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

What was happening in AI was I think the most extraordinary moment in the history of capitalism, the history of American business. Anthropic added 11 billion of ARR. The three highest-profile SaaS companies founded in the last 10-12 years are Palantir, Snowflake, and Databricks. And these three companies spent 10 years building their businesses. Anthropic added their combined businesses in one month. That's just nothing like that has ever happened in the history of capitalism. Forget my career. Just the flat-out history of capitalism, the history of business.

I would say broadly speaking there are two kinds of drawdowns. There are drawdowns where you're wrong, a company misestimates, your hypothesis was invalidated and you have to take your medicine and you crystallize that loss. And then there are drawdowns or periods of underperformance where you're underperforming because of companies you know really, really well and where you profoundly disagree with the price action and you can lean in and instead of crystallizing negative performance you can kind of build pent-up alpha, pent-up future performance. And for me that is what March felt like. It felt like the NASDAQ was selling off and at the same time what was happening in AI was I think the most extraordinary moment in the history of capitalism, the history of American business. And what I just mean by that is that Anthropic added 11 billion of ARR. And what is astonishing to me about this is that the SaaS and cloud revolution, it created we'll call it between 5 and 10 trillion dollars of value. And I would say arguably the three highest-profile SaaS companies to have kind of been founded in the last 10-12 years are Palantir, Snowflake, and Databricks. And these three companies have spent, employ thousands of people, tens of thousands collectively. They've all spent 10 years building their businesses. And Anthropic added their combined businesses in one month. That's just nothing like that has ever happened in the history of capitalism. Forget my career. Just the flat-out history of capitalism, the history of business. It's wild. And then Krishna comes on this show and shares some stats. 500% in DRAM.

So there's just no precedent for this. And we, you know, tech investors, you hear a lot of discussions about S-curves and investing in exponentials. I've just never seen an exponential like this. It felt even more extreme than DeepSeek, which was a very similar setup. If we go back to '25 and there's a huge sell-off on DeepSeek, which was very strange because the paper gets published 7 days before DeepSeek Monday. It got published, I believe, on a Monday that was a holiday in America. And I read it, I thought, hm, you know, this feels like it might not read that positively for the AI trade. I took action. We had DeepSeek Monday where AI really imploded a week later and that was really strange because by DeepSeek Monday it was super clear that this was going to be the most positive thing that had ever happened to compute demand. Prices in the AWS availability zones in Asia had already like doubled. You were seeing GPU availability go down. And this was just the first time we saw how much more compute-hungry reasoning models are during inference than non-reasoning models. And so that was a similar setup, but you had to do some work to see that. I mean, it's not that hard to say, oh wow, stocks are selling off. The price of DRAM is going vertical. The price of GPUs in Asia going vertical. GPU availability is going down. And then like two or three days later, you know, GPU prices in America started going up, GPU rental prices. All you had to do in March was just simply observe what was happening to Anthropic. There's all these people who seem to regret, you know, not buying during '22, not buying during COVID, not buying during DeepSeek. You had the same valuation setup at the beginning of April and an even clearer AI inflection. And so there have been all these chances to buy into AI. And then of course what complicated it was the Strait of Hormuz. I became a believer and am a believer that I think maybe one thing that the market was mispricing, and I'm no macro expert, I do do a lot of pro-national security investing, and so I do have access to people who are experts and are excited to share their thoughts and opinions with me, that the Strait of Hormuz being closed is actually relatively awesome for America.

Because particularly for the goals of the current administration. So electricity is a very important industrial or manufacturing input. The key input into American electricity prices which feeds into AI is Henry Hub natural gas on Bloomberg that was down 20%. And natural gas in Asia, Europe, everywhere else doubled or tripled. So our relative manufacturing competitiveness improved overnight. And for better or worse that is what the Trump administration seems to care about. They are very focused on America's relative position. And I think a lot of people had memories of the 1970s. And what made the 70s so traumatic was it wasn't just that prices went up, it's that there were actual gas shortages. And then you go through, okay, well the US economy is dramatically less energy intensive than it was. The United States is now the world's largest producer of oil and gas and we've become now the world's largest exporter of oil and gas. And then on top of that there's this relative manufacturing advantage. And so that made it I think easier to stay focused on AI fundamentals, stay focused on what were historically attractive valuations. I think on a relative basis tech essentially got as cheap as it's been versus the rest of the market has at any point over the last 10 years. And just think about that in the context of market efficiency. We have the most extraordinary moment in the history of capitalism that's wildly bullish for AI and you get a chance to buy AI at really attractive valuation.

