500 ft long on each
16:50side. You keep it in a sun synchronous
16:53orbit so those solar panels are always
16:56at the sun.
16:57And then because it's in an exactly sun
16:59synchronous orbit,
17:00the radiator, which extends behind it
17:02for hundreds of feet,
17:05This is a common criticism, yeah. How
17:06are you going to cool it?
17:07>> Yeah.
17:07I've spent
17:09a lot of time at Starbase
17:11over the years and I've talked to a lot
17:13of SpaceX engineers.
17:15And I do think it is the most talented
17:17group of engineers on planet Earth. And
17:20they're very confident they have solved
17:21this. And they're not always confident.
17:25Like I think probably, you know, there's
17:27some engineering that needs to happen to
17:29turn the Starship into a Mars colloidal
17:31transporter. Will they do that?
17:32Absolutely. What are they more focused
17:35on? I'd say probably, you know, the
17:37repair and maintenance. Those are the
17:39two big, you know, the two big
17:40responses. The radiator and the and how
17:42do you repair the whatever issue goes
17:44wrong in the rack. And the answer is
17:46like until you have probably an, you
17:49know, floating optimists, you don't.
17:51Now, I do think Starship is going to
17:53change the space economy in ways we
17:55cannot imagine, particularly if
17:56regulation becomes a constraint to data
17:58centers. None of it's going to matter.
18:00You're going to sell as much orbital
18:02compute as you can make.
18:04And then obviously you link these racks
18:07using lasers traveling through vacuum,
18:09which are already on every Starlink. And
18:11it's just it's just mind-blowing to me
18:14that SpaceX operates the world's largest
18:17satellite fleet, which is
18:19like 98 or 99% of all satellites in
18:22orbit.
18:24Every Starlink, they're cooling it
18:26today.
18:27And, you know, I think Starlink V3 is
18:29going to operate at 20 kilowatts.
18:31A Blackwell rack is only a 100
18:34kilowatts. And people talk a lot about
18:37density. Well, if you're connecting the
18:39racks with lasers through vacuum, you
18:42know, you can make the rack bigger
18:44physically. You're focused on weight,
18:46not size. In a data center on Earth
18:49where you're trying to connect racks
18:51ideally using copper, minimize lengths,
18:54etc., etc. Cabling is a big cost. Um,
18:57you do want that rack to be small cuz,
19:00you know, copper when you can, optics
19:01when you must. But in space, you know,
19:03there's all sorts of things that SpaceX
19:05can do that I think maybe some of these
19:07naysayers are not contemplating.
19:10But it's just they operate more
19:11satellites than They have a 20 kilowatt
19:13satellite today. So maybe you just scale
19:15that up to 60 kilowatts to start. They
19:18seem very confident they're going to go
19:19right to 100 to 120.
19:21And they also the same company now
19:24also operates the largest data center on
19:27Earth.
19:28They have the world's best hardware
19:29engineers and all sorts of people,
19:32almost all of whom are not smart enough
19:35or practical enough to work at SpaceX.
19:39Are these armchair skeptics?
19:41>> [laughter]
19:42>> You know, I don't want to quote Larry
19:43Ellison, but somebody was, you know,
19:45being skeptical. And Larry And Larry was
19:47just like, "Listen, he's out there
19:48landing rockets. I don't see anybody
19:51else landing rockets."
19:52And the reality is is that 10 years
19:54later, no other company is consistently
19:57capable of landing and fully reusing an
20:00orbital rocket.
20:02And none of this works makes sense
20:04without reusability. That means you have
20:06to land it. I would like to redefine
20:08orbital compute as racks in space.
20:11Not giant floating Pentagon-sized
20:14data centers in space, which is just,
20:17you know, that's silly. But you can, you
20:19know, what makes a data center is you're
20:20connecting these racks with lasers.
20:23So it'll be racks in space that are
20:24connected with lasers into a virtual
20:26data center.
20:27And And if you think about that state of
20:29the world, let's say that all happens
20:31and we're really good at getting these
20:33things up economically, running matrix
20:35multiplication all over space. What does
20:37that mean for terrestrial data centers?
20:39Someone once said, um
20:42you know, America was going to suck as
20:44hard as it can on every energy source it
20:47can get. And I just think the same is
20:49true of compute.
20:51It's why I'm probably less worried about
20:53like an
20:54edge AI bear case than I was.
20:57We're going to consume as much compute
21:01as we can.
21:03And
21:04inference, I think is very sensible for
21:07orbital compute. Training will be done
21:09on Earth for a long time. So, I don't
21:12think that this is super bearish for
21:14terrestrial data centers. I think those
21:16are going to be valuable for
21:18my lifetime.
21:21But, I do think if you are in this
21:23ecosystem of power production and
21:26cooling
21:27and you are massively ramping
21:30capacity and you know, a lot of these
21:33capacity ramps are going to be hitting
21:35just as I think, you know, all of the
21:38silly skeptics
21:39start to understand that orbital compute
21:41is very real. Like, I think it's worth
21:43thinking long and hard about that if
21:45you're one of those companies. And then
21:47all sorts of cool stuff is happening in
21:48the interim, you know, we're getting
21:50really good at repurposing jet engines,
21:52you know, there's that Boom Aerospace
21:54that is doing this. So, there's a lot of
21:56capitalism is hard at work
21:58on on watts. On wafers though,
22:02it's just this group
22:04of, you know, plenty older humans in
22:09Taiwan
22:10who are the most important humans in
22:12Taiwan. They are the overwhelming
22:14fraction of the country's GDP, water
22:16usage, electricity usage. They talk
22:19about the silicon shield. They all view
22:22themselves as inheritors of, you know,
22:25Morris Chang's sacred legacy. I vividly
22:27remember like visiting Science Park
22:30more than 20 years ago
22:33and, you know, talking to them, "Do you
22:35think you could catch Intel?"
22:37And they said, "This is such a beautiful
22:39dream, but it's a dream for our
22:41grandchildren." And they did it. Partly
22:44because of Intel's self-inflicted
22:46wounds,
22:47but just they don't they think very
22:50differently. You know, one reason, you
22:52know, Jensen flies over there so much is
22:55he wants them to expand capacity. I do
22:56think it's wild that Jensen has never
22:58had a contract with Taiwan Semi. They do
23:01business on what seems fair in
23:03handshakes.
23:04Just fascinating. No contract.
23:06It's going to be fair over time. We're
23:08partners. We're going to be fair to each
23:09other.
23:10And the truth is, you know, based on
23:12every every prior market precedent for a
23:16foundational new technology like AI,
23:18you've always had a bubble. You know,
23:20Carlotta Perez wrote this great book
23:21about this. And basically, markets are
23:24efficient. They correctly understand
23:26that this is a foundational new
23:27technology.
23:29There's what Simpson calls a breakdown
23:31in diversity.
23:33Everyone becomes bullish on this new
23:35technology.
23:36And I am beginning to worry a little bit
23:38about a diversity breakdown.
23:40And then you get a bubble. That bubble
23:44funds the build out of this new
23:45technology, but supply gets ahead of
23:48demand.
23:50And you get a crash, and it's a
23:51particularly severe crash if it's a
23:53debt-fueled build out like the year
23:552000.
23:56And one thing I'm really happy about
23:58really good about the current build out
24:00is it's still overwhelmingly funded out
24:02of operating cash flows, which is a a
24:04really important fundamental difference
24:06versus year 2000. As is valuation, as is
24:09the fact that every GPU is running at
24:11100% utilization when 99% of fiber was
24:14unutilized. So, there's all these
24:15fundamental differences.
24:17But we do have to History doesn't
24:18repeat, but it rhymes. And and as
24:20investor, we have to be very cognizant
24:21of it.
24:23And recognize that based on the last 200
24:27years, you know, forget the internet
24:28bubble. We had a railroad bubble. A
24:29canal bubble. We should expect a bubble.
24:33And
24:34that's terrifying. Like nobody wants a
24:36bubble. A bubble is terrible. Reason
24:38it's terrible is if you're valuation
24:39sensitive, you like massively
24:41underperform. You get fired by probably
24:44all your clients. George Vanderheide,
24:46who
24:48um is is is no longer with us, great
24:50uh Fidelity portfolio manager.
24:53He fought the bubble in '99.
24:55And he retired in two in early 2000 cuz
24:58I think he just couldn't couldn't take
24:59it.
25:00He knew it was wrong.
25:02And you know, his his clients were
25:04deeply skeptical. George, you're out of
25:06step. You know, he had he had white
25:08hair. He's truly great man.
25:10I only overlapped with him briefly, but
25:11he was a very important mentor and
25:13friend to my good friend and mentor
25:16Jennifer Yurig. So, I have a lot of
25:18Vander Heiden DNA through her.
25:21Like he was the same person who said
25:22being early is the same thing as being
25:23wrong. George retired cuz he can't take
25:26the underperformance and he can't take
25:29clients saying, "What's wrong with you?
25:31You don't get it." And he has like 40%
25:33of his fund did tobacco, 40% in home
25:37builders.
