我与 Dylan Patel 和 Stacy Rasgon 的对谈

All right. Hello, everyone. Y’all ready for this? This should be good. This should be fun. We’re going to talk about the industry, look ahead to 2027 and 2028, and hear these two gentlemen’s thoughts about where it’s headed. We’ll try to cover it all, from financing to GPUs and CPUs, down to foundry. Let’s get into it. I actually want to start with the finance side.

1. Financing the Buildout

Lately, a lot of the news has been about the cost of capital. We know hyperscalers are investing a ton, and free cash flow is going negative. Now we’ve got the 10-year Treasury over 5%. Even if you zoom out to CoreWeave, there’s talk about its bonds yielding 15%. My question is, should we be concerned? As companies invest more and more, is it becoming more and more expensive to finance this? I’ll start with Stacy.

Look, there are a lot of constraints in general that are moderating the pace at which this buildout can happen. These are power, supply, semiconductor tools, and now maybe financing. I’m a Wall Street analyst. If you look at the stocks, they’re still up a bunch on this, but they’re clearly way down from the peaks. The market is telling you something: it’s telling me they’re worried there’s a lot of uncertainty about the pace and sustainability of the buildouts, and maybe the peak is approaching. Clearly, financing is one of those concerns. It’s probably near the top.

You mentioned a number of different reasons. The hyperscalers were the most cash-generative businesses in the history of humankind. Now they’re burning cash and probably going to be raising equity. Meta just had a new share-class structure that, among other things, will enable it to do that. We saw Google raise money, and others did too. You mentioned CoreWeave’s bonds. That’s its unsecured debt. It’s not what the company is paying right now; that’s where the bonds are trading in the market. But that would suggest that if they go out to the market today to get new unsecured debt, that’s roughly what they’ll be spending. And it’s a lot of money.

I think these are all broad concerns right now. Though I think that, at least for the next year or two—which is as far out as I can look, frankly—we’re probably okay. But I do think it’s pretty clear that revenues need to grow to the point where this becomes less of a concern. Maybe that happens. I don’t know. SpaceX reported earnings this week—or last week; I can’t remember. He’s running out, clearly he’s running out in the spot market, and things are very, very constrained. The only thing we’re hearing is that nobody has enough compute.

He’s renting capacity out on short-term contracts at something like $50 billion a gigawatt, at those kinds of rates. I don’t think it’s actually that hard to imagine a scenario where this stuff gets covered and then some. But again, it comes down to how sustainable those rates are, and that raises broader questions. It really depends on what demand actually looks like relative to supply, and how long the supply-demand mismatch lasts. These are questions Dylan would probably have a better idea about than I do. But absolutely, it is a concern. It is one of many constraints out there that I think are moderating growth.

I’ll say one more thing, though. Nobody likes the idea that these constraints are moderating growth, but at the same time, it may be a good thing in some sense, too, because it stops things from going nuts. We saw this during the tech bubble. There’s dark fiber in the ground today that was overbuilt and never used.

Some of these things may actually give us stronger-for-longer behavior, because they almost force the demand that is there to push out supply even more and allow things to be better matched, rather than going completely nuts today and risking a crash down the line. So I don’t actually think constraints in general are necessarily a bad thing. But I do know how they’re perceived, and clearly they’re impacting the trajectory of the stocks in the near term.

Yeah, totally.

Echoing Stacy’s point, in any buildout, everyone’s going to want to build. When we look at any buildout over the last few decades, we’ve always had the industry double-order, triple-order, and build too much capacity. Supply chains are really bad at forecasting demand. As Stacy mentioned, there is a bit of self-moderation, especially on the finance side. If we go back a year or two, any cloud that wanted capital could get it. They would pay a pretty high interest rate, but they could get it.