I do think OpenAI and Anthropic are pretty different animals from a capital efficiency perspective. And Anthropic clearly has a dramatically lower cost per token than OpenAI. They just do. And you can just see that in the amount of money that they have burned to get to a roughly similar revenue scale. I think have they burned maybe 80% less than OpenAI. So as businesses, they clearly have very different structural ROICs. I think OpenAI is doing a lot. I think Sarah Friar is one of the most exceptional CFOs. I think they're doing a lot of things to try to improve this. And they've secured a lot of compute. That's another big difference. It turns out being aggressive really paid. But yeah, just Anthropic at 900 billion for 50 billion ARR and growing at ridiculous rates. Maybe a true statement is that if Anthropic had all the compute, they'd probably be doing well north of 100 billion dollars today, maybe 150. And I do, you know, they have clearly deprecated the intelligence of Claude. There's an analysis Claude even on Opus is generating 70% less tokens for the exact same question. And you know, as we talked about last time, token quantity equals quality of answer and quality of thinking at some level. And there is an intelligence density per token that also matters. I think I felt that as a user. So I think they would be doing materially more, 100, 150, maybe 200 billion. So you might be buying it at more like five times unconstrained, I'm going to make up a new word, unconstrained run rate revenue.

So I think it's wise. The future is uncertain. You are clearly in a very capital-intensive game even if you are Anthropic. I'm sure is at very positive gross margins on inference today. I think probably starts generating cash this year if they are not already generating cash, which I think is probably the case. But still you probably want to be able to raise more capital, access more compute. The world is uncertain. Ukraine is starting to really, really win. How is Russia going to respond? And you know, I think there's still a lot of uncertainty in Iran. All this uncertainty, I think, probably amplifies geopolitical uncertainty over Taiwan. So, it's an uncertain world. If I think about Elon, Elon has always made investors money. He treats it like a sacred covenant. And as a result, because he's made people money for now 20 years, he has a superpower. And that is he can essentially raise as much capital as he wants, whenever he wants. And I think it's wise that these companies are taking, I don't know if that's how they think about it, but I do think being focused on making investors money is wise and creates benefits that don't just last for like a year or two. They can last for the next 20 to 30 years.

Just never being greedy on valuation, never pushing valuation.

You know, my friend Antonio pointed out SpaceX compounded, you know, low 30% per year for whatever that was a decade. And that was just because Elon was, I think, focused on preserving the superpower and having trying to strike a fair balance between investors and employees. But I think it's wise. But could Anthropic raise money at probably at least a 100% premium to this rumored latest mark? Of course.

Yeah, I would say I think capitalism is going to solve the watts shortage absent big regulatory political blowback, which I think is a real possibility. The head of kind of data center infrastructure investing at one of the big PE firms, you know, I think Blackstone, Apollo, KKR said it used to be energy and chips were our biggest gating factors. Now it's zoning and approval, much more important. And I think a lot of companies are waiting till after the midterms to take action in terms of maybe workforce reductions. Nobody wants to be, you know, piñata during the midterms. But you know, you've seen a lot of companies that make turbines announce plans to significantly increase capacity. There's like two of these machines that can cast these big blades. We haven't made one in 80 years in the West. We don't know how to make them anymore, etc., etc., etc. All of that is true. And by no means am I minimizing, you know, the industrial engineering magic and artistry that goes into those, but capitalism is very good at solving problems like these over time. There's other sources of energy besides these turbines with a longer time frame. So I think the watts shortage will probably begin to alleviate '27, '28. And then I think orbital compute will really solve that. And I do want to like reframe orbital compute because I think when people hear data centers in space, which we discussed on our last episode, they picture a Pentagon-sized building in space. They're like, well we can't do that. That's not what it is. A Blackwell rack weighs 3,000 lbs. It's 8 ft high. It's 4 feet deep. 3 feet wide. It's racks in space. And SpaceX has showed you an illustration and it's a rack. That's the satellite. But it's probably about the size of a Blackwell rack. It has these solar wings that are probably 500 ft long on each side. You keep it in a sun-synchronous orbit. So those solar panels are always in the sun. And then because it's in an exactly sun-synchronous orbit, the radiator which extends behind it for hundreds of feet.