25:38And literally he under he probably
25:40outperformed the Nasdaq
25:43by like
25:4420 or 30X over the next 3 years, okay?
25:48And I have been optimistic that this
25:51fundamental shortage of wafers, which
25:53really today is controlled by Taiwan
25:56Semi, will prevent one.
25:58If Taiwan Semi did what Jensen wanted, I
26:00think Nvidia could sell $2 trillion of
26:02GPUs
26:04in 20 in 26 or 27. Maybe 2 and 1/2
26:07trillion. Maybe 3 trillion. But there is
26:10a limit where consumers would consume so
26:12much that you probably would be in an
26:15overbuild. And so, Taiwan Semi, if we
26:17don't get a bubble, like we need to
26:18throw a party for them because they will
26:20have single-handedly prevented a bubble,
26:22okay? You are starting to see companies
26:26go to Intel
26:28and Samsung. Let's just assume TSMC
26:30stays super supply constrained versus,
26:32you know, the latent demand. Like what
26:34what happens?
26:36one of,
26:36>> [snorts]
26:37>> you know, the history of markets is I
26:39don't know who, but one of Intel and
26:40Samsung, they're not going to stay
26:42disciplined. They will break.
26:44And then at some level that will force
26:47everyone else to break.
26:49So,
26:51like I think a lot of this may come down
26:53to the degree to which Taiwan Semi can
26:55maintain a lead over Intel and Samsung.
26:59And you got to remember it's whatever it
27:01is, it's 9, 12, 15 months.
27:02>> Sort of like the leading node edge, you
27:04mean? Exactly. You know, the pace at
27:06which they expand capacity.
27:09Like if I were to watch one thing to
27:10understand where there's a bubble, it's
27:11Taiwan Semi's capacity decisions.
27:14And I think there's a Goldilocks zone
27:17where they
27:19expand enough
27:21they make it hard for Intel or Samsung
27:24to really, truly emerge as like a um
27:28at scale second source with something,
27:31you know, well north of 30% market
27:34share.
27:35And yet they also keep this fundamental
27:38constraint on wafers
27:41that
27:42you know, helps us avoid a bubble. And
27:44then obviously, I think the Terra Fab
27:47um is going to play into this, too. Say
27:49more about that. For people that are not
27:50familiar.
27:50>> the Terra Fab. It's a SpaceX, I believe
27:53Tesla's involved as well. Um joint
27:56venture to build the world's largest fab
27:59here in America. And I'm I think they're
28:02going to be successful. One, they have a
28:04partnership with Intel, which is very
28:06important. Um because they're getting
28:08access
28:09to 50 years of institutional knowledge.
28:12That's just, you know, a 9 months, a few
28:14quarters, 12 months, three to five
28:16quarters behind the front. That's an
28:17advantage.
28:19It's also an advantage that I believe
28:21the Terra Fab is going to get attention
28:24from the A teams at all the semi-cap
28:26equipment companies. Like one big reason
28:27Taiwan Semi caught up is ASML and KLA
28:31Tencor and Lam Research and Applied
28:33Materials. They wanted them to catch up.
28:36They didn't They don't like having a
28:37monopsony.
28:38And so the A teams were in Taiwan
28:40working Intel made some mistakes.
28:43And presto. And so the A teams will will
28:46be here cuz of Elon's reputation in in
28:49hardware engineering.
28:51And then just to a degree that I think
28:54is uh maybe hard for people to imagine
28:57in America.
28:59Um where, you know, politics has
29:00replaced religion cuz Elon had his foray
29:02into politics that makes it hard for
29:04some people in America
29:06to see him clearly, which is sad because
29:09I do think
29:10you know, he's probably doing more for
29:11America than any other American. You
29:14know, he's single-handedly bringing
29:16manufacturing back to America. He's
29:18revived defense tech. SpaceX is in some
29:20ways the most important defense
29:22contractor in America.
29:24You know, what he's doing with Starlink
29:25is amazing for the world. He's creating
29:28all these blue-collar manufacturing
29:30jobs, which is like a goal I think of a
29:31lot of liberals
29:32and good for America. He's done more
29:34than any living human to decarbonize the
29:36world. And if you are upset about data
29:39centers on Earth for environmental
29:40reasons, well, here you go.
29:42>> [laughter]
29:44>> Uh so it's it's sad,
29:46but he is a living deity
29:49in China,
29:51Taiwan, South Korea, and Japan.
29:55And having watched him for a long time,
29:59what he's going to do is they're going
30:00to recruit the best people
30:03because the best engineers
30:05want to work for Elon,
30:07especially in hardware engineering. He's
30:10going to recruit incredible engineers.
30:12And then they'll be next to the next to
30:14Terrific Um they'll be a Taiwan town.
30:16Oh, these are your favorite restaurants?
30:19I'm going to move them and their whole
30:20staff
30:21from Taiwan to Texas. And we're going to
30:24make everything the way they like it.
30:26And then we'll have Japan town. Same
30:28thing. We're going to have Korea town.
30:29We're going to have all these things
30:31exactly [clears throat]
30:32but dialed to recruit the best
30:35engineers.
30:37And that's just not the way that
30:40the people who run Intel and Samsung
30:42think.
30:44So, he's going to have the best talent.
30:45He's going to have the A teams
30:47at the wafer fab equipment companies.
30:49He's He has Intel, which is important.
30:51It's so good for all of any
30:54administration's political goals.
30:56And I think it's different enough that
30:58it will not alienate Taiwan semi.
31:00>> And these have long lead times, right?
31:02So, like TerraFab is going to be pumping
31:03out Nvidia cheaper whatever GPUs
31:06whatever chips like quite quite a long
31:08time from now. Elon tends to do things
31:10differently. Everybody else is taking 3
31:11years to build a data center. He built
31:13one in 122 days.
31:15>> [laughter]
31:15>> You know.
31:16Samsung had to give him an office in
31:18their fab in Texas cuz he was so unhappy
31:21about like the pace at which they're
31:22expanding and building. We'll see. Are
31:25you surprised by
31:27You mentioned Deep Seek earlier. The
31:29simple reaction to that was, okay, these
31:30models are just going to get
31:3295% as effective for some tiny fraction
31:35of the cost of still Chinese open source
31:37models like we'll be able to use these
31:39for most of what we want to do. Fast
31:40forward it a little bit of time, you
31:42know, 2 years from now, there's no
31:44reason I have to spend a million dollars
31:46a year in my small little firm on on
31:47tokens or something. But then the actual
31:49reality seems quite different than this.
31:51And I'm curious why there's that
31:53dissonance in your mind.
31:54>> I do think it's the fascinating the
31:56returns to the frontier.
31:58All the economic returns to AI at the
32:01model layer,
32:02not all of them, but an overwhelming
32:04amount of them have been at the
32:05frontier,
32:07which is surprising to me.
32:10I think it's been surprising to a lot of
32:11people. And I think
32:14this is one of the most important
32:16questions to be answered, and you need
32:18to have a hypothesis on it as an
32:20investor. Are frontier tokens going to
32:23continue
32:24capturing the overwhelming majority of
32:27economic value created at the model
32:29layer?
32:30And it is surprising. Like I just I
32:31remember when Jim and I 3.1 Pro came
32:34out.
32:35And it was it was mind-blowing to me. It
32:37was so good.
32:39And today, it's intolerable.
32:41Intolerable.
32:43And, you know, there's probably a little
32:44bit of a dynamic where companies
32:46prototype with frontiers, then when they
32:48put something into production, you're
32:50hearing a lot of people do use for
32:51attacks or, you know, open source.
32:54But still, it is it is a fact today that
32:57the overwhelming majority of these
32:59economic returns come from frontier
33:00tokens.
33:01And that's surprising. And whether or
33:04not it continues, I think is a very
33:06interesting question.
33:08And I'm much more open-minded to that
33:10having had the experience I've had with
33:11Gemini 3.1.
33:14And then Opus.
33:16Um and then I do use Grok 4.3. It is on
33:19the Pareto frontier. Like the companies
33:21that are on the Pareto frontier are And
33:23this is, by the way, a big change in a
33:25consequence of what we talked about last
33:27time, Google losing their per cost token
33:30leadership as a result of making very
33:32conservative design decisions with TPU
33:34v8 to try and
33:35take it away partially from Broadcom and
33:37Nvidia
33:39um continuing to make aggressive
33:40choices. Uh but Google dominated the
33:43Pareto frontier. The Pareto frontier
33:44being intelligence first cost. And I
33:47think this is the most important thing
33:48to look at to analyze AI labs. Google
33:50dominated that 9 months ago. At every
33:53point on the Pareto frontier,
33:55OpenAI, xAI, and Anthropic were inside
33:59of them.
34:00Now, the Pareto frontier is dominated by
34:02Anthropic, OpenAI,
34:04and then Grok 4.3 is on the Pareto
34:07frontier. It's clearly like the,
34:09you know, the best lowest cost 500
34:11billion parameter model. And then Gemini
34:133.1 is like hanging onto the Pareto
34:17frontier. And if I were to bet, I'd bet
34:19that they're subsidizing that out of
34:20pride. I'd just say, well, one,
34:22violation of Richard Sutton's bitter
34:24lesson is for sure the biggest risk to
34:26this trade. To all of AI.
