SemiAnalysis' Jeremie Eliahou Ontiveros on the supply/demand dynamics of AI and data centers
Andrew WalkerJeremie Eliahou Ontiveros
Ontiveros rejects “unlimited” as an investment premise, but his measured forecast is still a widening AI data-center shortage. Global IT capacity grew roughly 4 GW annually from 2019 through 2023; in 2024, NVIDIA hardware alone added about 5 GW of demand and AI accelerators collectively added 7-8 GW. SemiAnalysis expects construction to lag that demand for several more years, leaving upside across the infrastructure chain.
DeepSeek does not break the compute-spending thesis because its efficiency sits within an already extraordinary trend. Ontiveros says that, in two years, a model with GPT-3 quality should cost about 1,200 times less to run inference on. He argues Gemini 2.0 Flash is both cheaper and better than DeepSeek. DeepSeek itself remained inference-capacity constrained, illustrating why cheaper intelligence can still produce demand for much more compute and data-center capacity.
Running out of easily available text data increases the need for compute rather than ending training scale. Synthetic-data generation, reinforcement learning, stronger validation, and using huge models to improve smaller ones all consume substantial compute; Ontiveros cites GPT-4o and Claude 3.5 Sonnet as products of that broader approach. Today’s largest clusters are roughly 100,000 Hopper GPUs and 130 MW, while leading labs are planning gigawatt—and in some cases 2 GW—sites toward 2027.
NVIDIA’s underappreciated inference advantage may be scale-up networking rather than its familiar training-software moat. The GB200 NVL72 connects 72 GPUs in an all-to-all NVLink configuration at unmatched bandwidth, with 3,000-4,000 copper cables in the rack-scale system. Longer-context reasoning models require more memory and networking, leading SemiAnalysis to think NVIDIA’s moat “might actually be stronger in inference than in training.”
The AI upcycle breaks when adoption and monetization fail to support physical investment, not merely when models become more efficient. Ontiveros expects hyperscaler capex to keep beating estimates by 20-30%, but gigawatt-scale sites eventually require more than user growth. The indicators he would watch are product traffic, available indicators of OpenAI revenue and perhaps Claude revenue, and whether consumers keep paying $20 or $200 monthly. At some point, he says, the investment will be too much for the market.
The shortage makes “time to power” the most valuable infrastructure product. Schneider Electric, Eaton, Vertiv, Taiwanese liquid-cooling companies, and smaller engineering or cooling businesses can all participate; much of the equipment is more commoditized than NVIDIA’s GPUs. Behind-the-meter nuclear solutions such as Talen-style arrangements look much trickier than expected.
Bitcoin miners possess scarce powered sites, but converting them into AI campuses is a team-and-capital problem rather than a simple equipment swap. A mining facility costs roughly $0.5 million per MW, versus more than $10 million per MW for an AI data center, and customers want credible operators. Walker’s metaphor captures the optionality—miners are “clown cars that fell into a gold mine”—but Core Scientific and Applied Digital stand apart because they built data-center expertise and industry relationships early.
Miners’ power is not automatically suitable for AI. Walker flags West Texas curtailment agreements and intermittent wind and solar power as potential problems for always-on AI workloads. Ontiveros says generators, on-site batteries, and the UPS systems already used in data centers can cover outages, with battery reserves potentially extended beyond the usual 5-10 minutes.
The United States remains the default location despite cheaper overseas energy because of speed, labor, supply chains, and experience. Roughly 75-80% of hyperscalers’ large-scale self-built data centers were already American, giving the country experience with 100 MW-scale construction. Ontiveros says that no other country “currently knows better how to build large-scale data centers than the United States.”
1. AI has broken the old data-center growth curve
Ontiveros begins by rejecting the word “unlimited,” then quantifies why the enthusiasm exists. Worldwide data-center IT capacity added roughly 4 GW annually from 2019 through 2023; NVIDIA hardware alone generated about 5 GW of incremental demand in 2024, while AI accelerators—including custom ASICs and a small AMD contribution—generated roughly 7-8 GW.
SemiAnalysis models demand by translating every purchased GPU, CPU, or other IT equipment into the power and physical space required to operate it. Since 2024, Ontiveros says, “pretty much all” of the market has been driven by AI accelerators, making generative AI a sharp break from the industry’s previous trajectory.
Its supply-side forecast for 2024, 2025, and several subsequent years says operators are not building quickly enough. The investable implication in Ontiveros’s framing is straightforward: a growing deficit leaves room for companies supplying data-center capacity, electrical systems, cooling, and faster access to power.
2. DeepSeek sits below the existing efficiency curve, not beyond it
Walker’s central pushback is the fiber-overbuild analogy: data usage can compound for decades while investors who build at the peak still lose money. If AI developers begin optimizing power consumption as capability gains slow, today’s Manhattan-scale facilities could become stranded excess capacity.
Ontiveros answers that optimization is already proceeding at a staggering rate. He says that, in two years, a model with GPT-3 quality should cost about 1,200 times less to run inference on; DeepSeek therefore does not change the trend.
His sharper comparison is Gemini 2.0 Flash, which he describes as cheaper to serve and better than DeepSeek. The competitive achievement is real, but not evidence that frontier developers suddenly require less infrastructure than previously expected.
DeepSeek’s own constraints reinforce the point. Ontiveros says it had “zero capacity to serve inference,” turned down new-user requests, operated with a maximum batch size, and had very low interactivity. He also links its CEO’s meeting with China’s number-two political official and the following day’s announced $140 billion of state subsidies to a desire for more compute—not less.
3. Scarce real data makes training more compute-intensive
Today’s largest GPU clusters contain roughly 100,000 Hopper GPUs and consume about 130 MW of IT power. Over the following two to two-and-a-half years, Ontiveros says major labs were planning individual gigawatt-scale sites, with certain projects reaching 2 GW toward 2027.
Walker asks why clusters must grow when “most of the internet” has already entered training corpora. Ontiveros’s short answer is precisely because real data is scarce: labs must spend compute generating synthetic data, using reinforcement learning to build reasoning capabilities, and performing increasingly expensive quality checks.
Walker preserves the key failure mode: heavily synthetic corpora might recursively amplify falsehoods until a model confidently invents an Apple product launch from 1942. Ontiveros does not deny the validation problem; his answer is “scaling law again”—spend more compute to create stronger checks and higher-quality synthetic data.