Now, if you don’t have really high-grade offtake, you’re not able to build if you’re some random cloud. Ultimately, the hyperscalers can take their free cash flow down to zero, raise some debt, and their earnings are obviously growing. So their capacity to build out grows, but they can’t do north of $2 trillion in capex next year. They can’t do $2.5 trillion in capex next year. That moderates how much they can spend. And when you look at the neocloud industry, which accounts for non-hyperscaler demand, nearly half of the stuff being built out, how are they going to get it financed if they don’t have offtake from a hyperscaler?

The market’s already saying, “Hey, if you’ve got offtake from Anthropic or OpenAI, we’re going to price it at 13%, 14%, 15% debt.” We saw that recently with Volta. They’re making a cluster for Anthropic, and their debt is basically a 12% coupon, but they sold the bonds for 95 cents. So it ends up being like a 15% interest rate. If we look at everyone who isn’t getting offtake from these companies, hyperscalers, or the major AI labs, how do you build that capacity? How do you build capacity to serve smaller enterprises? How do you build capacity to serve new startups? You literally can’t, unless it’s coming from a hyperscaler.

That’s Elon’s whole gambit: he’s making all this capacity that he can sell at very high prices on short-term rental contracts to startups and AI labs. But ultimately, it is quite a dampener. We’ve seen Nvidia recognize this as a very big problem for them. They have five customers who can buy as much as they want, or as much as they have capital for, but what about the other part of the market? And how do they enable that?

They’ve gone and done really creative structures with Sharon AI and Fermi and a few others. They’ve taken revenue shares on the cluster in order to guarantee that it’s financed. They’ll say, “Hey, we’ll rent the cluster from you at a guaranteed price of $3.50, which means the cloud makes zero money, and then if you rent it to anyone else above that price, we get a revenue share.” We’re starting to see more and more of these structures. In fact, even to build a cluster for Anthropic, Lambda had to do a revenue share with Nvidia.

In this case, it wasn’t 50%, like Sharon AI and Fermi had, but it was still a 10% revenue share. So it’s really tough for these clouds to get the capital to build, get to the point where they build everything they want to build, and reach the oversupply scenario, just because of the financing. We’re seeing a shift from long-term loans and debt to something shorter-term. Broadcom has done the same thing with Anthropic, where they’re basically leasing the chips to them. It’s rent-to-own. They’re stretching out the payments over a few years.

The whole name of the game here is that if anyone wants to get capacity and they’re not one of the big companies, they’ve got to come with a deal, a structure that has Nvidia, Broadcom, or AMD backstopping these clusters. Ultimately, the financing is a big dampener. The hyperscalers also have very large off-balance-sheet and on-balance-sheet commitments. Meta has $1.4 trillion of commitments for compute and data centers, and so on and so forth.

That can then get financed by the whole market: Meta can get a loan and buy a data center, rent a data center from someone, or rent chips. Then those companies can turn around and get capital for it. But ultimately, there’s a big struggle for anyone to do more than that. A lot of people have worried about even Meta’s financial structure. Meta is completely fine, but even the capital markets are getting spooked by that. Financing is part of the holy trinity: land, power, shell; chips; and financing.

That’s good. That’s interesting. So it’s getting more difficult to finance. The hyperscalers are fine, but maybe it’s a little bit harder for neoclouds. We’re seeing these new creative structures come around to help with that. If we set finance aside and go back to the traditional bottlenecks that are ultimately limiting the pace of growth in the industry, talk to me about where you see things in 2027 and 2028. ASML says they’re sold out in 2028. TSMC always has more demand than supply, even though they’re doubling capacity for 2 nanometers. And then people talk about power as a bottleneck.

Dylan, I know your team recently—maybe this week—put out a post about up to 50 gigawatts coming online next year. Maybe I’ll start with you. How would you stack-rank these different parts of the semiconductor supply chain, and where should we be looking for constraints?

Yeah. On the semiconductor side, the biggest constraint, besides memory, is substrates. Substrates, PCB materials, substrate materials—all of these components were severely underinvested in. Many of the companies in Japan that create these materials are more conservative and unwilling to invest heavily in capacity. So we’ve seen tremendous bottlenecks in what can be made over the next two years, for accelerators as well as CPUs and every other component in the supply chain.

