Transcription
So, first, this is a bet-the-farm bet by Lisa Su, right? She's given away 10% of the company if the compute gets deployed. She said a couple weeks ago, "We're in year two of 10 of a compute buildout across the country." So, I I I have a couple of charts here. Why would she make a bet-the-farm bet, Jason? Right, that's an important question.
Remember, Nvidia just did a huge deal with OpenAI. They didn't give away any of their company. In fact, they got the right to buy >> part of OpenAI. So, look at this chart. This is pretty wild. Just in '22, in 2022, two and a half years ago, Nvidia and AMD had basically the same revenue, $25 billion, right? This year, Nvidia will do 10x that at roughly, you know, $210, $230 billion, and AMD will do $33 billion, right? Not much more than they were doing when the companies were tied in 2022. So Nvidia's captured nearly 100% of the incremental AI data center revenues over the course of the last 2 and 1/2 years. And I think there's this notion that somehow Nvidia just popped out this special AI chip. But I think if you listen to Jensen, understand, study the company, it's because they have this ecosystem of software, networking, extreme code design. The unit of compute is no longer the chip, right? It's the entire data center, which is composed of, you know, five to 10 different chips. Performance per watt, right? Power being the constrained resource here is everything. And Nvidia's just crushed the competition.
So now put yourself in the shoes of Lisa, right? And she's clearly not on the wave. This tsunami has come. Jensen's riding it. He's capturing 100% of the wave, and she's not even yet on the wave. Her MI 350 was just not competitive. And so they have one shot. Either the MI450 gets adopted and they get back into the game, right? Or or or or they're out. She's a total warrior. I think she believes in the 450. She went to her board and she said, "Listen, this is we got to take the shot here. We got to bet the farm. If it works, she's going to get $150 billion of incremental revenue just from OpenAI, right, for building out 5 gigawatts." And on top of that, of course, it could unlock a lot of the other market because now it will have validated that they have a a a chip that works. But it's far from a done deal, far from a conclusion whether the 450 is going to work. Can it compete against Vera Rubin? Can it compete against uh Reuben Ultra? All super important questions. So that's one, you you know that that's the AMD deal.
Why have we switched from saying, "Oh, Colossus bought, you know, XAI's data center, 100,000 H100s, 200,000 H100s" to now framing these in gigawatts? I think it's important for people to understand why that's now how deals are being framed, just in the last 3 months changed. >> Yeah, I think I think there are a couple of constraining features. Number one, chips change, right? So a 100,000 H100s is not, you know, is not apples for apples with a number of GB200s, you know, Grace Blackwell 200s or GB300s. And so it's hard to talk about them. You're comparing apples and oranges when you're comparing these data centers. So a unifying metric of compare, right, is the the the power that goes in. Everything starts with the power. It's the constrained resource. So we can compare a gigawatt of power because it's, remember, it's power in and it's tokens out, and we can compare that over time and normalize across all these different chips. And it's not just these two companies, you got Tranium by Amazon, TPU by Google, Cerebrris, Gro, etc. There are a lot of folks in this game. And so again, I think what you saw this week in this flurry of announcements, right, is that you have the market leader that's basically captured 90 to 100% of the incremental demand for the biggest thing that the the data center and the chip market has ever seen. And every other player is looking at what they have to do to have a shot to capture part of this wave. They're high-risk, high-reward bets.