Larger models also need not become economical consumer products themselves. Labs can use them to fine-tune and improve smaller models, which Ontiveros identifies as part of the path to GPT-4o and Claude 3.5 Sonnet. The frontier model’s value can therefore reside in teaching cheaper models rather than directly serving every query.
4. Untapped video offers another order-of-magnitude data pool
Before companies pay building owners to record the physical world, Ontiveros sees a much larger near-term corpus sitting unused. Models ingest YouTube transcripts and other textual derivatives, but generally do not train on the underlying video itself.
Incorporating movies and YouTube video would provide “orders of magnitude more data” than web-text sources such as Common Crawl. Walker notes that the corpus is skewed—a thousand MrBeast videos do not represent ordinary life—but Ontiveros’s point is scale: multiple years of existing video remain available before labs need schemes to pay people to generate new physical-world data.
User platforms also generate more data for their providers. Every interaction with ChatGPT, Claude, or similar platforms is stored on the provider’s servers, making broad distribution valuable as both a monetization surface and a continuing source of usage data.
5. NVIDIA’s inference moat rests on networking
Walker states the bear case in full: NVIDIA is richly valued, semiconductors are cyclical, and Microsoft, Amazon, Google, Apple, and other large customers can build competitors. They may not need state-of-the-art performance everywhere; an accelerator delivering 90% of the performance or trailing by 18 months could capture a meaningful share.
Ontiveros agrees that NVIDIA’s software advantage may matter less in inference than in training, but says the market underappreciates its overall engineering depth. Hardware and software are familiar layers; high-bandwidth networking is the third and increasingly decisive one.
His best specimen is the GB200 NVL72: 72 GPUs connected all-to-all through NVLink, at bandwidth no hyperscaler solution then matched. The rack-scale design uses “three or four thousand copper cables,” illustrating the systems-engineering burden behind what can look superficially like a collection of interchangeable chips.
Reasoning models and longer context windows require more memory; scaling memory requires a powerful scale-up network. That causal chain leads SemiAnalysis to the contrarian conclusion that NVIDIA’s moat “might actually be stronger in inference than in training,” even if its software differentiation narrows.
6. Adoption and revenue determine when the upcycle ends
Ontiveros’s cyclical stance is blunt: “If it’s an up cycle, things are going to go up, and if it’s a down cycle, things are going to go down.” Rich valuation matters less while earnings estimates keep being revised upward, which is why SemiAnalysis remained constructive on NVIDIA, power infrastructure, and other AI-exposed companies.
He expects hyperscaler capex to keep exceeding forecasts—not by 5%, but potentially 20-30%. DeepSeek does not alter that view because Google already had a model Ontiveros considered better and cheaper; its more immediate effect might be pressure on the high margins charged for state-of-the-art APIs.
What could break the cycle is deteriorating adoption. Ontiveros would monitor ChatGPT, Gemini, and Claude traffic alongside whatever indicators are available for OpenAI revenue and perhaps Claude revenue. The initial rush reflected the prospect of the next billion-user-plus platform; the physical expansion planned for the next two years ultimately needs revenue.
The analogy is Google buying pre-revenue YouTube: hyperscalers historically secure users first and construct monetization later. AI may combine increasingly capable free models such as GPT-4o Mini or Gemini 2.0 Flash with premium tools, advertising, APIs, and $20 or $200 subscriptions—but at some point, gigawatt investments require those experiments to work.
7. Time to power spreads value beyond generators
Ontiveros prefers data-center infrastructure as a capex-driven proxy for GPU deployments. Schneider Electric, Eaton, and Vertiv dominate portions of electrical and cooling equipment, while Taiwanese liquid-cooling companies and smaller local engineering or cooling-tower businesses can benefit because much of the equipment is more commoditized than NVIDIA’s GPUs.
The forecast deficit makes “time to power” especially valuable. Talen-style behind-the-meter nuclear arrangements initially looked like an answer, but Ontiveros says they proved much trickier than expected, forcing developers to consider other solutions.
He also agrees that power producers and natural gas are relevant areas to examine, but his personal focus is the infrastructure required to support GPU demand. The supply-and-demand mismatch can make solutions valuable that would not be economically viable in a normal environment.
8. Bitcoin miners own power, but AI conversion costs twenty times more
Miners historically optimized for abundant stranded electricity in West Texas, Wyoming, or North Dakota. That “power first” history created assets hyperscalers now value: a 500 MW AI project can require roughly $5 billion, while Walker’s example is that buying an existing site could bring it online in six months rather than waiting until 2029 to build it independently.
The physical mining shell is not an AI data center. Ontiveros estimates mining infrastructure at roughly $0.5 million per MW, versus more than $10 million per MW for AI, especially when the miner must outsource much of the work—about a 20-fold step-up before buying the GPUs. Customers consequently need confidence in a miner’s team, technical execution, and ability to deliver.
Walker presses the strongest objection: Core Scientific signed its landmark CoreWeave arrangement months earlier, yet few miners had secured comparable 100 MW-plus commitments despite apparently desperate demand. If the assets were such obvious “gold mines,” knowledgeable insiders should be converting faster and protecting equity rather than issuing shares aggressively.
Ontiveros offers two explanations rather than dismissing the concern. Some people in Bitcoin mining remain almost religiously committed to Bitcoin and expect it to reach $1 million; others lack the senior data-center hires and customer relationships needed for a multibillion-dollar project. “At some point, you’ve got to choose.”
9. Core Scientific and Applied Digital have the clearest execution paths
Core Scientific’s advantage was not merely power. It had an earlier relationship with CoreWeave and experience in Ethereum mining, signed a 16 MW Austin deal with CoreWeave before the flagship arrangement, hired experienced colocation personnel, and built the relationship early. That history helps explain why CoreWeave paid most of the capital expenditures for the conversion. Walker says the rumored earlier offer implied roughly a $1 billion market capitalization, while the eventual contract was worth roughly $1.8 billion in net present value.
Ontiveros calls Applied Digital the credible number two. Its 100 MW data-center capacity was already built at an approximate cost of $1 billion within a broader 600 MW North Dakota site; it had new Macquarie backing and a nonbinding letter of intent that Ontiveros thought was likely to become a roughly 400 MW deal.