34:28Now, the closer someone is to AI, the
34:30more skeptical they are this will occur.
34:33One thing I think contributed to
34:34weakness in March was, you know, a much
34:37more stupid version of deep seek, which
34:39is a thing called turbo quad.
34:41And turbo quad is some Google memory
34:43optimization that was written up in a
34:45paper a year ago. And then during the
34:47middle of an agreement, while Google was
34:49negotiating
34:51with Micron, Samsung, and Hynix to sign,
34:53you know, some LTA that would lock in
34:55really high prices for a long time,
34:57they released this. You know, what
34:59people do is always more important than
35:00they say, and they just kind of
35:01publicize it on X.
35:03And it goes viral. Like, "Oh my god,
35:05DRAM is cooked. Here's this DRAM
35:07optimization." I was unable to find a
35:09single AI engineer on planet Earth who
35:12believed that turbo quad would have any
35:14impact on DRAM demand. But nonetheless,
35:17a violation of Richard Sutton's bitter
35:19lesson, you know, more compute will
35:21always outperform human algorithmic
35:23ingenuity. More compute and data are
35:24chin beyond Chinchilla optimal, I guess
35:27what what people increasingly do today.
35:29That's a real risk, man.
35:31And I think the people who are building
35:33these models are skeptical of that risk.
35:36The reason I am a little less skeptical
35:39is I think we are very close to ASI. And
35:42who knows if the bitter lesson holds for
35:44400 IQ models. Just, you know, or maybe
35:48we get a temporary
35:51period where these, you know, if you get
35:52to ASI, the first thing it wants is
35:55probably to be smarter and have more
35:56resources. How does it do that? It makes
35:58itself more efficient. I think that
36:02is an actual risk that humans, the
36:05bitter lesson, literally I believe
36:07includes humans in it.
36:09So, we're about to find out whether the
36:11bitter lesson will find out if applies
36:13to a 300 IQ AIs, then 400, then 500, and
36:16600. And at some point, we may have like
36:20a temporary violation of the bitter
36:22lesson
36:24based upon AI and ASI. So, I'm curious
36:27how you think about some other parts of
36:30the innovation around the model,
36:32continual learning and memory being two
36:34that people seem to be most focused on
36:36as things that might create yet another,
36:38you know, new paradigm that we would
36:39enter. What do you think about the role
36:41of those two things?
36:41>> Yeah, well, I think we've done a lot
36:43with memory through these harnesses. And
36:45it turns out that harness engineering
36:48is
36:49not as important as the model, but it
36:52really matters. And these harnesses and
36:54these models are increasingly being
36:56co-developed. One of the big things a
36:58harness does, we used to think of it as
36:59like a a runtime that the model operates
37:03in. It knows where the pool tools are.
37:06It you know, it creates context, memory,
37:09state. Um, you know, has very specific,
37:13you know, prompts or instructions. And
37:16just Makes a huge difference. Even
37:18[clears throat] simple versions. It
37:19makes an incredible difference. And I
37:20think the last time I was on here or one
37:22of the other times I just said like,
37:23"Hey,
37:24as an investor, it's very important that
37:27you pay for the $250 a month version to
37:30get like your own intuitive sense."
37:32That's no longer possible. To understand
37:34what Frontier AI is capable of today,
37:37even for like a non-coding use case, you
37:40need to have Claude code or Codex. And
37:42you need to be on an enterprise plan.
37:44And the reason for this is and this is
37:46another, I think and
37:48this is another dynamic that's enabled
37:50by Google losing their
37:52cost leadership, is these AI models just
37:55shifted to usage-based pricing. And if
37:58you're on that $250 or $300 or $280 a
38:01month plan or whatever it is,
38:03you're getting severely rate limited.
38:05You're getting a lobotomized version of
38:07the AI.
38:08Because like we talked about, Claude now
38:10produces 70% less tokens. You want the
38:13tokens that Claude and its harness
38:15really think it needs to produce to get
38:17you a good answer, you need to be on a
38:19usage-based plan. And by by way,
38:21this is
38:22so bullish for AI. I was a telecom
38:24analyst in '05 to '07.
38:26And cellular had been a great growth
38:28industry really for last 10 years. And
38:30the reason was
38:31you had a combination of fixed pricing.
38:34You had 900 minutes for whatever it was,
38:37and then usage-based pricing over that.
38:39And when did cellular stop being a great
38:42growth industry? When everybody just
38:43went to all you can eat.
38:45And And by the way, long distance is the
38:47same thing. AI is just shifting from all
38:49you can eat to pay by the drink. And it
38:52turns out people really like to talk to
38:53their friends long distance. They really
38:55like to talk to their friends on the
38:56phone. And people really like to use AI.
39:00And particularly now that one person can
39:01have 100 agents working. So, I think
39:03this shift to usage-based pricing
39:06is probably why you will see OpenAI and
39:10Anthropic exceed well over $200 in ARR
39:13this year. Because not only is more
39:15compute going to become online, but
39:17they're going to be able to push
39:19frontier token pricing with these usage
39:21enterprise models.
39:23But it's it's sad. It's sad for the
39:24world. And cuz it just means if you
39:26can't afford that, you're not at the
39:29frontier. But yeah, continual learning,
39:30man. I mean, if we solve that How do you
39:33conceptualize that? Like There's so many
39:34mysteries about the human mind. Like
39:36we're such sample efficient learners
39:40relative to AI.
39:42Like I forget what it is, but like an AI
39:44needs
39:44>> Orders of magnitude. Yeah, many orders
39:45of magnitude. Now, we have a crude
39:47variant of continual learning today when
39:50something is verifiable. And that's just
39:53you know, reinforcement learning during
39:54mid-training.
39:56But yeah, continual learning is a model
39:57that dynamically adjusts it its weights
40:00or adjusts in some way in real time.
40:03Like as a human, That's what you do.
40:05Yeah, like if I the first time I touch
40:08or, you know, put my hand in a fire,
40:10I've learned I never put it in there
40:12before.
40:13That model today needs to put its hand
40:15in the fire a million times
40:18and then have, you know, the designers
40:20effectively put a fire in the next
40:23training run or an RL gym for it to
40:26learn. I think it has to be dynamically
40:28updating the weights, but I think people
40:31are working on really smart techniques
40:33beyond this.
40:34But if we get that,
40:37then we have a really fast takeoff. And
40:39people seem
40:42confident that continual learning
40:45is kind of just around the corner. And I
40:47do think this is like the third big
40:50question. Bitter lesson violation as
40:52result of ASI are less likely. Human
40:55ingenuity. Will frontier tokens still
40:57command the premium they do?
40:59And will you get continual learning and
41:01if so, when?
41:02As your business scales up, everything
41:04gets more complex, especially your
41:05compliance and security needs. With so
41:07many tools offering band-aids and
41:09patches, [music] it's unfortunately far
41:10too easy for something to slip through
41:11the cracks. Fortunately, Vanta is a
41:14powerful tool designed to simplify and
41:16automate your security work [music] and
41:17deliver a single source of truth for
41:19compliance and risk. There's a reason
41:21that Ramp, Cursor, [music] and Snowflake
41:23all use Vanta. It frees them to focus on
41:25building amazing, differentiated
41:26products, knowing that compliance and
41:28security are under control. Invest like
41:30the best listeners get a special offer
41:31of $1,000 off Vanta
41:33>> [music]
41:33>> when you go to vanta.com/invest.
41:37I know firsthand how complex the tech
41:38stack [music] is for asset management
41:39firms. And seemingly every new tool and
41:41data source makes the problem even
41:43worse, adding more complexity, more
41:45headcount, and more risk. [music]
41:46Ridgeline offers a better way forward,
41:48one unified platform that automates away
41:50that complexity across portfolio
41:52accounting, reconciliation, reporting,
41:53[music] trading, compliance, and more,
41:55all at scale.
41:57Ridgeline is revolutionizing investment
41:58management, helping ambitious firms
42:00scale faster, operate smarter, and stay
42:02ahead of the curve. See what Ridgeline
42:04can unlock for your firm. Schedule a
42:05demo at ridgeline.ai. [music]
42:07What is the role of new chip companies
42:10in all of this? Like we talked a lot
42:11about Nvidia and you know, their their
42:13sort of relationship with TSMC and Intel
42:15and all these sorts of things. There's a
42:17thousand flowers blooming, I think
42:19literally probably a thousand flowers
42:20blooming, trying to create a new chip to
42:23address some part of this bottleneck.
42:26I'm curious how you process this space,
42:28this opportunity, what role it will
42:29play, what role they'll play. So, I
42:31think this is good and healthy for the
42:32world. It's good for Jensen, too.