I think substrates are the biggest bottleneck of all the bottlenecks, including land, power, shell, semiconductors, and tools.

I’d say the next bottleneck would be memory. Following that, there are a lot of bottlenecks around the desire to ramp optics to a really, really large degree, with co-packaged optics or near-package optics. There are humongous constraints on ramping that. We simply cannot ramp enough, even if everyone wanted to do CPO in 2028, or even if one major company wanted to do CPO in 2028. They could not satisfy their volumes in terms of lasers, fiber attach, and all the different components that go into it.

As far as land, power, and shell, power is definitely a big constraint. But one thing our research shows, which I think is counterintuitive to much of the market, is that next year there are data center constraints, but a lot less than people realize. Nearly 50 gigawatts of data centers are being built globally: 7 or 8 gigawatts in China, about 35 in the U.S., and a bunch in the rest of the world. Australia, India, and Malaysia are some of the big places, but data centers are being built at a tremendous pace all over the world.

While power is a huge bottleneck, people can find creative ways around it, whether it’s turbines, reciprocating engines, or fuel cells on site. All of these things are helping alleviate the power bottleneck. So I think power is actually a lot less of a constraint than most people think when they say data centers and power are the biggest constraints. They’re constraints, but I think they’re a lot less significant than substrates, memory, lasers, and other things in the optics supply chain. People can ramp power capacity over a shorter time span. Those solutions are significantly less complex than the semiconductor supply chain.

Yeah, you bet. I think it’s really interesting on the power side. Jensen and Hock at Nvidia and Broadcom have given some pretty sizable guidance. Jensen said he’ll grow his revenues by at least 70%; Hock said 100%. Those are actually constrained numbers, and they’re looking primarily, I think, at land, power, and shell as their constraints. Maybe that’s because they’re semiconductor companies. They know what they can make in terms of semiconductors. But on the land, power, and shell side, I agree there are ways to get around this.

Although even for some of the things going on site, I don’t know what the lead time is on a turbine these days. Musk says he’s going to start building his own turbine blades to get around that. I think the memory side is really interesting. The memory stocks are all trading at 2, 3, or 4 times forward earnings, and pricing has gone through the roof. I don’t cover the memory stocks, by the way. A colleague of mine does, but take a look at SanDisk: they guided to quarterly EPS higher than the stock price was 24 months ago. It’s really amazing, because pricing has gone through the roof and they’re all trading at single-digit P/Es. That’s how you trade things at the peak of a cycle, and everybody’s worried that, especially as you get into 2028, supply will start to come online and pricing will collapse.

I sit here and always wonder what happens if it doesn’t. Look inside these racks. Look inside an AI rack: 70% or 80% of the silicon area inside that rack is memory. It’s mostly DRAM. There are other issues when you’re making high-bandwidth memory, the memory that goes into the AI chips. There’s something called a trade ratio, but it takes 3 or 4 times as many wafers to make a gigabyte of high-bandwidth memory as it does to make a gigabyte of standard DDR5 DRAM that would go into a PC or smartphone.

So you’re sucking up wafers from that. We need a ton of memory. And this feeds into one of the other bottlenecks, which is semicap: the actual tools to build this stuff. We had a company report this week. They gave a preliminary outlook for WFE, wafer fabrication equipment, for next year. They said more than $200 billion. I think a lot of investors thought it was going to be a lot higher than that, which I found really surprising because I just don’t think it’s possible. We don’t have the fabs.

There’s another constraint: we don’t have clean rooms. They’re all full. They’ve actually been squeezing as much as they can out of the existing footprints. That’s why the WFE estimates at the beginning of this year were maybe $120 billion. We’ll probably end the year at $155 billion. That’s all from finding better ways to use the existing footprints, but now they have to build, and it takes years to put the stuff into the ground. Clean rooms do start to come online a little in 2027, more in 2028 and 2029. At that point, I guess we’ll know if there’s a supply-demand mismatch or not.