IREN is his more speculative candidate because it has 1.4 GW secured in West Texas. Walker flags the area’s curtailment agreements and intermittent wind and solar power as potential problems for always-on AI workloads. Ontiveros says generators, on-site storage, and the UPS batteries already used in data centers can address outages; the usual five-to-ten-minute battery reserve could potentially be extended to 20 or 30 minutes, or whatever is appropriate.
Other examples include TeraWulf’s roughly 70 MW contract and Hut 8’s Louisiana project, which Ontiveros describes as 100 MW in 2025 and, with some uncertainty, expandable to roughly 200 MW by 2026. Ontiveros thinks Riot and Marathon likewise remain committed to mining. Raw megawatts therefore do not substitute for the teams and relationships needed to win AI contracts and access the 15-20-times-EBITDA valuations associated with colocation businesses, sometimes higher.
10. The United States keeps winning on execution speed
Walker wonders why hyperscalers do not chase nearly free natural gas in Qatar or abundant resources in places such as Brazil. Ontiveros’s answer is time to market: roughly 75-80% of hyperscalers’ large-scale self-built data centers were already in the United States, so its supply chain, labor pool, grid power, and experience support 100 MW-scale construction.
Cheap energy cannot substitute for specialized electrical, mechanical, and plumbing labor. Walker also raises the risk that a government could seize a $5 billion data-center-and-GPU investment. Ontiveros says that is a major risk, but argues that even without considering it, labor makes overseas construction difficult.
The leading AI labs, hyperscalers, and surrounding ecosystem are American, reinforcing the domestic concentration. Ontiveros’s conclusion is that no other country “currently knows better how to build large-scale data centers than the United States.”
Full transcript
Before we get started, a quick disclaimer: Nothing on this podcast is investing advice. Please consult a financial adviser and do your own research. That's always true, but we were talking ahead of time about a long list of companies we can talk about today, so we might end up hitting 50 different companies. Please remember that we're not recommending any of them.
I'm super excited to have you on today. You come highly recommended, obviously, and you're over at SemiAnalysis. We started planning this podcast before the DeepSeek stuff even came out, and AI, semiconductors, data centers, and everything around them have been the hottest areas of the market to talk about and debate for the past year. They've gotten even hotter and more interesting over the past 2 weeks after DeepSeek, so we have a lot to talk about.
I'm going to start with my main interest in AI and everything around it. Aside from how I can use it as an investor to improve my work, my main interest is data centers and electrification. I know that's an area you specialize in, so we can get into different companies and industries, but broadly, if I said that most investors I talk to are still working from the thesis that AI equals unlimited demand for data centers—that you should buy nuclear energy and buy anything you can because there's going to be unlimited demand—what would you think about that trend? How are you thinking about it evolving as we sit here? We're recording on February 5, 2025.
Unlimited is a pretty strong word. At SemiAnalysis, we do have a comprehensive tracker of supply-and-demand dynamics. Demand is basically this: Every time someone buys a GPU, or even a CPU—whatever IT equipment they want to put into a data center—there is corresponding demand for power and data center space. That's something we track very closely.
Since 2024, and going forward, everything has been driven by generative AI. NVIDIA hardware and all the custom ASICs are driving the whole market. AMD contributes a little bit, but it's pretty much all AI accelerators.
To give some perspective, the global data center industry from 2019 to 2023 grew by about 4 gigawatts per year. That's the amount of IT capacity added annually. In 2024, NVIDIA alone added 5 gigawatts of demand. AI is roughly 7 to 8 gigawatts, including all the custom ASICs and everything else. It's a complete change in trend, and it's all driven by GenAI.
That's the demand side. On the supply side, which we spend a lot of time tracking, the question is whether people are building data centers quickly enough to power AI and provide the space and power where they can place their GPUs. The answer, based on our forecasts for 2024, 2025, and a few years beyond, is no. They need to build data centers faster, which means upside for the companies exposed to that supply chain.
Let me start with a basic question. DeepSeek highlighted this concern for me. I worry as an investor about the power companies people are buying—Talen, Vistra, Constellation Energy, and all these other companies that have become investor darlings because of what seems to be unlimited demand. Talen, in particular, has a great nuclear asset.
Right now, there's an AI race, and these companies don't have to optimize for anything. They're just trying to go, go, go and get the best model. But I worry that if we run this out 12 or 24 months, at some point you have to start optimizing. The gains get smaller, and you start optimizing. It seems like a natural first place to optimize would be to say, "We're literally consuming Manhattan-level power in one data center. How hard would it be to design one that uses a little less power?"
You could get into a really overbuilt scenario. We've seen it before with fiber. Data demand went straight up for 20 years—I think it was growing at about 30% from 2000 to 2020—but if you built fiber in 2020, there was a huge oversupply. I worry that we're building all these data centers, not to mention bringing all this new energy online, and then 12 months from now we start optimizing and realize we way overbuilt.
That's a totally fair point. I would start by highlighting that the progress on the efficiency of inference for AI models is tremendous. Everyone is freaking out about DeepSeek, but in reality, DeepSeek isn't changing the trend. If anything, DeepSeek is below the trend.
The trend is that, in 2 years, a model with the quality of GPT-3 will cost about 1,200 times less to run inference on. The amount of improvement in the AI space is gigantic. Something we also like to highlight is that everyone is freaking out about DeepSeek, but if you look at other competing models, Gemini 2.0 Flash is much cheaper to run inference on than DeepSeek and is actually a better model.
People are freaking out about DeepSeek, but it's nothing out of the ordinary. It's very good, to be clear, but it's not fundamentally better than the best models from Google, OpenAI, and others. The trend of massive improvement was always the case.
People talk about the Jevons paradox, but the trend that matters here is the overall scaling law. You want to have the best model, and we've seen tremendous improvement in the capabilities of those models. Maybe you've tried Deep Research recently; the things it can do are getting pretty insane. You've surely seen the benchmarks for o3, including the ARC-AGI-1 benchmark. Those are pretty incredible as well.
The point is that we're going to keep building better and better models. That means more demand for training, but even more demand for inference. There's also a funny thing to note: Despite all its efficiency, DeepSeek has zero capacity to serve inference. It's extremely constrained.