42:35Um you know, because a different
42:36administration might take a different
42:39view. Competition, I think, is good for
42:40everyone. In in tank design, they talk
42:42about the iron triangle. The iron
42:44triangle's tank design is that all
42:46designers of a tank, they have to make
42:47trade-offs between attack, defense, and
42:50mobility. And you know, for obvious
42:51reasons. The more defense you have,
42:53which is your armor, the heavier the
42:55tank is, the less mobile it is. So, you
42:57have to live in this triangle and make
43:00trade-offs. Okay? Like the Merkava in
43:03Israel, it's optimized for defense.
43:05Russian tanks and like the Leopard are
43:08generally more optimized for mobility.
43:10Chip design is the same.
43:12And you
43:13there there are these fundamental
43:15constraints imposed by the laws of
43:17physics as embedded in the Taiwan semi
43:20design rules that you need to live
43:23within. And
43:25you have TPU, Trainium, and AMD, which
43:29are all
43:30um
43:31you know, essentially trying to be a
43:33better GPU.
43:35And today, I think probably Trainium is
43:38doing the best. Now, nobody's a better
43:39GPU. But Trainium is is, I think,
43:42they're you know, they're they're
43:43tugging on Superman's cape.
43:46And and this is that I'm starting yet.
43:48The Trainium 3 needs to ramp into
43:50production cuz it has a switch scale-up
43:52network, which you really need to
43:53economically inference MoE models. You
43:56know, a lot of companies have a Taurus
43:58architecture. Um that that's where
44:00Google was. And AMD, we'll see. The
44:02MI450, we we we don't know yet. We'll
44:05see. We probably know more about Tridium
44:073 than the MI450. But, that's a hard
44:09game to play. So, you have to do
44:12something different. And you have to do
44:15something different that is also hard to
44:19do. So, I think the best path for these
44:21startups, like my rule of thumb is 1%
44:24market share is going to be worth 100
44:25billion. 100 billion is a pretty good
44:27venture outcome. I think what Jensen
44:28would say is like, "Okay, if something
44:30somebody does something different and it
44:33gets to 1 or 2 or 3% share,
44:35we'll make that chip."
44:37And that's that's coming for everyone.
44:40But, if you're trying to make a better
44:41GPU, good luck. If you were doing
44:43something
44:44different, it also needs to be hard to
44:47do. And you can make different
44:49tradeoffs, you know, the disaggregation
44:51of prefill and inference really have
44:53opened the aperture
44:55um for making these different pre
44:56tradeoffs because you can make very
44:58aggressive tradeoffs for decode,
45:00aggressive tradeoffs for prefill.
45:02Prefill being taking in the context,
45:03decode being, you know, write the
45:05output. Yeah, I have a great colleague
45:07named Andrew Fox who said, "Picture, you
45:09know, a British naval ship from the 18th
45:10century. Prefill is loading the cannon,
45:13decode is firing it." And what prefill
45:15literally is is just the model
45:16understanding the question, the prompt,
45:19and then kind of keeping track of its
45:20own deco- if if it's own answer. And
45:22that is fundamentally a memory capacity
45:25bound problem. Decode is the process of
45:27generating new tokens and that is memory
45:29bandwidth constrained.
45:31And so, if you're a chip designer, this
45:33gives you a richer canvas to to paint
45:35on. But, even so, it needs to be hard
45:38cuz if you make different tradeoffs in
45:40that iron triangle to optimize for
45:42memory capacity and they're not hard
45:44tradeoffs to make, well then, Nvidia is
45:47going to make those same tradeoffs.
45:49They get better prices from Taiwan Semi
45:51than you're ever going to get. Um and
45:54good luck. Good luck. And they have the
45:56advantage of working with every model
45:58company and optimizing in designs. By
46:00the way, another very funny thing is if
46:02you're a VC
46:04and you're investing in a semiconductor
46:06company that is telling you they are
46:08going to have an advantage cuz of a
46:09Taiwan semi process that they have
46:12special access to. I promise you
46:15that Jensen saw that process
46:18when it was a twinkle in Taiwan semi's
46:21eyes
46:22and it they know more about it than this
46:25little company with
46:26200 people can imagine. Taiwan semi,
46:29everybody supply chain is showing Jensen
46:31everything. The same way they're showing
46:34Amazon everything, AMD everything,
46:36TPU everything. And that's another
46:38reason don't go try to make a better
46:39GPU.
46:40So you can do something different. You
46:42can paint in the prefill canvas. You can
46:43paint in the decode canvas.
46:46But you also have to do something hard
46:48because if it gets to scale,
46:50you're going to have those four
46:51companies as very fast followers. My
46:53firm was a was a um
46:56venture investor in Cerebrus. What
46:58Cerebrus has done is something hard and
47:00fundamentally different. Wafer scale
47:02computing.
47:03And it it comes with a set of tradeoffs.
47:06But that
47:07architectural decision they made was
47:09hard
47:10and lets them do something that no one
47:13else can do. And we'll find out how big
47:16that is. And you know, they're working
47:17on really cool things like um one of the
47:20problems Cerebrus has, once you start
47:22needing to glue a lot of chips together
47:24and scale up networks or scale out
47:25networks,
47:27you need a lot of IO. And IO is bound by
47:30what's called the shoreline, the sides
47:32of the chip. And so Cerebrus has an
47:34overwhelming ratio of on-chip computer
47:37memory relative to shoreline IO.
47:40Well, they're really smart people. They
47:41did something really hard. They're
47:43trying to see if they can put an optical
47:44wafer right on top of that. And then
47:46that solves that problem. Um I'm sure
47:48they're looking at hybrid bonding of
47:50DRAM, you know, to get around these
47:52alleged limitations that are not true. A
47:54Cerebrus machine can theoretically run
47:56any size model. So there are of models
47:58where they're much better than other
47:59sizes.
48:01So, Cerebras, what I think is
48:02interesting is they did something
48:03different that's hard to do. Really hard
48:05to do. Wafer scale computing. So, I do
48:07think there's a role for these. And, you
48:10know, I would just encourage them all.
48:12Make a different trade-off.
48:14And
48:15try and do something hard.
48:17Cuz
48:18everybody's going to get funded after
48:20the Cerebras IPO. It's not going to be a
48:22problem. But, it took it took Cerebras
48:24three generations of chips
48:27to get it right.
48:29And it's really hard. Like, Andrew
48:31Feldman, the CEO, you can just see
48:35how hard it was what he did and that
48:39whole team did to get where they are
48:42today.
48:43And they need to have the grit to do
48:45that, the resilience. This first chip is
48:46a failure. It happens. Can you come back
48:48and make a second chip? But, the one
48:50last thing on this topic, this is going
48:51to be amazing for the useful lives of
48:54GPUs and may single-handedly save
48:56private credit.
48:57>> about that. What do you What do you mean
48:58by the private credit?
48:59>> Well, just, you know, private credit,
49:01they're in pain from these SAS loans.
49:02And however much they're marked down,
49:04they probably need to be marked down
49:05more. Cuz if the public companies are
49:06struggling to to adapt, how's like a
49:08debt laden company going to going to
49:10adapt? Um and invest in what is a very
49:14different margin structure business.
49:16But, there's a lot of private credit in
49:18GPUs, too. They were underwriting that
49:20to I think three or four years.
49:22And but the disaggregation of inference
49:24means
49:26that I think these GPUs are going to
49:28have 10 or 15-year lives. The AI
49:30skeptics are like, "Oh, these companies
49:32are all cooking their books. You know,
49:33the useful life of
49:35GPU is only a year or two. The useful
49:36life of a CPU is only four years cuz the
49:38rapid technological change." No. What
49:41rapid technological change has done with
49:44the disaggregation of prefill and
49:45inference
49:46is mean that you you know, you can put a
49:48Cerebras system or Groq LPU's that
49:50Nvidia acquired
49:52and effectively in front of a hopper or
49:54even an ampere use that hopper and
49:56ampere for prefill and extend the useful
49:58life of that GPU
50:00until it melts. Now they do melt. They
50:02do melt so they have a time but you know
50:04maybe you don't have to run them
50:06as fast. This is going to be really good
50:08for the whole private credit industry.
50:10It's going to help finance the AI build
50:11out cuz if you can start to finance GPUs
50:14at more like you know
50:165% or 6% instead of I think CoreWeave's
50:18lowest financing was like low sevens.
50:20That actually mathematically changes the
50:22cost of finance this build out. We had
50:24this technological innovation that
50:27it's going to lower the cost of
50:28financing, extend the useful life of
50:29computer on earth. And then I do think
50:31the one last thing that's interesting
50:32about that is um
50:35my friend Jamin from Coatue just did a
50:37podcast and Coatue had a deck and they
50:39talked about hey
50:41you know the sellers of shortage are
50:42doing so much better than the buyers of
50:43shortage. Buyers shortage being you know
50:45the the hyperscalers.
50:48But if you own a giant installed base of
50:52what is currently in shortage that's
50:55also a very very good place to be. And
50:57we're hearing you know CPUs are way more
50:59important than they were in an agentic
51:00world. They do all these things around
51:02orchestration, tool calls, etc. etc.