But if demand profiles continue at the rate people seem to think they will, memory may still be a constraint. You talked about substrates, PCBs, and other things. Again, going back to the tool vendors, they’re all complaining about short-term supply constraints that are actually impacting their own ability to deliver upside even in the near term. It’s exactly that kind of thing: PCBs, semiconductor chips, and even other components, valves and mass flow controllers. Simple stuff is affecting other parts of the space as well.

We saw a little bit of this during COVID. I do think the semicap players, at least, and most of the folks in the supply chain learned some lessons from COVID. I think they’re managing that sort of thing better now than they used to. But at the same time, we’ve got all kinds of constraints in the spaces that need to get fixed before the other constraints can actually get fixed. So you have these layers to things. I don’t really see it changing. I think things are going to be tight for a while, frankly. I think this year is fine. I think next year is going to be the same thing. 2028 will probably still be tight. 2029, I don’t know. That turns into a demand question. But I suspect we’re going to be tight for a while.

Okay. So if we’re tight for the next couple of years, and we’ve talked about prices skyrocketing, with the memory industry making margins of 80%—

Higher, higher.

Yeah, 88%, 90%, right? So do you see all those players, everyone constrained up and down the whole stack, continuing to charge for their value?

Well, it’s funny. In this industry, pricing coming down was just taken as a given. The semiconductor industry as a whole was historically a massive value destroyer. It never covered its cost of capital. All of the value flowed to the consumer. I did a piece of work at McKinsey many, many years ago. It was called “Value Creation in the Semiconductor Industry.” We defined value very classically: economic profit minus WACC times your average invested capital.

We aggregated it across about 150 different companies, 12 different subsectors, over a full cycle. I think the initial work was from 1996 to 2004. Over that period, collectively, the industry destroyed $5 billion in aggregate economic profit. If you took Intel out of the mix, it was $50 billion destroyed, which was a big number back then. The industry didn’t really start covering its cost of capital, meaning that all the economic profit it had created before that was flowing directly to the consumer. It was a wonderful time to be a customer for semiconductors.

But after 2008, there was a lot of rationalization and consolidation, and at least we started to come in and meet the cost of capital. Then COVID happened, and what the industry discovered during COVID was that it has pricing power. I think that was something it didn’t feel in its bones before. We had input costs going up, and I cover a bunch of companies. I didn’t have a single one experience margin compression because of input costs going up. They all raised their prices to offset it, and they raised their prices at their margin structure.

So if their cost went up by 100, they raised prices by 160 if they had a 60% gross margin, and they had no problem passing it along. I’ve said that inflation saved the industry during COVID. If you look at industry unit shipments in 2023 or 2024, they were flat compared with 2019, and blended ASPs for the industry were 50% higher, even if you take out memory. We’re seeing that on steroids right now. Clearly, memory pricing has gone through the roof. The industry, by the way, is going to do $1.5 trillion this year, maybe more. We’re going to be up probably 100%.

Seventy or 80% of the growth we’re seeing this year is memory ASPs. A big chunk of the rest is also non-memory ASPs. Units are up 10% or something like that, whatever it is. But the industry is now flexing that pricing power a little more, and by and large, I think customers are accepting it. They don’t really have any choice. So, yeah, it may stay tight. If it stays tight, I think we’ll probably continue to see this.

Yeah, I would say the semiconductor industry is very interesting in that some people flex their pricing power in the good times and suffer really badly in the bad times, while others just sort of eat margin the entire way through. If you look at TSMC over the last decade, they’ve risen from 45% gross margin to 60% now. It’s been a very steady climb up. Yes, they’ve taken more when times were good, but they haven’t given much back when times were bad. And you look at memory: they’ve gone from basically negative margins just a few years ago to now, you know—

People forget we’re coming off the worst memory cycle since the tech bubble just 24 months ago.

Negative gross margins.

When you look across the supply chain, some of these folks are taking tons of pricing because of the structure of the market. TSMC, where they’re a near monopoly and a very long-time partner, has not really jacked up pricing even though they could. Whereas memory companies, if Apple’s going to screw them in the downturn, they’re going to screw Apple over in the upside of the cycle. That’s pretty clearly what’s happening.