The other thing I would say is, and again, shut me up and anybody can jump in if they want to, but I want to decompose. We're hearing these estimates of 100 gigawatts, 4 to 5 trillion of buildout over the course of the next four years. If you look at that same chart I just showed you, that's not what Wall Street estimates are, right? There is a lot of disbelief on Wall Street as to what's going to get built out. So if you look at that estimate, you know, starting in 2027, it basically flatlines '27 through '29 for Nvidia, right? Less than 10% CAGR in terms of the growth for Nvidia. They have by 2029 them doing $360 billion in revenue compared to $210 billion of revenue this year. To put it in gigawatt perspective, that means going from 4 to 5 gigs per year in '25 to 9 gigs in '29. That's a big step up, but a long way from the 100 gigawatts you were hearing, you know, bantered about on CNBC this week. >> That's for a different reason. Nick, put this put this quote up there, which is this famous quote from Nathan Rothschild where he said, "I care not what puppet is placed upon the throne of England to rule the empire on which the sun never sets. The man who controls Britain's money supply controls the British Empire, and I control the British money supply."
Why is that quote so interesting as applied to AI? I think that what you're going to see, and you're seeing it in those revenue graphs, is that there is a traffic jam that's happening in growth where it won't be the ability to actually build next-gen silicon, but it'll be the energy inputs that will constrain it, and it will be the ingredient inputs that constrain growth. And so the companies that then control those elements of the supply chain will actually come into power and rise to power. So what is one example of this? One example of this is when you look at the architecture of Nvidia's chips, one of the things that they have made a huge bet on is HBM. And this is a memory structure. And what's so interesting about that is when you look at the HBM market, it is effectively Nvidia who takes up the majority, the majority majority, and then Google. And now AMD's next architecture actually needs to sit on top of HBM. Now they're going to have to go and step into that supply chain and try to ask for share. Where will that share come from? And this is where the people that then control that supply, SK Hynix and Samsung, will have leverage. Now, this is what's interesting about a deal that OpenAI announced two or three weeks ago. Sam was in Korea, and you saw him shaking hands with SK Hynix and Samsung. And my initial thought then was, why is OpenAI doing a deal with the memory maker? And then it occurred to me, wow, he's buying forward capacity on HBM because now he can allocate that share. So when I saw that deal and I saw that equity, it reminded me of this Rothschild quote. Sam has allocation, and now Sam can allocate allocation, and then as a result get a tax. And I believe that the warrants and the equity are effectively that. What's a different example? Brad mentioned this before, but energy will be the gating item beyond a shadow of a doubt. And so if you control electrons, any form of electrons, hydrocarbon to electron, electron to electron, photon to electron, doesn't matter, you will then be in a position to start asking for equity, upside, participation in these companies in a way that you could never do before. You would just have been a linear member of the supply chain, a low-margin participant. So when I look at where we are, I think that the really interesting question now is to ask what is the second and third-order degree inputs that are critical to allowing the big foundational model makers, to allowing Nvidia, to allowing AMD, Broadcom to thrive. And that is where I would start to look because those that control those resources are going to dictate the pace and the scale of this AI expansion.
Okay, coming around the horn here to you, our ZAR of AI, David Saxs. When you see this massive amount of deal-making occurring, it's got to warm your heart a bit. American exceptionalism at work here. People swinging for the fences. What's your take on Brad and Chamat's overview of where we're at here in the fall of 2025 in the AI race? >> Yeah, I I don't really like to take sides on these deals for the obvious reason that we want to be supportive of everybody and just have a healthy environment for competition. So, as long as there's >> investment going on and as long as there's a lot of competition, those are good things. That's what we want to see. So, I tend to think this OpenAI AMD deal is evidence of that. There was the OpenAI Nvidia deal. There was the Nvidia XAI deal. There's just a ton of investment going on, and that's what we want to see. So that's all really good news. >> I was struck by Brad's chart about Nvidia's revenue in the outy years. The amount of capex or investments going to go into data centers and compute more than a few years from now. I think that's really hard to predict. >> Impossible, right? Yeah. >> Very hard because on the demand side of the compute, it's going to depend on new applications that get created. So, for example, the demand for tokens depends not just on AI chatbots, but on the new agents that are coming along. You have these new video generation models, Sora, and now XAI's just released something. So we don't really know what the demand for tokens is going to be. I think it's going to be huge. I think there's a lot of applications that haven't even been invented yet.