51:03etc. The biggest CPU fleets in the world
51:06sit at the hyperscalers. So I think some
51:08of these hyperscalers may have
51:10you know may may catch up a little bit
51:12to the sellers of shortage.
51:13>> I want to talk about this idea of
51:15different and hard applied outside of
51:17the infrastructure piece of this. So now
51:20you're starting to interact with new
51:21founders, um existing CEOs and founders
51:24that have to adjust to this new world.
51:26What are you seeing like the most AI
51:28native founders that aren't building
51:29chips or infrastructure or models but
51:32just people using this technology to
51:33build other stuff. How do they feel the
51:36most different to you if if you've
51:37observed differences?
51:39Well one I do think this is just for
51:40chip design. To me it's always been a
51:41fundamental question for venture.
51:44So there are different ideas that are
51:47obvious to everyone on planet Earth as
51:48soon as they hear it. And if that's
51:50where you are in venture, if it's not
51:52hard to do, if it becomes obvious to the
51:54world before you have built
51:58scale, scale is the ultimate advantage,
52:00you're in trouble. And the great thing
52:02Amazon had was,
52:05you know, it was obvious to a lot of
52:07people, but it wasn't obvious to the
52:08retail CEOs. And Amazon, they were very
52:11smart.
52:13Any e-commerce company that VCs invested
52:16in, they would destroy.
52:18They'd be like, "Oh, that's so cute.
52:20We're going to We're going to take our
52:21margins in that to negative 10,000%."
52:24And that's what like like the guys at
52:26Wayfair, they did something hard. And
52:27Amazon tried to kill them and they
52:28failed. Those were like tough
52:30operationally,
52:31like really competent CEOs. For me in
52:34venture, I always look,
52:35is this going to be obvious to the world
52:38before this company could build scale?
52:41Or is this both not obvious, different,
52:45and really hard to do?
52:47I think a lot of founders are really
52:49struggling with this
52:51in AI. Like I think people are
52:56becoming worried, you know, today in
52:58that in Jensen's five-layer cake of AI,
53:02you know, the profits, they're accruing
53:03to energy, they're accruing to data
53:05centers, they're accruing to chips,
53:07they're accruing to models, not really
53:09accruing to the applications. Cursor and
53:12Cognition,
53:13you know, got to a scale. You know, they
53:15focused on coding, you know, 18 months
53:18ago the people were focusing on coding.
53:19OpenAI was doing everything under the
53:21sun. The people focused on coding were
53:22Cursor, Cognition,
53:24and
53:26Anthropic. And it was really righteous
53:27focus on code.
53:29Um I'm John Massaad, the founder of
53:31Replit, tweeted something that I thought
53:32was so smart. Just it was something
53:34like, you know,
53:36bitter lesson adjacent is the fact that
53:38coding might be the shortest path to ASI
53:41and useful AI. Cuz if you really go to
53:43coding, you can write yourself code to
53:45do anything. And so I think it was
53:46really smart of those companies to focus
53:48intensely on coding. And I think they
53:50all probably got to
53:52a scale where they they have a place. I
53:54think Cognition is doing something
53:55really, really different. But I think a
53:57lot of founders are really struggling,
53:59man. They're really struggling.
54:03And you know, I think they're trying to
54:04get confidence that in niche areas
54:08that they can get to them and get like a
54:12you know, a data moat
54:13before the model companies get to that
54:16niche. Or that it's a small enough niche
54:18that the model companies won't do it
54:20themselves, but it can still produce a
54:21different outcome.
54:22>> Is this related to what you would call
54:23like the token path? I know you've used
54:25that phrase with me before. Yeah, he
54:27comes from a guy at Altimeter, Jamin
54:29Ball. He just said, "If you're a
54:30software company or an AI company of any
54:32kind, you have to be in the token path."
54:34So Databricks, that's in the token path.
54:36Comparable companies are in the token
54:38path. If you're not in the token path
54:41and you're not in some really niche
54:45thing,
54:46life may be hard. And even for these
54:49vertical niches,
54:50I think if you talk to the people at the
54:53model companies,
54:55they're even skeptical of some of these
54:58because all of the data that's, you
55:00know, being generated in these niches
55:02come from humans. But then you're
55:03betting that you're able to use that
55:05proprietary data in this narrow vertical
55:08to train a model that's lower cost than
55:11the Frontier Labs can ever get to. Maybe
55:13that's a good bet. But I just think you
55:14have to be very, very careful. Now, on
55:17the other hand, if the returns to these
55:20frontier tokens relative to other tokens
55:22come down,
55:24there's going to be an explosion in
55:26value creation at the application layer.
55:29And I think another really important
55:31point is
55:34I have a belief
55:37that whenever he wants
55:39Jensen can probably get pretty close to
55:41the frontier.
55:43With his own model. With his own model.
55:45They're doing some really cool things in
55:46pneumatronics.
55:47>> to monetize your compliment as Sklansky
55:49would say.
55:50>> don't think he wants to do that. That is
55:53what OpenAI
55:55and you know, Anthropic are kind of
55:57trying to do to him
55:59unsuccessfully.
56:01But so it's just like he's a very
56:02logical thinker. This is the logical
56:04counter move.
56:06And I think you will see that like open
56:08source frontier, which today consists
56:11of, you know, Chinese models with stolen
56:15American tokens, you know, somebody told
56:16me that like Deep Seek
56:19uh the latest one or maybe the original
56:21one was only 150,000 reasoning traces.
56:23There's many ways to launder this if
56:25you're Chinese company.
56:27You know, you can hit all these
56:29different APIs. You can make it hard.
56:31Now, the American labs are working
56:33really hard on anti-distillation
56:34technology, but I I I just think Chinese
56:37open source, they're doing really
56:39impressive things in a very resource
56:40constrained way, but there's a lot of
56:42distillation. And this is why
56:45I think in addition to there not being
56:46enough compute to serve Mythos,
56:49just they did not want it to be
56:52distilled. They wanted to use Mythos,
56:55you know, distill it themselves, use it
56:57to RL their next model, whatever it is.
56:59And then I think what they and
57:02eventually I think if OpenAI gets to,
57:04you know, economics I feel good about
57:05anyone on the frontier will do is just
57:07say
57:08you know, there's going to be some very
57:10interesting game theory because it's it
57:12is it's a new kind of prisoner's
57:13dilemma. You know, we talked about the
57:14old prisoner's dilemma being just around
57:16like, "Hey, you you're in a prisoner's
57:18dilemma where you have to spend." The
57:20new prisoner's dilemma is going to be if
57:22you were at the frontier, do you release
57:24that model via API or not?
57:26And if everyone at the frontier agrees
57:30not to do that, then Chinese open
57:32sources
57:33quickly
57:34if one person defects, they're going to
57:37have the best model,
57:38they're going to have a lot of revenue
57:40and cash flow, and then of course
57:41resources equal intelligence, so they'll
57:43start to pull ahead and then that will
57:45lead to, you know, everybody else
57:47releasing it. So it's a new game theory.
57:49It's kind of the same game theory that
57:50you have with Taiwan Semi,
57:52Samsung and Intel. The reality is like
57:54if if a company like Nvidia were or AMD
57:57were to ever really really use one of
58:00these other foundries, that foundry
58:02would get better really quickly. So I do
58:04think
58:06Jensen is going to keep open source
58:09a certain time frame behind the
58:11frontier. I think that's going to be a
58:14very interesting thing to watch. And
58:15then by the way, open source gets
58:17monetized. There's this misnomer that
58:18open source is free. Open source tokens,
58:20they cost energy, they can, you know,
58:22they cost energy to produce, you need to
58:23make up on GPUs, and the open source
58:25model companies almost always get a
58:27revenue share. How are you preparing a
58:29trade ease for the world of Mythos 3,
58:33Mythos 4? We're just trying to over
58:35invest in cyber security, you know,
58:36something I've like, you know, said in
58:38multiple forums and I really believe is
58:40you everybody needs to have a safe word.
58:43Everybody needs to go
58:46leave your digital devices behind,
58:47literally go to the ocean and have a
58:49family safe word or a company safe word.
58:52And it can't be one that can be like
58:53socially engineered. And this is just to
58:55avoid like cyber crime where like what
58:57looks like your son or your daughter or
59:00your your grandparents or your parents
59:02or whatever FaceTimes you,
59:04it's an utterly accurate
59:08simulation of them.
59:09They know everything and can extrapolate
59:11based on what they've said, what they're
59:12likely to say,
59:14and says, you know, wire me a million
59:16bucks. That's defensive. What about What
59:18will you still be able to do that it
59:19won't be able to do, I guess? On the
59:21analytical side. So it's a good
59:22question. I did just have I just watched
59:24The Last Samurai and I asked um at my
59:27firm to watch it. And The Last Samurai,
59:29if you haven't seen it, I highly
59:30recommend watching it. It's actually a
59:32movie that's aged really well. Tom
59:33Cruise movie from 20 years ago, you
59:35know, the conceit is Tom Cruise is this
59:37like bitter, washed-up Civil War veteran
59:39who's actually a very good soldier. He's
59:41bitter and washed-up cuz he feels like
59:43he participated in negative actions
59:45against the Native Americans. He's hired
59:47by Japan to train It's during the Meiji
59:50Restoration. And he's hired by the
59:52modern elements of the Japanese
59:54government to train like an army of
59:56peasants
59:57how to fight the samurai. There's a
59:59first battle, of course the samurai win
1:00:01even though they don't have guns.