To build a gigawatt in 2024, it was roughly $40 billion across chips, land, power, shell, everything. Now it’s north of $60 billion. The price to build a gigawatt of compute has skyrocketed. Of course, not many people were building a gigawatt—or really, no one was building a gigawatt—in 2024, and a lot of people were building gigawatts in 2026.

You’re getting more out of the gigawatt now, too, by the way.

Yeah. And that gigawatt’s profitability is huge. I think everyone is recognizing that profitability extends. A lot of the mechanics I mentioned earlier mean you can’t build a cluster as a cloud provider unless you have offtake from a hyperscaler, which means you’ve locked in pricing ahead of time. Elon is exploiting that. If chips, land, power, and shell are constrained, and I build these things but don’t sign a contract until they’re actually up, I can make $50 billion per gigawatt instead of $15 billion.

Anthropic, in its recent disclosures, is making on the order of $70 billion to $80 billion per gigawatt off of its inference capacity. The AI labs are kind of forced to eat the cost. Yes, their margins are 75% to 80% when they’re spending $15 billion per gigawatt for compute, but if they’re spending $50 billion, their margins are much smaller. So it’s not clear who’s going to eat all the cost. Nvidia has had to keep its margin roughly the same, slightly decreased, but it’s had to eat the memory pricing increase and other costs in the supply chain, like substrates, optics, and all these other things.

The prices they charge their customers have soared despite their margins being flat. We’re seeing something really interesting: who gets to keep their margin, who gets to expand it, and who’s going to eat margin. I think it’s a very fun time to be on the pricing side, because you’re discovering in real time that you can sign a deal at many different prices. The best salespeople are going to get really insane pricing, and hopefully they’re rewarded for it.

Hilarious. Okay, Stacy, I kind of like what you said about COVID. It was almost like a dress rehearsal for AI, for the industry to start realizing, “Hey, we can charge for our value here.” And Dylan, you make an interesting point about a system builder like Nvidia having to decide, as the prices of all the components go up: Do we pass it on, or do we keep our margins the same?

Now, I’m feeling a little tension. There’s still going to be buildout all the way through. We’re not sure what’s going to happen in 2029, but clearly supply is limited and demand is high for the next couple of years. And yet we started by saying it’s getting harder and harder to finance. I’m wondering if there comes a point where those people at the top of the stack—the Metas and the model lab companies—push back and say, “This is just getting too expensive,” or, “Beggars can’t be choosers.”

Remember, you have to look at more than the margins and the cost. You also have to look at what they’re getting out of it. I’ll take Nvidia as an example. Their pricing goes up with every generation, but what you’re getting out of it is going up more than the price, which I think is important. You went from Ampere to Hopper.

Yeah.

The ASP roughly doubled, and performance went up 10x or whatever. Dylan would have better numbers for exactly what the performance increase was, but it was a lot. It was a very NPV-positive sort of thing. No problem. You can pass that along.

You go from Hopper to Blackwell, and it’s a little harder to tell. The Blackwell architecture is much different, but call it roughly a 50% increase in ASP per GPU. Performance went up anywhere from 3x to 30x, and maybe more given some of the later optimizations and software. Again, no problem.

You go from Blackwell to Rubin, and it’s the same thing. The cost of the racks is probably doubling, and so is the cost per gigawatt. I think Nvidia said that, for its own reference, it was something like $25 billion for Blackwell and $40 billion for Rubin. But again, performance is up 10x. That’s what needs to happen. You need to keep giving your customer value, because ultimately that’s what’s important. It’s how much performance you’re getting out of that gigawatt of compute relative to how much it costs. As long as that gap is getting bigger, I think it’s still not easy, but it’s palatable to do this.