I've spent a lot of time at Starbase over the years and I've talked to a lot of SpaceX engineers and I do think it is the most talented group of engineers on planet Earth and they're very confident they have solved this. And they're not always confident. Like I think probably, you know, there's some engineering that needs to happen to turn the Starship into a Mars colonial transporter. Will they do that? Absolutely. What are they more focused on? I would say probably, you know, the repair and maintenance. These are the two big, you know, the two big responses, the radiator and how do you repair whatever issue goes wrong in the rack. And the answer is like until you have probably an, you know, floating Optimus, you don't. Now, I do think Starship is going to change the space economy in ways we cannot imagine. Particularly if regulation becomes a constraint to data centers, none of it's going to matter. You're going to sell as much orbital compute as you can make. And then obviously you link these racks using lasers traveling through vacuum which are already on every Starlink. And it's just mind-blowing to me that SpaceX operates the world's largest satellite fleet which is like 98 or 99% of all satellites in orbit. Every Starlink they're cooling it today. And you know, I think Starlink V3 is going to operate at 20 kW. A Blackwell rack is only 100 kW. And people talk a lot about density. Well, if you're connecting the racks with lasers through vacuum, you know, you can make the rack bigger physically. You're focused on weight, not size. In a data center on Earth where you're trying to connect racks, ideally using copper, minimize lengths, etc., etc. Cabling is a big cost. You do want that rack to be small because you know copper when you can, optics when you must. But in space, you know, there's all sorts of things that SpaceX can do that I think maybe some of these naysayers are not contemplating, but it's just they operate more satellites than you want. They have a 20 kW satellite today, so maybe you just scale that up to 60 kilowatts to start. They seem very confident they're going to go right to 100 to 120. And they also the same company now also operates the largest data center on earth. They have the world's best hardware engineers and all sorts of people almost all of whom are not smart enough or practical enough to work at SpaceX are these armchair skeptics. You know, I don't want to quote Larry Ellison, but somebody was, you know, being skeptical and Larry was just like, 'Listen, he's out there landing rockets. I don't see anybody else landing rockets.' And the reality is that 10 years later, no other company is consistently capable of landing and fully reusing an orbital rocket. And none of this works makes sense without reusability. That means you have to land it. I would like to redefine orbital compute as racks in space, not giant floating Pentagon data centers in space, which is just, you know, that's silly. But you can, you know, what makes a data center is you're connecting these racks with lasers. So it'll be racks in space that are connected with lasers into a virtual data center. And if you think about that state of the world, let's say that all happens and we're really good at getting these things up economically, running matrix multiplication all over space. What does that mean for terrestrial data centers? Someone once said, you know, America was going to suck as hard as it can on every energy source it can get. And I just think the same is true of compute.

We're going to consume as much compute as we can. And inference I think is very sensible for orbital compute. Training will be done on Earth for a long time. So I don't think that this is super bearish for terrestrial data centers. I think those are going to be valuable for my lifetime. But I do think if you are in this ecosystem of power production and cooling and you are massively ramping capacity and you know a lot of these capacity ramps are going to be hitting just as I think you know all of the silly skeptics start to understand that orbital compute is very real, like I think it's worth thinking long and hard about that if you're one of those companies. And then all sorts of cool stuff is happening in the interim. You know, we're getting really good at repurposing jet engines. You know, there's that Boom Aerospace that is doing this. So, there's a lot like capitalism is hard at work on watts. On wafers, though, it's just this group of, you know, flinty older humans in Taiwan who are the most important humans in Taiwan. They are the overwhelming fraction of the country's GDP, water usage, electricity usage. They talk about the Silicon Shield. They all view themselves as inheritors of, you know, Morris Chang's sacred legacy. I vividly remember like visiting Science Park more than 20 years ago and, you know, talking to them. Do you think you could catch Intel? And they said, 'This is such a beautiful dream, but it's a dream for our grandchildren.' And they did it. Partly because of Intel's self-inflicted wounds, but just they think very differently. You know, one reason, you know, Jensen flies over there so much is he wants them to expand capacity. I do think it's wild that Jensen has never had a contract with Taiwan Semi. They do business on what seems fair in handshakes. Just fascinating. No contract. It's going to be fair over time. We're partners. We're going to be fair to each other. And the truth is, you know, based on every prior market precedent for a foundational new technology like AI, you've always had a bubble. You know, Carlota Perez wrote this great book about this. And basically, markets are efficient. They correctly understand that this is a foundational technology. There's what Mokyr calls a breakdown in diversity. Everyone becomes bullish on this new technology. And I am beginning to worry a little bit about a diversity breakdown. And then you get a bubble. That bubble funds the buildout of this new technology, but supply gets ahead of demand. And you get a crash and it's a particularly severe crash if it's a debt-fueled buildout like the year 2000. And one thing really happy about, really good about the current buildout is it's still overwhelmingly funded out of operating cash flows which is a really important fundamental difference versus the year 2000. As is valuation, as is the fact that every GPU is running at 100% utilization when 99% of fiber was unutilized. So there's all these fundamental differences, but we do have to, history doesn't repeat, but it rhymes. And as investors, we have to be very cognizant of it and recognize that based on the last 200 or 300 years, you know, forget the internet bubble. We had a railroad bubble, a canal bubble, we should expect a bubble. And that's terrifying. Like nobody wants a bubble. A bubble is terrible. The reason it's terrible is if you're valuation sensitive, you like massively underperform. You get fired by probably all your clients. George Vanderheiden, who is no longer with us, great Fidelity portfolio manager, he fought the bubble in '99 and he retired in early 2000 because I think he just couldn't take it. He knew it was wrong and you know his clients were deeply skeptical. George, you're out of step. You know, he had white hair. He's truly a great man. I only overlapped with him briefly, but he was a very important mentor and friend to my good friend and mentor Jennifer Urick. So, I have a lot of Vanderheiden DNA through her. He was the same person who said being early is the same thing as being wrong. George retires because he can't take the underperformance and he can't take clients saying what's wrong with you? You don't get it. And he has like 40% of his fund in tobacco, 40% in homebuilders. And literally he underperformed, he probably outperformed the NASDAQ by like 20 or 30x over the next three years. Okay. And I have been optimistic that this fundamental shortage of wafers which really today is controlled by Taiwan Semi will prevent one. If Taiwan Semi did what Jensen wanted, I think Nvidia could sell two trillion dollars of GPUs in '26 or '27, maybe 2.5 trillion, maybe three trillion, but there is a limit where consumers would consume so much that you probably would be in an overbuild. And so Taiwan Semi, if we don't get a bubble, like we need to throw a party for them because they will have single-handedly prevented a bubble. Okay, you are starting to see companies go to Intel and Samsung.