1:00:02He fights valiantly, so the samurai
1:00:04decide not to kill him, take him to
1:00:06their village. He becomes a samurai. It
1:00:07feels like the Civil War to him. So he
1:00:09fights on the side of the samurai.
1:00:12And at the end, he's massacred by a
1:00:14peasant with a machine gun.
1:00:16And like the machine gun is here.
1:00:18And if we do not all
1:00:21become masters of the machine gun, we're
1:00:23going to get mastered. So I am trying to
1:00:25become a master of the machine gun. And
1:00:27then,
1:00:28you know, I'm optimistic there's a long
1:00:31period of time where just like if you
1:00:33were a 50-year-old samurai veteran of
1:00:36many wars, I fought many wars, Master
1:00:39Dwarf.
1:00:40Um you will have advantages using the
1:00:42machine gun. And I'm optimistic as a
1:00:44lifelong student of investing, I'm going
1:00:47to be able to master the machine gun,
1:00:49this new technology, um integrate it
1:00:51into my own process, integrate it into
1:00:53our firm's process
1:00:55in ways that, you know, let me
1:00:57contribute value as a human being for a
1:00:59long time. But, you know, like everyone,
1:01:01like, you know, I have agents running
1:01:03all the time now.
1:01:04>> What's your most useful agent?
1:01:05>> useful agent, honestly, is as And I
1:01:07think I told you this, and I don't want
1:01:09to hurt your business, but my single
1:01:12most useful agent is a really good
1:01:15summary of the points that would be
1:01:17interesting to me from podcasts. There's
1:01:20like 6 hours a day of stuff that I feel
1:01:23like it's in my job description to
1:01:24watch. You know, every time every time
1:01:27somebody from OpenAI, xAI,
1:01:30Google,
1:01:32you know, Cursor,
1:01:34Fireworks, Space 10, let's say nothing
1:01:37of like Jensen, Elon, Dario.
1:01:40Um,
1:01:41I feel compelled to watch and I just
1:01:43don't have that much time. And there's
1:01:46some real needles in haystacks. There's
1:01:48a set of things I always like to see
1:01:49like I'm very sensitive to management
1:01:51compensation. What are they incentivized
1:01:53to do? They do they have stupid RSUs? Or
1:01:56do they have PSUs? And if they have
1:01:57PSUs, what are those PSUs incentivized
1:01:59to do? I think systems that do a very
1:02:01good first pass at that.
1:02:03And you know, that saves people a lot of
1:02:06time. It frees them up for more creative
1:02:08work than like, you know, going through
1:02:10the proxy, pulling the PSU thing,
1:02:14looking at how it's changed versus all
1:02:16the proxies cuz there's signal in that.
1:02:18And that's very labor intensive and
1:02:20that's so good for an AI. And there's
1:02:21obviously all sorts of same things
1:02:23within investing. This is the most
1:02:24exciting, thrilling time to be an
1:02:26investor.
1:02:28And there is and it is I am a little I'm
1:02:30getting a little bit worried.
1:02:32>> The diversity breakdown thing?
1:02:33>> Yeah. I'm getting Say just like a little
1:02:35bit more about like the kinds of people
1:02:37that are
1:02:37>> know of anyone like me who's not really
1:02:39bullish
1:02:41on DRAM. No one. No one. There's all
1:02:43these interesting things happening with
1:02:44AI right now.
1:02:46So, one is cross-sectionally the
1:02:48valuations do not make sense.
1:02:50They just flat out do not make sense.
1:02:52They cannot all be true. You have
1:02:54semi-cap equipment companies trading at
1:02:5640 times next quarter's annualized
1:02:58earnings and DRAM companies trading at
1:03:00mid single digit. At the peak of the
1:03:02last cycle, that was like five versus
1:03:0412. At one point it was like three
1:03:06versus 45. Those can't both be true. And
1:03:10yes,
1:03:11semiconductor capex business models have
1:03:13improved more than the memory business
1:03:14models. We don't know how much HBM
1:03:17is going to improve memory business
1:03:19models yet. Yes, they have some element
1:03:21of recurring revenue with parts and
1:03:23maintenance,
1:03:24but it's not worth a thousand percent
1:03:26multiple gap. I think it's hard to
1:03:27square like the valuation of something
1:03:29like Nvidia, which is still, you know,
1:03:31in in in early April was essentially as
1:03:34cheap as it gets relative to the market
1:03:36like in the last 10 or 12 years or
1:03:37whatever it is, and very cheap absolute.
1:03:40It's very hard to square that valuation
1:03:42with something like GE Vernova's
1:03:44valuation.
1:03:46Cuz it builds in like
1:03:48un- unfathomable amount of share loss
1:03:50for Nvidia. So, valuations
1:03:52cross-sectionally are really different.
1:03:54Because we are in shortages,
1:03:58the lowest quality companies are doing
1:04:00the best.
1:04:01So, if you're an oil and gas investor
1:04:04or, you know, a mining investor, natural
1:04:05resources
1:04:07investor, and you're, you know, you're
1:04:08well-versed in thinking of costs, this
1:04:10is very intuitive to you. In a real bull
1:04:12market for a commodity, the commodity
1:04:14suppliers with the highest cost go up
1:04:17the most because it's the most
1:04:19beneficial to them. They go from on the
1:04:21verge of bankruptcy to just gushing
1:04:22cash.
1:04:23And this is, I think, one reason
1:04:25commodity investing is really, really
1:04:26hard because quality outperforms during
1:04:29the cycles, but you get all of the
1:04:31outperformance during the downturns when
1:04:33the high cost guys that mooned during
1:04:36the shortages and the commodity bull
1:04:37markets, you know, go bankrupt or
1:04:38whatever. You're seeing that happen in
1:04:40every industry.
1:04:41The lowest quality players in, you know,
1:04:44these different industries that are
1:04:46hated and detested
1:04:49by the hyperscalers and the buyers cuz
1:04:51they have high costs, they're
1:04:52unreliable, the parts fail at a high
1:04:54rate, etc., etc. They're sold out and
1:04:57raising prices. Um
1:04:59And then that activity gets the interest
1:05:01of like these retail accounts on X, and
1:05:04these stocks get bid to the moon.
1:05:07Whereas some of the higher quality
1:05:08expressions
1:05:10have like actually really
1:05:11underperformed.
1:05:13And, you know, as an investor, it's it's
1:05:15hard because you know
1:05:17within a like
1:05:20a shadow of a doubt
1:05:22that that thing that's moved, you know,
1:05:2410x
1:05:25in 3 months or 6 months
1:05:27is going to go right back down subject
1:05:30to what they do with all the cash. But
1:05:32like these little quality companies
1:05:33really do smart stuff with cash. And so
1:05:35it worries me a little bit that people
1:05:37who were very skeptical a year ago are
1:05:39no longer skeptical. But then I just
1:05:41contrast that with like the valuations
1:05:44of these like high-quality companies,
1:05:47which are just not extended, and it
1:05:49makes me feel better. But it does kind
1:05:51of feel like, you know, I just thought
1:05:52it was funny in '24 and '25 that anyone
1:05:55asked about an AI bubble or talked about
1:05:57it. Cuz it's like you have this nuclear
1:05:58bubble and this quantum bubble right
1:06:00here, right in front of you. What are we
1:06:02talking about? This is so real.
1:06:04Some of that nuclear quantum silliness
1:06:07has maybe spread into more speculative,
1:06:10lower-quality, smaller-cap names
1:06:14where if you have a big presence on X or
1:06:16Reddit, it's easy to move them. And that
1:06:19frightens me a little bit. But I just
1:06:21wish there were more AI bears, like I
1:06:23wish there were more memory bears. You
1:06:25know, one reason I'm
1:06:27you know, Astera is a stock I've been
1:06:28close to a long time.
1:06:30There's a lot of bears on that. I love
1:06:32that. Great, you know, I first invested
1:06:35in the series C. Good luck thinking
1:06:37you're going to price that you know,
1:06:39differentially from me. You know, good
1:06:40luck thinking that's a copper loser. And
1:06:42then there's also you can feel the
1:06:45baskets in the market in the leverage
1:06:46baskets. And what baskets you're in is
1:06:49really important, you know, copper,
1:06:51optical, DRAM, NAND. Um and a very
1:06:54interesting thing that's happened this
1:06:56year um is in '24 and '25 the AI trade
1:06:59traded together.
1:07:01So like you could be long GPU compute,
1:07:05scale-up networking, and optical scale
1:07:07across
1:07:08and like short power that trade worked
1:07:11from like a risk management sense cuz
1:07:13you know I'm very factor aware.
1:07:15That all blew out in Jan- January of
1:07:17this year.