Hypothetically, if they were to roll out Feynman and it were only 20% better, then I’d start to worry a lot about pricing and margins. This is always a huge controversy for a company like Nvidia, because it does have margins up in the 70s on a massive revenue base. People wonder: You’ve got competition, ASICs, other GPU vendors, and other things—how can they hold those margins in the wake of all that? In my view, as long as they can keep improving performance at a significantly higher rate than their ASPs are going, I don’t worry. They have to raise ASPs because their own costs are going up.

Things are getting more complicated, so you have to keep that performance-to-cost equation working for the customer. As long as they can do that, I don’t really worry about their margins. But if the pace of innovation slows, then I would worry an awful lot. They’re aware of this, too. I went to the Hot Chips conference in 2023, and I didn’t understand anything there. I’m a process engineer by background, so all the computer science and chip design stuff goes over my head. But Bill Dally from Nvidia, their chief scientist, gave the keynote. It was the one thing there I understood.

What he was talking about was exactly this: GPU performance improvements. He made the point that over the last 10 years, they’ve improved GPU performance by 1,000x. It was a whole bunch of different levers that they pulled, but they got there. He said that over the next 10 years, they want to improve performance by another million times. So far, they’re on that trajectory, and if they can do that, then I don’t worry as much. If they or anybody else falls off, then I worry a lot, because ultimately somebody has to pay for it. In a perfect world, that somebody is the customer, and there’s a lot of value from it. If the costs go up at an unsustainable rate, the suppliers have to start absorbing some of it, and then it’s a dogfight to see who does and who doesn’t.

Yep. Gotcha, Dylan.

That’s certainly the trend: all of these companies are trying to make chips. OpenAI showed, I think for the first time ever in the industry, that you can make a first-generation chip that’s really good. They let us into the lab. We tested it on our open-source InferenceX benchmark, running all these open-source Chinese models, and the performance was better than Blackwell. That’s a first-generation chip, and they used AI heavily in the design and programming process. Anthropic is starting its chip program, and it wants to do it in half the time OpenAI did.

They want to get a chip. These are aggressive goals, but if models are able to help them really accelerate this process, then you’ve got Google selling TPUs externally. Amazon’s Trainium is really quite good now and fully usable in PyTorch. Meta has a very aggressive roadmap, and so does Microsoft.

If you look at how I said it went from $40 billion to build a gigawatt to $60 billion, and you remove all the margin from the chip design houses—Nvidia, Broadcom, Marvell, and so on—you’re actually at about $20 billion a gigawatt. If you were able to do that, then you could pay for 5 gigawatts with 1 gigawatt’s worth of compute running inference. The economics here are obviously very aligned. As Stacy said, these companies just have to run like crazy. You can’t afford to be only 2x better in performance per dollar per year, because you’re going to get completely smashed by these other companies.

If Nvidia can keep up this incredible pace, sure, it deserves to continue to earn margin. But if the other competitors—and their main major customers—are able to design chips and deploy them, then they can scale their manufacturing. Memory without a markup, TSMC wafers, substrates, optics, and so on are really, really cheap.

Those companies have a lot of room to run on pricing, especially if things get more disaggregated and Nvidia loses its position or Broadcom loses its position. That’s why those two stocks trade at basically their lowest multiples ever. A lot of people are worried about that. Is Meta going to disentangle Broadcom? Is Google going to disentangle Broadcom? Are both of these companies going to stop buying Nvidia and just insource completely? I think the component vendors are going to be able to jack prices up like crazy.

The question is whether Broadcom and Nvidia can continue to earn these incredible margins. At least for the next two years, it looks like yes. But beyond that, is it off the table now? At the end of the day, chip design is very, very complex, but models have solved software, and programming accelerators is actually really easy now. It took us 6 months to get Nvidia inference working for InferenceX, our open-source, largest open-source inference benchmark. It took us 6 or 9 months to get AMD working. We got Google working in a few months, and we’ve got Amazon working in a few months.

We got OpenAI working in a month and a half. The pace from chip works to being able to run software on it is rapidly compressing, and now these companies are all turning their eyes to chip design. Can that actually work? It’s similar in the sense that it’s very verifiable, like software. Obviously, there are a lot of challenges with data and scaling, but the big question for the semiconductor industry is whether design becomes more commodity-like, as software has, in which case everyone can design chips.