DeepSeek has 2 problems. On the training side, its CEO met with the No. 2 person in the Chinese Communist Party, and the next day China announced $140 billion of state subsidies. It's pretty clear that DeepSeek said, "I need subsidies to build more compute." On the inference side, it can't serve new users. It has to turn down new user requests, and it operates with a maximum batch size and very low interactivity. Its user interface is not very good because it has to maximize the capacity of its GPUs.
The trend is that even if you're building the most efficient model—like people think DeepSeek is doing—you still want much more compute, many more data centers, and much more capacity overall.
I guess I'm just a journalist. Every time I email you or someone like you and ask what you think about something, the DeepSeek news, for example, everybody freaks out. It was about 2 weeks ago that every journalist freaked out, and I'm sure that every specialist knew about it 2 months before it filtered through to the general public.
When did you first hear about DeepSeek?
DeepSeek actually started popping up about a month ago.
If I were to talk to you and ask whether you're at all worried about scaling laws at this point, or what worries you about the entire AI trade, what would be on your mind?
The way we think about it is that we split it into 2 phases: training and inference.
If you think about training alone, over the next 2 years all the big AI labs still have massive plans. A nice way to visualize that is that today the largest GPU clusters are about 100,000 Hopper GPUs. In terms of power, that's roughly 130 megawatts of IT power. All these companies want to scale that massively, and they have plans for gigawatt-scale data centers, including 2-gigawatt-scale facilities for a few specific sites, toward 2027.
In 2 to 2½ years, they want to scale their biggest single cluster from 100,000 GPUs and 130 megawatts to gigawatt scale. That's massive scaling, and that's for training alone. We're talking about tens of billions of dollars in capital expenditures, which they partially have the funds to finance at this point.
At some point, though, they're going to have to face adoption of the underlying technology and generate inference revenue.
Can I pause you there and ask a really stupid question? I've always wondered about this. You mentioned that we're scaling from roughly 100,000-GPU clusters to gigawatt-scale clusters. I keep hearing that we're near the end of the amount of data available to train these models on.
If I told you that, why do we need to go from 10,000 to 100,000, or whatever the numbers are, when there's only so much more data we can put into the models? Yes, the world generates more data, and we can go find books from the 1500s, but most of the data has already been input into these training models. Most of the internet is in them. Why do we need to go from 10,000 to 100,000 when it seems like the amount of new data we can add is getting smaller? Does that question make sense?
The answer is exactly that there's not enough real data, so we have to use a lot of compute to generate synthetic data. When you go from 10,000 to 100,000, a lot of that additional compute is used to generate synthetic data and train on it.
It's not only synthetic data. Techniques for adding more reasoning capabilities also use reinforcement learning, which uses a lot of synthetic data. Another interesting area is pre-training scaling laws—building models with more parameters and more data. One issue with pre-training scaling laws is that at some point the models become too expensive to run inference on.
You can use those huge models to fine-tune smaller models and make the smaller models much better. That's a technique all the AI labs have used. That's how they came up with models like GPT-4o and Claude 3.5 Sonnet.
That's why I have experts on, so I can ask super-generous questions. You read about this stuff all day, while I read about a lot of other things.
If you start having models that run more and more on synthetic data instead of real-world data, don't you risk the synthetic data becoming corrupted? I understand that they have checks, but you can't check everything in a model. If 90% of the data is synthetic and 10% is real, don't you risk ending up with models that believe the sky is red or the ocean is green?
Those would be obvious things to correct, but we're all familiar with the phenomenon where you ask Google, "What was Apple's big product launch in 1942?" and it tells you Apple was excited to launch the Catapult that year. You wonder where that came from, and no one can point to it. Doesn't synthetic data create an increasing risk of hidden, phantom answers?
The simple answer is the scaling law again: You just need to spend more compute to have better checks and make sure the synthetic data is of better quality.
Let me ask another question about data. Most of the data right now is stuff that's online. Do you foresee a rush to generate more data tokens, where companies say, "Facebook has cameras and interactions everywhere. Google has cars driving around. Forget that—I'm going to put a camera on every corner and in every building so I can record everything happening in the world"?
Maybe ChatGPT could say, "We really need a lot of real-time physical data, so we're going to go to every building owner and say, 'We'll pay you $1,000 per month if you let us record everything and collect the audio and video.'" Is that something you could see happening?
You could argue that using platforms like ChatGPT, Claude, or whatever is one way to contribute more data to those companies. Every time you have an interaction, it is stored on their servers. Having a platform with many users is very valuable because it generates much more data.
I get that. One argument for Facebook is that it has all this private data—the interactions, direct messages, and everything you like on Instagram. Facebook has more data on you than anyone else, which is great for advertising.
I was talking more about bringing it into the real world. I could imagine a world where ChatGPT says, "We really need a lot of real-time physical data, so we're going to go out and pay people to record everything happening in the world." That would generate enormous amounts of additional data in some way, shape, or form.
There's a step before that, which is starting to use all the video data that's already out there.
You mean YouTube data and things like that?
That's right—all the movies and all the YouTube videos. We don't really use that data currently. We use transcripts and things like that, but we don't actually use the video data itself.
Using all of YouTube would be orders of magnitude more data than what people use from internet text data, like Common Crawl. That would be the first step. We need to use the existing video data, which is massively larger than what exists in text format.
YouTube is interesting because you and I have completely different YouTube feeds, and you'd never even realize it. There are millions and millions of hours of video uploaded to YouTube every day, but it isn't really being used for training yet.
I would also argue that if you started analyzing 1,000 MrBeast videos for data, that would be interesting, but it's very different from me sitting in my closet interviewing you, or petting my dog on the street. It would be a very skewed view of the world.
Imagine all those great podcasts that YouTube could train on. It would learn a lot, especially about investing. There's just a tremendous amount of data out there in video. Again, it's orders of magnitude more than what exists in text format, so before we get to your scenario of paying people to generate data, there are multiple years of video data that we're yet to uncover.
I want to ask about some more data center questions, but let me completely switch gears for a moment.
Matt Levine called it the most market-moving short report ever—the short report on NVIDIA that came out the weekend of DeepSeek. People think it was partly the reason NVIDIA opened down 10% or 15%, losing $600 billion of market capitalization.