1:07:18It's like you know scale up networking
1:07:21would go crazy while scale out was going
1:07:23down or DRAMs massively underperforming
1:07:26NAND and HDDs which had not happened.
1:07:29So these cross-sectional correlations
1:07:32within AI
1:07:33really fell apart and you had to get
1:07:36very fine-grained. You couldn't hedge
1:07:39your memory
1:07:40anymore with like some semi-cap
1:07:43equipment or NAND. Everything
1:07:46cross-sectionally
1:07:47really changed and in a very interesting
1:07:50way in January.
1:07:52And I think maybe one reason for that
1:07:53was you know the AI got to a quality
1:07:57where it was all of a sudden really easy
1:07:59for a bunch of people to get really
1:08:00smart on these different subsectors,
1:08:03start trading them, and then they get
1:08:05put into baskets and those baskets in
1:08:07the
1:08:07>> Yeah, creating price efficiency. Yeah.
1:08:09Yeah, exactly. And then it's like if you
1:08:11like I think some of the biggest
1:08:13opportunities outside of these higher
1:08:14quality names that I think can compound
1:08:16for a long time
1:08:18and they're safe unlike these low
1:08:19quality names which are terrifying is in
1:08:21names that are miscategorized.
1:08:24Like Astera was in a lot of copper loser
1:08:26baskets.
1:08:28Astera their biggest product is going to
1:08:30be a switch. You use both copper and
1:08:32optics to connect switches to
1:08:35accelerators.
1:08:37>> [laughter]
1:08:37>> And so definitionally
1:08:39if you're a switch company or an
1:08:41accelerator company, you cannot be a
1:08:43copper loser because you're going to be
1:08:45on the other side of that connection. I
1:08:47I wonder if you could riff just for like
1:08:49a sentence or two on each of the major
1:08:50companies. I feel like I always forget
1:08:52to ask you like Google, Microsoft,
1:08:53Amazon, you know, the the the major
1:08:55players that are public that all the
1:08:57conversation is centered around these
1:08:58exciting new companies.
1:09:00>> Yeah. So Google uh it was incredible
1:09:02last year because they had that TPU
1:09:04advantage which is now gone. The reason
1:09:05I think they're still in a great
1:09:06position is just they have the most
1:09:08compute of everyone. We talked about the
1:09:10value of installed bases being higher as
1:09:13a result of shortages.
1:09:15They have the biggest installed base of
1:09:16compute.
1:09:18I am a little surprised
1:09:21by
1:09:24their inability and Google IO is this
1:09:28is this week.
1:09:30And
1:09:31um
1:09:32like if they don't release something
1:09:35that even slightly leapfrogs
1:09:39OpenAI
1:09:40and or Claude
1:09:43like that that's interesting and it's
1:09:45not a disaster for Google. It's just
1:09:48interesting and it just means this
1:09:49Nvidia effect we discussed is even more
1:09:51powerful than maybe I'd imagined but I'm
1:09:53very curious to see what the Pareto
1:09:55frontier looks like literally in five
1:09:58days after Google's announced its new
1:10:00stuff. This is a big card for them but
1:10:02Google you know between the amount of
1:10:05data they have and the YouTube data is
1:10:07actually really genuinely valuable. It's
1:10:09actually
1:10:10it is valuable in a world of robotics.
1:10:12The amount of compute they have and you
1:10:15know the search business they have.
1:10:17Google's never not going to be in a good
1:10:18position and then you see that with GCP
1:10:20going crazy. You got to give Zuckerberg
1:10:23a immense credit.
1:10:24What he's done in terms of making meta
1:10:26an AI first company internally
1:10:29and I do think he is the only one of
1:10:31those true internet giants to have done
1:10:33that.
1:10:35And I give him a lot of credit for that.
1:10:37I give him a lot of credit for paying up
1:10:41when he did for you know all those you
1:10:43know those billion dollar contracts that
1:10:45talent.
1:10:46And news I think it was a really big
1:10:48upside surprise.
1:10:50You know it was the first model from MSL
1:10:54and it's not on the Pareto of frontier
1:10:57with you know XAI Google's one entrant
1:11:00and then open AI and Claude but it's
1:11:02pretty close. That was very impressive
1:11:04to me. So I think that is in a
1:11:07better position. Still not as strong of
1:11:09an absolute position as Google but like
1:11:11they're better position and rates of
1:11:13change matter more than level as you
1:11:15know in markets particularly over short
1:11:17like three year time frames over like
1:11:19long time frames level of competitive
1:11:21advantages tends to dominate but even
1:11:23within that you know the changes changes
1:11:25are really matter.
1:11:27Amazon I think is in a really strong
1:11:29position because of tranium. You're
1:11:30going to see like real P&L efficiencies
1:11:33from robotics over the next 18 months in
1:11:35their retail business. I actually think
1:11:37Nova their internal models are not where
1:11:40Muse is but they're better than they get
1:11:42credit for. Microsoft I think Satya is a
1:11:45really brilliant man but you know in in
1:11:48investor
1:11:49conversations people just don't talk
1:11:52about him the way that they did. I I
1:11:53like Satya. I admire him. I think he's
1:11:55an exceptional CEO.
1:11:59And I give him a lot of
1:12:01credit for the decisions he's made but
1:12:03you know he did go from we're going to
1:12:04make Google dance to being the product
1:12:07manager of co-pilot
1:12:08in like three years. I I would love to
1:12:11know during the coup attempt against
1:12:12open AI
1:12:14does Satya regret his decisions?
1:12:17Does Satya wish that he had supported
1:12:20Ilya
1:12:21and instead of Sam and that kind of Ilya
1:12:25and Mira were really running open AI
1:12:28today. In his heart of hearts I would
1:12:30love to know.
1:12:32Cuz I think the Microsoft open AI
1:12:33partnership might look very different
1:12:37in that world. I think that's a very
1:12:39interesting question that we'll never
1:12:40know the answer to.
1:12:43But I give him a lot of credit like he
1:12:45is what he is doing now
1:12:48he's taking risk.
1:12:50So they could earn you know this goes to
1:12:52the decisions you have to make in that
1:12:53cone of uncertainty are not only
1:12:55how much you spend,
1:12:57but what you're going to spend it on.
1:12:59I think Microsoft flinched
1:13:03for like a moment in early 25. You know,
1:13:06they have this algorithm, we spend this
1:13:08much CapEx dollars, we get this return.
1:13:10That algorithm was kind of off.
1:13:13And if you flinch, you lose position.
1:13:15You lose all these allocations, and it's
1:13:17difficult to get it back. So, they
1:13:19flinched.
1:13:20And now the decision Satya is making,
1:13:22which the market has punished him for,
1:13:23but I think is the right decision,
1:13:26is we're going to use our compute,
1:13:28rather than making, I mean, who knows
1:13:30how fast Azure could be growing if
1:13:32they're willing to just sell GPUs to
1:13:34OpenAI.
1:13:35We're going to use our compute
1:13:37internally to make our own products
1:13:39better. You know, one reason Copilot is
1:13:41so bad, or has been so bad, is just one
1:13:43enough compute available. They're fixing
1:13:44that.
1:13:46He's the product manager of Copilot. I
1:13:47do think he's a great CEO.
1:13:50And they're trying to use their compute
1:13:52to train their own models.
1:13:54I don't I am a little skeptical that
1:13:56they have the right team to succeed
1:13:57there, but, you know, they can
1:13:59certainly, like, just like Meta, they
1:14:01can afford
1:14:03to hire maybe a maybe a different team.
1:14:06But I think he's making good decisions
1:14:08that are risky decisions
1:14:11to position Microsoft from for this
1:14:13world where frontier models are are no
1:14:16longer API accessible.
1:14:19And I think it's a really courageous
1:14:20decision that I give him a lot of credit
1:14:22for, and he is forgoing, I mean,
1:14:24Microsoft probably be an $800 stock
1:14:26today if they were using their GPUs to
1:14:28serve OpenAI
1:14:30solely OpenAI and Anthropic's capacity
1:14:32instead of using them for their own
1:14:34products. So, I give him a lot of credit
1:14:36for making a great decision. What's
1:14:38really interesting
1:14:40is the degree to which these companies
1:14:42are outward-facing
1:14:44in their decisions. The two companies
1:14:46who are the most deeply engaged with
1:14:48startups are Amazon and Nvidia by a
1:14:51mile. Then there's a really intense
1:14:55engagement with Google. They're next
1:14:57most intense.
1:14:58Broadcom is engaged in a different way.
1:15:01They're just, you know, everybody's
1:15:03favorite ASIC supplier. Like it's, you
1:15:05know, if you're a startup, it's
1:15:06considered like a level up if you get to
1:15:08work with Broadcom for your second gen
1:15:10chip. And it's considered mana from
1:15:11heaven if Broadcom works with you for
1:15:13their first gen chip. And then you see
1:15:15essentially zero engagement with
1:15:19startups
1:15:20from AMD, Microsoft, and Meta. And I
1:15:23just, yeah, I mean, when I say zero,
1:15:25it's a little.