Obviously, there’s the pesky fact that you have to pay $50 million for all the fixed costs of masks and EDA tools and all this other stuff. But the number of people designing chips is going to explode just because programming them is much easier. If designing them becomes much easier, margins have to fall. We’ve seen SaaS companies get obliterated in the market, and we’ve seen Nvidia and Broadcom really trade sideways despite their earnings growing at 2x-plus year-on-year. We’ll see what happens, but that’s the big question.

So, Stacy, I liked your point that it’s not about the fact that it’s going from $25 billion to $40 billion to $60 billion. It’s just: are you squeezing out more tokens per dollar? That’s all that really matters. And then, Dylan, you brought up XPUs and OpenAI. Very interesting example. How should we think about the growth of GPUs and the explosion of XPUs over the next 2 years? There are all sorts of AI accelerator startups taking different architectural bets. Is it going to look any different in a couple of years?

For XPU companies, there are sort of 2 paths. If you’re a hyperscaler, you can take a more conservative design approach and do similar things to Google or Nvidia, and it works out. But if you’re a startup, you have to do really crazy things. That’s why, in the first generation of AI hardware startups, the ones that did the craziest things, Cerebras and Groq, were the most successful. As we look at the current crop of startups popping up, only the ones that do really crazy things that are very different and can’t immediately be copied by Nvidia, Google, or AMD are the ones that have a shot at succeeding.

The nice thing is that so much capital is flowing, and a lot of really crazy ideas are being tested by companies trying to build some really insane formats for XPUs. In 2023 and 2024, Nvidia swallowed up a huge percentage of total HBM, total CoWoS, and so on in the industry. As we looked at 2025, it started to moderate, but then in 2026, Nvidia lost share. They’re using a lower percentage of the total CoWoS wafers and a lower percentage of total HBM. And again, next year they lose share.

Ultimately, the number is growing so fast that their earnings are still doubling year-on-year next year, or more. But this is a big thing that’s already happening: Google’s taking share, Amazon’s taking share, and AMD’s taking share.

On a unit basis, is what you’re—

Sorry?

On a unit basis. Yeah.

Yeah. I mean, units are always difficult. Who makes a big chip, a small chip? But CoWoS wafers and HBM are kind of hard to measure. And revenue, right? Dollars spent on foundry wafers, on HBM, on these sorts of things. These guys are taking share. Ultimately, Nvidia’s base is growing so fast that they’re still going to double earnings.

As we go up to 2028 and 2029, maybe growth rates aren’t as high. Maybe revenue isn’t increasing 100% year-on-year. Then what happens? Does growth moderate to the point where Nvidia actually loses share entirely, or are they making some crazy bets for their next-generation architectures past Rubin? Maybe that’s to Stacy’s point: they just have to run really fast, and that’s how they defend themselves.

This is one of the reasons Nvidia has moved to an annual cadence for chips. It wasn’t just for them, of course; it was for everybody else to try to keep up. They’ve got a scale that makes them hard to catch. People always talk about moats. Nvidia’s got a lot of them. There’s the CUDA software, although we can have a debate about whether CUDA is actually a moat anymore or not. There’s also the ecosystem and install base of infrastructure, and even their balance sheet and financing are a moat.

They’re leveraging AI to speed up their own design. The industry isn’t doing anything there that they’re not aware of or using. But it’s a race. Jensen, at least, is paranoid. He took the old Intel lessons to heart. He’s looking over his shoulder every day.

Yeah, this is absolutely why these stocks are trading where they are, right? They’re trading cheaper than they ever have. I wrote a note—this was probably 9 or 10 months ago—but we were looking at Nvidia relative to the SOX index, which is a broad index of semiconductor stocks. As of when I wrote the note, I think there were only 13 days in the last 10 years when Nvidia had traded cheaper relative to the broader semiconductor industry. That was 9 months ago. It’s cheaper today.