The basics of the report were that NVIDIA has had great success, but success attracts competitors. For various reasons, NVIDIA's future is not as bright as the stock market is giving it credit for. One reason was that all the big technology companies and hyperscalers—Amazon, Apple, and Google—are trying to build NVIDIA competitors.
As we move from training to inference, people aren't sure NVIDIA has such a big advantage over its competitors. There are all kinds of reasons you can list. I'd love to hear your thoughts on NVIDIA.
I'm in a privileged position because I'm part of the SemiAnalysis team. We're 23 people with a lot of technical backgrounds and very different areas of expertise. We have an AI engineer who knows everything about model architectures, networking experts, and so on.
We answer these questions from a technologically informed point of view. What we see when analyzing the engineering of all these systems and comparing it with common market narratives is that NVIDIA's engineering talent tends to be underappreciated.
People know that NVIDIA is a great hardware designer. That's not a secret. People increasingly understand that it has a great software moat, although there is a discussion about whether that software moat applies to inference.
The third area, which I think is less often discussed, is networking. Our position at SemiAnalysis is that NVIDIA's moat might actually be stronger in inference than in training. Even if we agree with the market position that the software aspect is less important for inference than for training, everything related to networking—especially the scale-up network—is more important.
Think about the engineering process behind building the NVIDIA GB200 NVL72, where you have 72 GPUs interconnected in an all-to-all configuration with NVLink at incredible bandwidth. That bandwidth is unmatched by any hyperscaler. The amount of engineering involved is pretty incredible.
If you look at the rack-scale solution, you have 3,000 or 4,000 copper cables. It's a state-of-the-art system, and no one else has a solution as good as that. That's a huge moat for inference.
As we move into more reasoning models, especially when reasoning models have a longer context window, they need more memory. To scale memory, you need that scale-up network, and that scale-up network is one of NVIDIA's fortes.
Let me pause there. I have no view on the NVIDIA stock, so I'll be honest about that. It seems like a great company that's growing rapidly, but it is priced pretty richly, and semiconductors overall are very cyclical. Every semiconductor company has a cycle at some point.
One point in that short report was that nobody is questioning whether NVIDIA is a great company. It's just priced very richly for what is ultimately a cyclical industry. The second point was that Microsoft, Amazon, Apple, and Google are not stupid. NVIDIA is earning huge margins, so why couldn't they recreate 90% of NVIDIA's performance?
Why do they need 100% of state-of-the-art performance all the time for every AI model? Why can't it be 90%? If I gave Microsoft's engineers the NVIDIA technology from a year ago, couldn't they recreate it within 6 months? Why isn't Microsoft creating its own stuff 18 months behind NVIDIA, with half the cost of the capital expenditures, and then using NVIDIA's state-of-the-art technology for the other half?
I know I'm talking about hypothetical worlds, but the underlying question is this: Nobody disputes NVIDIA's moat, but the stock is quite richly priced. Why can't there be a world where NVIDIA is still a great company but the stock is overvalued?
That's why investing is hard. I think my friend Byrne Hobart from The Diff said about 2 years ago that the way you'll know AI is here is when NVIDIA's stock starts going up and all your smartest friends who are into AI start buying calls.
I think about you and the SemiAnalysis team all the time because Doug said that demand for power is unlimited and that you should just keep buying. But at some point, where does it end? Where's the limit?
I used to be a buy-side guy. You can invest however you want, but generally you want to be with the cycle. If it's an up cycle, things are going to go up, and if it's a down cycle, things are going to go down.
Even if a company is richly valued during a strong up cycle, nobody cares. It keeps going up because earnings estimates keep getting revised upward. That's why we generally have a positive stance on NVIDIA and many other AI-related stocks, including power companies. The earnings are going to rise so much more than consensus expects.
I would briefly talk about hyperscaler capital expenditures. You've seen Google's capex, Meta's capex, and Microsoft's capex. Microsoft hasn't given full-year guidance, but at the end of calendar 2025, when we look back at what people project today versus what they actually did, I think the answer will be that capex went up massively.
It keeps beating estimates—not by 5%, but by 20% or 30%. That's how we see the trend playing out in this up cycle: very strong upward revisions in what people are spending.
Those are words of wisdom. If we're in an up cycle, numbers are going up and you can buy everything. But nobody has a crystal ball. When does the down cycle start?
I do think there were people who thought that Project Stargate would be announced, with $500 billion from Masayoshi Son, Oracle, and Trump, and that everything was going straight up. People saw it as a huge check and a silly number, and said, "This is your license to buy anything."
Then DeepSeek came out, and a lot of people got worried. My first question is about the timing. Stargate was announced after Sam Altman had committed to an $80 billion check, and Mark Zuckerberg had talked about spending $50 billion or more in capex. I would imagine they knew about DeepSeek before making those commitments, because they continued to see upside.
Would you agree or disagree?
I fully agree. As I said, Gemini 2.0 Flash is actually better and cheaper than DeepSeek, so DeepSeek doesn't change the trend. It might have an impact on margins because the companies want to charge high API prices for state-of-the-art models, but overall it doesn't change the trend.
You just said we're in an up cycle, the numbers are going up, and you can buy everything while the numbers are going up. What breaks that?
A couple of things could break it. First, it's very important to carefully track adoption of AI. You can do serious work, like looking at website visits for ChatGPT, Gemini, Claude, and whatever else. You can also use whatever tools are available to track OpenAI's revenue and perhaps Claude's revenue. The overall adoption obviously matters.
I think the reason this has taken on such proportions is that we've somehow seen the AI work. Hyperscalers built their businesses around users first, not revenue first. You had acquisitions like YouTube in 2007 or 2008, where people said it was way overpriced because YouTube had no revenue. In Google's mind, it was a great platform with great users. They wanted the users first and would figure out the revenue model later.
WhatsApp was a similar story. You could argue it was less successful, but Instagram has obviously been extremely successful. Hyperscalers are user-first and revenue-second, and it turned out to work pretty well for them.
If you apply that logic to AI, it doesn't work perfectly because AI is more capital-intensive. But in terms of users, ChatGPT demonstrated 200 million users in about 2 months. It's pretty clear that GenAI has the capability to become the next billion-user-plus platform.