Well, one of, you know, the history of markets is I don't know who but one of Intel and Samsung, they're not going to stay disciplined. They will break and then at some level that will force everyone else to break. So like I think a lot of this may come down to the degree to which Taiwan Semi can maintain a lead over Intel and Samsung. You got to remember it's whatever it is, it's 9, 12, 15 months.

Exactly. You know, the pace at which they expand capacity. Like if I were to watch one thing to understand whether there's a bubble, it's Taiwan Semi's capacity decisions. And I think there's a Goldilocks zone where they expand enough, they make it hard for Intel or Samsung to really truly emerge as like a at-scale second source with something, you know, well north of 30% market share. And yet they also keep this fundamental constraint on wafers that, you know, helps us avoid a bubble. And then obviously I think the Terafab is going to play into this too.

The Terafab, it's a SpaceX, I believe Tesla's involved as well, joint venture to build the world's largest fab here in America. And I think they're going to be successful. They have a partnership with Intel which is very important because they're getting access to 50 years of institutional knowledge that's just, you know, nine months, a few quarters, 12 months, 3 to 5 quarters behind the front. That's an advantage. It's also an advantage that I believe that Terafab is going to get attention from the A-teams at all the semicap equipment companies. Like one big reason Taiwan Semi caught up is ASML and KLA and Lam Research and Applied Materials. They wanted them to catch up. They don't like having a monopoly. And so the A-teams were in Taiwan working. Intel made some mistakes and presto. And so the A-teams will be here because of Elon's reputation in hardware engineering. And then just to a degree that I think is maybe hard for people to imagine in America where, you know, politics has replaced religion, because Elon had his foray into politics that makes it hard for some people in America to see him clearly, which is sad because I do think, you know, he's probably doing more for America than any other American. You know, he's single-handedly bringing manufacturing back to America. He's revived defense tech. SpaceX is in some ways the most important defense contractor in America. You know, what he's doing with Starlink is amazing for the world. He's creating all these blue-collar manufacturing jobs, which is like a goal, I think, of a lot of liberals and good for America. He's done more than any living human to decarbonize the world. And if you are upset about data centers on Earth for environmental reasons, well, here you go. You know, so it's sad, but he is a living deity in China, Taiwan, South Korea, and Japan. And having watched him for a long time, what he's going to do is they're going to recruit the best people because the best engineers want to work for Elon, especially in hardware engineering. He's going to recruit incredible engineers. And then they'll be next to Terafab, there'll be a Taiwan Town. Oh, these are your favorite restaurants. I'm going to move them and their whole staff from Taiwan to Texas and we're going to make everything the way they like it. And then we'll have Japan Town. Same thing. Then we're going to have Korea Town. We're going to have all these things exactly but dialed to recruit the best engineers. And that's just not the way that the people who run Intel at Santa Clara think. So he's going to have the best talent. He's going to have the A-teams at the wafer fab equipment companies. He has Intel which is important. It's so good for all of any administration's political goals. And I think it's different enough that it will not alienate Taiwan Semi.

Elon tends to do things differently. Everybody else has taken three years to build a data center. He built one in 122 days. You know, Samsung had to give him an office in their fab in Texas because he was so unhappy about like the pace at which they're expanding a building. We'll see.

I do think it's fascinating the returns to the frontier, all the economic returns to AI at the model layer, not all of them, but an overwhelming amount of them have been at the frontier, which is surprising to me. And I think it's been surprising to a lot of people and I think this is one of the most important questions to be answered and you need to have a hypothesis on it as an investor. Are frontier tokens going to continue capturing the overwhelming majority of economic value created at the model layer? And it is surprising. Like I just remember when Gemini 3.1 Pro came out and it was mind-blowing to me. It was so good. And today it's intolerable.