1:15:27And I just wonder about that decision.
1:15:30Because some of the best teams
1:15:35are no longer at big public companies.
1:15:37They're at these smaller startups.
1:15:39And I think it's going to end up being a
1:15:41pretty big advantage for Nvidia, AMD,
1:15:44Google right behind them to have this
1:15:47engagement
1:15:49that you just don't see from these other
1:15:53um hyperscalers. As we wrap up, I'm
1:15:54curious for you to riff on any other
1:15:56like out there knock-on effects that
1:15:58you've started to think about for this
1:16:00giant trend. We've talked about the
1:16:01specific companies in a lot of detail
1:16:03that this most impacts. We talked a
1:16:05little bit about the application layer
1:16:06and what would have to happen for there
1:16:08to be more value accruing to that layer
1:16:09of the stack. I'm curious like any other
1:16:11just fun knock-on things that you've
1:16:13been thinking about as this world
1:16:15changes so quickly.
1:16:15>> Yes, and it is wild. I mean, at the
1:16:17application layer, forget value
1:16:18accruing, just value has been destroyed.
1:16:20AI has net destroyed. Even if you count
1:16:22Cursor Cognition, the most successful AI
1:16:25natives, value has been
1:16:27trillions of dollars of value has been
1:16:29destroyed by AI at the application
1:16:31layer. And just in this context, I do
1:16:33think it's a little it's something we
1:16:35need to be aware of. The companies that
1:16:36are doing the best today that are are
1:16:41kind of their values increase the most
1:16:42that are creating economic value are the
1:16:45companies with the highest ratio a
1:16:47highest effective ratio of utilized GPUs
1:16:51per human.
1:16:52And you know, maybe this just means that
1:16:54every human's going to get a lot of
1:16:55GPUs. But I think that's an interesting
1:16:57fact that we kind of need to be
1:16:59cognizant of. I will just say and maybe
1:17:01this is a little dark. I am more more
1:17:03more and more worried about personal
1:17:05safety. And I worry about this a lot
1:17:07more for people who are you know, have a
1:17:10much bigger public presence and are much
1:17:12more associated with AI. But I really
1:17:15worry about personal safety. I hope
1:17:16nothing tragic happens, but like there
1:17:18is this upsurge in political violence
1:17:21here in America. And as AI increasingly
1:17:24becomes political, I worry that's going
1:17:26to get directed at more and more AI
1:17:28political leaders, you know, just
1:17:29whatever we can agree you know, whatever
1:17:31whatever I may think or may not think of
1:17:33open AI. Like I think it is terrible
1:17:35that someone threw Molotov cocktails
1:17:38at Sam Altman's house. I am worried that
1:17:41we are headed into a higher variance
1:17:45higher beta
1:17:48higher risk world because of AI. And
1:17:51that's for me as an individual and then
1:17:53you know, for people who are big players
1:17:55on the chessboard. Think about what it
1:17:57means geopolitically. Like we're
1:17:59watching the Ukrainians are really
1:18:01starting to win.
1:18:02And the reason they're winning I I think
1:18:04is not really because they have better
1:18:05drones. I think they do have better
1:18:07drones. That's part of it. I think the
1:18:08reason Ukraine is really winning is they
1:18:10have the best battlefield AI.
1:18:12Outside of probably America and Israel.
1:18:15And has China has our adversaries begin
1:18:20to process that
1:18:22like how do they respond? Like if the
1:18:24United States because of its edge in AI
1:18:28um it's great if you're America.
1:18:31But it is destabilizing for the rest of
1:18:34the world. Something I think a lot about
1:18:35is creating a charity to just like
1:18:37educate the world on how awesome the
1:18:39West has been. Slavery was endemic to
1:18:41essentially almost every civilization
1:18:43and slavery was really ended by the
1:18:44British Empire. Tell that story. Um
1:18:48but America after 1945
1:18:51we had the nuclear bomb, no one else had
1:18:53it.
1:18:54We could have controlled the world
1:18:56forever.
1:18:57Instead we rebuilt Germany and Japan
1:19:00and now who are America's most reliable
1:19:03allies? Israel, South Korea, Japan.
1:19:05That's a testament to like the American
1:19:07spirit in our country. We didn't take
1:19:08over the the world. You know, there were
1:19:10these fears, you know, that were
1:19:11documented at the time that the American
1:19:13generals
1:19:14and you know, MacArthur was a little bit
1:19:16of an American emperor in Japan
1:19:19but um we're just going to take over the
1:19:20world and they could have and they
1:19:22didn't. They came home, we demilitarized
1:19:26and then you had this, you know this
1:19:28period of of great global stability
1:19:30between, you know, a scary there were
1:19:31terrible wars. Yeah, you had the Pax
1:19:33Americana.
1:19:34So maybe it's not destabilizing. Maybe
1:19:36it leads to the another Pax Ameri-
1:19:39Americana
1:19:40informed by our AI dominance and I'm so
1:19:43optimistic that AI is going to be
1:19:45amazing for the world. There's someone
1:19:47like me whose daughter was diagnosed
1:19:49with a very rare mutation.
1:19:51There's no cure.
1:19:53He was able to assemble a lot of
1:19:54resources. He was able to get a lot of
1:19:56compute from the labs. Um we were made
1:19:58aware of what was happening.
1:20:01Spun up a immense amount of agents, came
1:20:03up using AI with a drug on the market
1:20:07that can actually impact his daughter's
1:20:08disease
1:20:09and then has spun up a company to cure
1:20:12it.
1:20:13And like her life is already
1:20:16immeasurably different because of AI. So
1:20:18I'm like an AI I'm like an AI optimist
1:20:21maximalist, but I also just acknowledge
1:20:23it's like an event horizon.
1:20:25It for sure I think it's going to be a
1:20:27discontinuity. We need to navigate as
1:20:29soci- as society. I think the Luddites
1:20:31are going to be wrong, but we need to be
1:20:33like really thoughtful in how we address
1:20:36their concerns. We need to make sure
1:20:38that it's good for everyone. Like it is
1:20:40a little dystopian that now the best AI
1:20:42is only available to people with a lot
1:20:44of money. Like we need to solve that. We
1:20:46need to approach this with humility,
1:20:48recognize there's a lot of uncertainty,
1:20:49and be thoughtful. When I do this with
1:20:51you, I tell people afterwards, I'm like,
1:20:52"May you find something that you love as
1:20:54much as Gavin loves markets and
1:20:56companies and capitalism and history."
1:20:59Uh on display today as always, Gavin,
1:21:01thanks so much for your time. [music]
1:21:02Thank you. Thanks, Patrick.
1:21:07You know how small [music] advantages
1:21:08compound over time? That's true in
1:21:09investing and just as true in how you
1:21:11run your company. Your spending system
1:21:13is your capital allocation strategy.
1:21:15Ramp makes it smarter by default. Better
1:21:17data, better decisions, better economics
1:21:18over time. See how [music] at
1:21:20ramp.com/invest.
1:21:22As your business grows, Vanta scales
1:21:23with you, automating compliance and
1:21:25giving you a single [music] source of
1:21:26truth for security and risk. Learn more
1:21:28at vanta.com/invest.
1:21:30Every investment firm is unique, and
1:21:32generic AI doesn't understand your
1:21:33process. [music] Rogo does. It's an AI
1:21:35platform built specifically for Wall
1:21:37Street, connected to your data,
1:21:38understanding your process, [music] and
1:21:40producing real outputs. Check them out
1:21:41at rogo.ai/invest.
1:21:43The best AI and software companies from
1:21:45OpenAI to Cursor to Perplexity use
1:21:47WorkOS to become enterprise ready
1:21:49overnight, not in months. Visit
1:21:51workos.com to skip the unglamorous
1:21:53infrastructure work and focus on your
1:21:55[music] product. Ridgeline is redefining
1:21:57asset management technology as a true
1:21:58partner, not just a software vendor.
1:22:00They've helped firms 5x in scale,
1:22:02enabling faster growth, smarter
1:22:04operations, and a competitive edge.
1:22:06Visit ridgelineapps.com [music] to see
1:22:08what they can unlock for your firm.

About this transcript

This page contains the full transcript of Watts, Wafers, and the Future of AI Infra | Gavin Baker by Invest Like The Best, generated from the public captions YouTube serves with the video. The transcript has 14,375 words across 2,357 segments, with the original timestamps preserved so you can click any line to jump to that moment in the embedded player.

What you can do with it

Use the transcript to take notes, quote the speaker, build a study guide, generate a summary with ChatGPT or Claude via the YouTube Summary tool, or export it as a timed subtitle file with YouTube to SRT. You can also re-open it in the transcriber to translate the transcript into 100+ languages.

Free YouTube transcript tool

YouTube2Text is a free YouTube transcript generator — no signup, no daily limit. Paste any YouTube link and get the full transcript instantly, with timestamps, click-to-jump, translation to 100+ languages, AI prompts for ChatGPT, Claude, and Gemini, and exports to TXT, SRT, VTT, or Markdown.