That's why everybody is in such a rush. A billion-user-plus platform means tens of billions of dollars in potential revenue, perhaps even hundreds of billions. The consumer adoption of these tools is extremely strong, and having hundreds of millions or billions of users was enough to support the spending level we're seeing today.
However, what we forecast will happen over the next 2 years—and there is meaningful physical evidence of it, especially in data centers—is that they're building huge gigawatt-scale data centers, many gigawatt-scale data centers dedicated to AI. At some point, to justify those investments, you're going to need more than users. You're going to need revenue.
Tracking whatever indicators you can to understand adoption is crucial. If people stop using ChatGPT, or if these companies struggle to sell $20-per-month or $200-per-month subscriptions, at some point that will simply be too much for the market.
I just paid for my first ChatGPT subscription. I started using it more, and I realized I needed to commit to it. If you're a professional in almost any intellectual field—including investing, although sometimes I say yes and sometimes I say no—you probably need to be using it constantly or you're falling behind pretty rapidly.
But is a subscription ultimately how you pay for these things? Bloomberg is a subscription product, although it's also a networking tool. Most of these tools, including Google Search, ultimately monetize through advertising. I'd be surprised if subscriptions are how they all monetize, especially if data is the way they get better.
One way to generate proprietary data is to have more searches and more usage go through your platform. How do you get the most searches? You make the product free and monetize it through advertising.
I would say that the business models people are exploring right now are generally based on having a free model with decent capabilities. That would be GPT-4o Mini or Gemini 2.0 Flash.
Those free models will see their capabilities increase. You can think of the free version as an advertising tool and a way to build a massive user base—hundreds of millions, perhaps billions of users. On top of that, once users see the capabilities, you can sell all kinds of additional products. Advertising could definitely be part of it.
As someone with your finger on the pulse of this more than 99.9% of the population, what AI tool do you use the most?
I like Gemini.
Have you tried Deep Research?
It's pretty incredible. If you want to do a quick stock initiation report, it's fairly decent. There are some good prompts you can use with it.
One of the reasons I paid for ChatGPT is that over the past couple of weeks I was researching things, started playing around with the prompts, and used the reasoning function. Some of the outputs I got would have taken me a month to put together. It was really incredible. I decided that, as an investor, I had to use it.
For me personally, the 2 areas that are most interesting are power and data centers. Power can mean the actual power producers—Talen, Vistra, Constellation Energy—or it can mean whether we're going to have to fund all this additional demand with nuclear, natural gas, or even coal. All of those are interesting.
The other area is the rush for data centers, which brings us to the Bitcoin miners in particular. Let's start with power. We talked earlier about the demand for power and all the construction that's happening. When you think about the power side, what are the most interesting plays? Is it what everybody likes to do and just buy the power producers? Or is there a way to play the underlying commodities?
We have such a supply constraint that we need all this additional power. We can't retire coal plants, or we might need more natural gas. How do you think about playing that?
I personally look at data center infrastructure, which is a pretty good proxy for GPU demand. It's a capex-driven type of business, so everything related to electrical equipment and cabling equipment is interesting. There are big companies like Schneider Electric, Vertiv, and Eaton.
If you want to be creative, you can look at the Taiwanese stock market. There are some liquid-cooling stocks with fairly high exposure to liquid cooling, so there are interesting plays there. There are also smaller companies with exposure to MEP engineering for data centers or cooling towers for data centers.
There's a fairly large supply chain. The thing is, those pieces of equipment are generally commodities. They don't have the level of product differentiation that NVIDIA has with GPUs. Even if Schneider dominates the electrical and cooling market, and Vertiv and Eaton are major players, there's room for smaller companies to play in local ecosystems. Everyone can benefit in that area.
That's one area because I personally look at equipment, but you're right to also mention power producers and natural gas.
Vertiv is what I was going to say.
Exactly. I think an area you want to look into is time-to-power solutions. In our forecast, we have an extremely granular view of the supply-and-demand dynamics of the global AI and data center market. We forecast that the deficit will increase in the coming years despite the data center boom.
That means if you want to solve the problem, you need time-to-power solutions. For a while, people thought behind-the-meter nuclear, like the Talen-type deals, would be the solution. It's actually much trickier than expected, so you have to look at other solutions.
This supply-and-demand mismatch is creating value that would not be economically viable without the mismatch. That obviously affects the miners because they're not traditional data center companies. They might have a few people from Digital Realty or Equinix, but they don't have the full experience of building these solutions.
In a normal environment, people would never go to those miners, especially for a 500-megawatt project, because it's a huge amount of capital. We're talking about a $5 billion investment. But in this environment of supply-demand mismatch, the miners have historically been power-first.
They found stranded power available in large quantities at extremely low prices, which is why they went to West Texas, Wyoming, North Dakota, and similar locations. The miners have the power and can provide time-to-power solutions. Suddenly, they're a very valuable asset to hyperscalers and other companies deploying GPUs.
Let me ask about the Bitcoin miners. For people who don't know, mining Bitcoin is similar to AI in that you take some GPUs or Bitcoin miners, put them in a facility, run a lot of power through them, and get Bitcoin. It's a very commoditized and difficult business.
A lot of these companies woke up and said, "There are all these AI companies desperate for places with cooling, lots of power, and huge facilities where they can put GPUs. We're Bitcoin miners. We can take the Bitcoin miners out, put AI GPUs in, and suddenly we have an AI data center."
Those facilities are getting valued at huge multiples. Core Scientific did this last summer. It struck a deal with CoreWeave, and the stock went up about 4 times in a year because it went from being a mediocre Bitcoin miner to being AI data infrastructure.
That was Core Scientific, but there are tons of Bitcoin miners out there. I know some of them well and others less well. I'd love to talk about the Bitcoin miners as potential AI plays.
Before getting into specific miners, I want to add something to the previous question. It's all about time-to-power. We forecast a deficit that will increase in the coming years despite the data center boom. That means you need time-to-power solutions.
For a while, people thought behind-the-meter nuclear, like the Talen deals, would be the answer. It's much trickier than expected, so you have to consider other solutions.
The value created by this supply-demand mismatch affects the miners. They're mediocre miners, and they don't have the data center expertise. They might have a few people from companies like Digital Realty and Equinix, but they don't have the full experience of building those solutions.