These are the things that investors worry about. It’s beyond just worrying about overall demand sustainability. If the demand’s not there, we’re all screwed, regardless of whether it’s Nvidia or anybody else. Within the group itself, this competition is clearly their biggest threat, because they’re the forerunner by a mile and everybody’s trying to eat their lunch.

But at the end of the day, that’s semiconductors. There’s always somebody trying to eat your lunch. It’s up to you to make sure you can stay ahead, and that’s something that will probably get harder and harder as you go forward, just naturally.

Sure. So when we’re talking about GPUs and XPUs and unit share and all these competitors gaining share, we’re really focused on the data center and hyperscalers, pretty much. What do you think about enterprise on-prem as an opportunity for someone like Nvidia?

Well, actually, even today Nvidia splits its revenue out by hyperscale and non-hyperscale. It’s not exactly clean, because some of the labs are in non-hyperscale, but some of the hyperscale spending is going to support the labs, or vice versa. So it’s not clean, but they’re about 50/50 right now. I think the labs, directly and indirectly, account for roughly 20% of Nvidia’s revenue right now. It’s sizable, but I’d say Nvidia’s revenue base is maybe even a little more diverse than we had thought, which is why they’re giving that presentation now.

At the end of the day, though, you clearly need enterprise adoption. That’s where this is. If we don’t have enterprise adoption, there’s no industry. It’s still early days, but I think we’ve got at least one killer application for enterprise, which is coding. Agent coding revenues have gone vertical. Does anybody write software by hand, like artisanal code, anymore? I don’t know that they do. I think everybody uses it.

Yeah. What’s coding? Right.

I don’t know. I also think that on the enterprise side, people are looking at revenues, but some of the actual impact may not show up as revenue. It may show up as productivity. It’s harder to measure, but I’ll give you a dumb personal anecdote. I’m a financial analyst. We build models. One of the guys who works for me has been using Claude to build models for his own personal account. We tried this among the team a few months ago, and it gave us garbage. But he said, “I think it’s working.” I said, “Okay, I want you to—we’re going to get on a Teams call, and I want you to throw it up. We’re going to have it build a model for a company we already cover, from scratch.”

It one-shotted a model in an hour and 20 minutes. It used 10 years of history, went through all the different resegmentations, captured it all, and even forecasted it. It listened to what the CEO said on earnings calls and tried to use logic rather than randomness. It was something that would have taken one of my guys a week to build and would have taken me 3 months to build. It did it in an hour and 20 minutes, so I could do more with less.

There are other implications of that. Are we going to be in a world where AI adoption in the enterprise is rampant and we have 20% unemployment? I don’t know. There are other factors to think about here, but I think we’re still early. Most of the anecdotal evidence we’re hearing right now—because this is the overarching debate—is about whether there’s ultimately a return from using this stuff. I think anecdotally we’re coming down pretty solidly on the side of yes, there is a return, but it’s absolutely still early. A lot of companies are still in the experimentation phase. They don’t have the expertise to know what to do or what they need to do, so it takes time. But I’ve been encouraged by what I’ve seen.

The only thing I would say about enterprise adoption is that I don’t think enterprises are going to adopt that much on the hardware side and run their own models and such. I think the efficiencies of running batches and the scale economics of models mean that most people are just going to buy tokens and implement them, combining them with their own data. Ultimately, that’s the real way most people are going to consume AI.

So it still leaves this sort of stratification. I think the industry has gone more and more in that direction. If you go back 2 decades, everyone bought their own infrastructure and deployed it themselves, and now most of the spending is in the cloud.

Most of the spend is going to be on buying tokens. Even pharma companies and hedge funds—most of their compute and AI spending is going directly to models, not to buying compute and running open-source models. There will be providers who take that on, but I’d love to talk more about this, although we’re a bit over time. I think it’s very fun.

I see a big zero minutes left. Okay, I guess that means we’re done. Okay, guys, this was awesome. I learned a lot. This was great. Thank you so much.

You bet.