In a normal environment, nobody would go to those miners for a 500-megawatt project, because it requires a huge amount of capital—around a $5 billion investment. But in this environment, the miners have historically been power-first. They found stranded power available in large quantities at very low prices, which is why they went to West Texas, Wyoming, and North Dakota.
The miners know how to secure power and can provide time-to-power solutions. Suddenly, they're valuable assets to hyperscalers and other companies deploying GPUs.
I've called them clown cars that fell into a gold mine. They built these huge facilities because they thought Bitcoin was going to a million dollars, often without much regard for cost, economics, or competitive analysis. Then the economics came down, and suddenly people realized, "You have 300 megawatts of power. If we build that ourselves, we might not be online until 2029. We could buy your facility, throw the Bitcoin miners away, and be online in 6 months."
If you go back roughly a year, remember that CoreWeave wanted to buy Core Scientific. They were effectively saying, "I'll buy this company and get roughly a gigawatt of available power, then develop it myself." It was a cheap way to secure that power.
The rumor is that CoreWeave offered to buy Core Scientific, but the contract they ultimately struck was worth twice what CoreWeave had offered. CoreWeave might have offered a purchase price implying a $1 billion market capitalization, while the contract was worth roughly $1.8 billion in net present value. Core Scientific kept the upside from Bitcoin mining on its other assets as well.
It was strange. If CoreWeave had offered $1.8 billion, maybe it could have had a deal. Instead, they signed a contract worth more than the purchase price.
That's essentially what happened. There are a few other companies we've looked into, including fuel cells, which we can discuss later if you want.
Tell me about one Bitcoin miner you're most bullish on as an AI player. As a Bitcoin mining business, I understand that it isn't very good, but as an AI play, which one do you think will successfully make the leap?
It's all about how much power you have available and how much uncovered upside remains. The easy answer would be Core Scientific, but they've already made the leap.
The most speculative one, but one that I think makes sense, is IREN. They have 1.4 gigawatts secured in West Texas. It remains to be seen exactly how they manage that, but just having that amount of power is valuable.
The downside is that, based on IREN's earnings calls, the management team seems hesitant about how much they want to move into AI versus Bitcoin mining. They still seem very committed to developing the Bitcoin mining business. I think that's also true of other companies like Riot and Marathon. In my personal opinion, they should go all-in on AI because the valuation of a colocation business is much higher.
You're talking about 15 to 20 times EBITDA for companies like Digital Realty, and sometimes more than 20 times EBITDA. You're talking about 15 years of revenue visibility and the same amount of revenue, but with a valuation of 20 times EBITDA.
Let me prove that I've done the work over the past few months. You mentioned IREN. There was a Culper Research short report in July 2024 that I thought was very good on the economics of Bitcoin mining.
I think one of the issues with IREN is that, although it has 1.4 gigawatts of power, a huge amount, all of the West Texas miners had to sign curtailment agreements with the government to be allowed to build. If the retail load is too high, they have to shut down and support the grid.
A lot of that power is also intermittent, coming from wind and solar. When you put all of that together, and when you think about the Bitcoin miners that say they're going to become AI plays, I think intermittent power is a major issue. You can't use intermittent power for AI, whether it's curtailment, solar, wind, or whatever else, because AI needs to be on 100% of the time.
You're going to spend $100 million a year on power and put $5 billion worth of GPUs in the facility. If you're offline for 5 minutes, the cost of the GPUs completely overwhelms the power cost. Given those issues, do you think IREN can actually become an AI data center, or does it have certain assets that can overcome them?
Generally, the issue isn't that difficult to overcome. Most data centers in the world have backup generators, which are meant to cover periods when you don't have grid power. If there's a grid power failure, the backup generators maintain power.
You can build more sophisticated systems. Some people have started deploying on-site battery systems to manage those peaks. Another thing to keep in mind is that when you build a data center, it's not just about the diesel generators. There are also batteries inside all data centers as part of the UPS system.
Typically, people have 5 to 10 minutes of battery capacity, which is just enough time for the generators to turn on. You could imagine going further if it makes sense—20 minutes, 30 minutes, or whatever is appropriate. Overall, the issue can be overcome fairly easily.
Let me push back in a different way. The Core Scientific-CoreWeave deal was announced in June 2024, and the Talen-Amazon deal was announced in March 2024. If you're a Bitcoin miner, you've had 6 to 9 months to try to make the AI transition.
These AI companies are desperate for power and facilities. Six to 9 months isn't a long time in the grand scheme of switching a company over, but it is a long time when there's a gold rush. We still haven't seen anyone other than Core Scientific sign a definitive, 100-megawatt-plus AI contract.
IREN has experimented with buying GPUs and doing AI, and Applied Digital has talked about doing the same. But if you choose any of the other miners, you're betting that they can make the transition. Why haven't we seen another miner land a major contract if any of these companies are going to make that leap?
TeraWulf did sign a contract. I think it was 70 megawatts, so it isn't insignificant.
Applied Digital is probably the closest. They have a real data center that's already built in North Dakota. It's a 600-megawatt site, and 100 megawatts of data center capacity is already built.
It's not a mining data center; it's a real data center. They paid roughly $10 million per megawatt for it, so they paid about $1 billion for the data center. It's essentially ready. They'll need more capital to build the other data centers, which are supposedly covered by their nonbinding letter of intent. They recently got capital backing from Macquarie, so Applied Digital is a credible number two behind Core Scientific.
I think Applied Digital is likely to get its 400-megawatt deal finalized. Why have those companies succeeded while others have struggled? I would argue that it's about the technical capabilities within the team.
Both Applied Digital and Core Scientific already had experience in the colocation business. Before the flagship deal, Core Scientific signed a 16-megawatt deal with CoreWeave in Austin, Texas. I don't remember whether that was in 2023 or early 2024, but it was early 2024.
The point is that those companies saw the trend earlier than others, hired people accordingly, and already had decent industry relationships. By "hired," I mean they hired senior people from the colocation companies. They built good teams, which is ultimately what matters. It's a human business.
If you look at a company like IREN, it needs to hire more technical expertise. It probably needs to be more focused and streamlined about exactly what it wants to do. In its last earnings call, the company said it wasn't sure whether it wanted to focus on Bitcoin mining or AI data centers. At some point, you have to choose.
This is the other thing that makes me want to smash my head against the wall. I'm not blaming you; I'm ranting, and I'd love to hear your response because I think there's an interesting question here.
Any of these miners say, "We're not sure whether to focus on Bitcoin mining or AI." If you're a Bitcoin miner, you're going to be valued at $100 or $200 per kilowatt, or whatever the number is. If you're an AI data center, look at the deals people are signing. Even $1,000 per kilowatt would be low.
There's a gold rush. You've got insiders who should know the opportunity best, but they're not enthusiastic about making the transition, and they're not protecting their equity. A lot of these companies have issued shares at an incredible pace. Some of them probably increased their diluted share count by 4 times in 2024, and I'm not exaggerating.
If making the switch to AI could make the stock go from being worth 200 to 2,000 overnight—not financial advice, just talking about the magnitude—why are the people who should see the gold mine not seeing it? Are we wrong, or are they doing something we don't understand?
That's a fair point. I would start with something that might be controversial. Many people involved in Bitcoin mining, and in Bitcoin generally, have a religious way of thinking about it. I don't know whether that's the right word, but think of the laser eyes.
Some of them said as recently as 6 months ago, "We're never going to do AI. We're Bitcoin miners to the core." The AI numbers became so large that they started saying they might consider allocating some of their assets to AI. To their credit, they are coming around.
I agree with what you're saying. I personally like Bitcoin, so I think they're right about several things in their overall pitch. If you're in their mindset, you see Bitcoin going to $1 million, so Bitcoin mining is the business to be in.
At these current prices, if you assume Bitcoin goes from $100,000 to $1 million overnight, these companies would be value-neutral between Bitcoin mining and switching to AI. My view is that they should make the switch to AI and then buy Bitcoin in their personal accounts, but that isn't how they see it.
Another issue is that people are generally bullish on AI, but there is still some skepticism. There's also a lot of fear. The capital required to build an AI data center is much higher than what these companies are used to building.
When you build a Bitcoin data center, I'm talking only about the physical infrastructure, not the hardware, it typically costs about $500,000 per megawatt. When you build an AI data center, especially if you don't have much internal capability and have to outsource a lot of the work, you're talking about more than $10 million per megawatt.
That's a 20-times factor. These companies are used to building a Bitcoin mining facility for $500,000 per megawatt, and now they need to build an AI data center for more than $10 million per megawatt.
But you don't have to do it on spec. CoreWeave paid for most of the capital expenditures for Core Scientific. If these assets are so good, why can't the miners find someone to fund the construction?
NVIDIA could write a check to Applied Digital or Core Scientific. You could find somebody to write a check and cover the capex. Why haven't we seen another proof point yet?
I would go back to the industry relationships. I think Core Scientific and CoreWeave had a partnership for several years. Core Scientific was involved in Ethereum mining a few years ago, so they knew each other well.
I'm sure they had internal discussions before doing the deal, including negotiations over how much capital CoreWeave would contribute. Companies like IREN, Marathon, and others probably aren't as involved in the data center market. They may be hiring now, but they don't yet have the industry relationships.
You need to become familiar with the industry. I personally go to a lot of data center conferences. I was at PTC a few weeks ago, which is a pretty large conference. I met people from Applied Digital and Core Scientific, and Hut 8 was there as well. They had a credible project.
A lot of the newer companies weren't at those events. They're new to the industry, so they need to become familiar with it. That's work Core Scientific started much earlier. You need credibility and relationships.
This has been a super interesting conversation. Just to wrap up the miners: I've done a decent amount of work on Applied Digital. I really like the North Dakota asset, although I'm not a super expert on it.
It sounds like you think it's a pretty solid asset.
I think it's a pretty solid asset. They've done a lot of dilution, but when you start talking about a 600-megawatt asset, you realize how much it could be worth if all 600 megawatts were applied to AI. It's a big number.
It sounds like you like the IREN asset as well. Any other miners you want to touch on? We're coming up on an hour, so I don't want to drag this on forever.
Very briefly, IREN is the more speculative one. It has a lot of power, but I'm not sure about the human capital and whether the company has the right people to execute.
Hut 8 has an interesting project in Louisiana. It has an interesting data center that is 100 megawatts in 2025 and expandable to, I think, 200 megawatts by 2026. It already has a solid site plan and secured power, so I think that's a pretty solid one as well.
Let me ask you a completely left-field question. Most of the data center activity is happening domestically in the United States. Forget Europe, which has power issues, but I'm always surprised that we haven't seen a company say, "Qatar has basically free natural gas. We're going to build a kajillion-dollar, jillion-megawatt plant there."
You could say the same thing about Brazil, which has huge energy resources. I'm surprised by how domestically focused the AI data center buildout has been. I'm sure I'm missing a few things, but why do you think it's been so focused on the United States?
You're 100% right. I would say it's mostly about time to market.
An interesting way to visualize it is to look at where the large-scale data centers are today. Let's go back to 2022 or 2023. Where were the large-scale data centers? Most of them were self-built by U.S. hyperscalers, and 75% to 80% of those large-scale data centers were in the United States.
Large-scale data centers in the United States are nothing new. The country has the experience of building 100-megawatt-scale data centers. It has the supply chain, a decent amount of labor, and a decent amount of grid power. The United States is simply a good place to build data centers.
One risk I've heard is that you build a data center, spend $1 billion on it, put $4 billion worth of GPUs in it, and then the government shows up with machine guns and says, "Thanks for the $5 billion investment. We'll be taking that now."
That's a big risk, but even without considering that, labor is an issue. You might want to build in the Middle East, but getting specialized labor isn't easy. Low-level labor is available, but specialized electrical, mechanical, and plumbing labor is more difficult to find.
You hear the horror stories from Qatar when it hosted the World Cup and what it had to do with labor to build all those stadiums. That shows how difficult it can be to build overseas.
The data center infrastructure has specific requirements, and no other country in the world currently knows better how to build large-scale data centers than the United States. The big AI labs, hyperscalers, and the surrounding ecosystem are all American, so it's all happening in the United States.
How can people find you if they want to get in touch or learn more?
I have a very underfollowed Twitter account with almost no posts. LinkedIn is another option.
The best way to become familiar with our work is to check out SemiAnalysis. The website has incredibly deep content, and the other people on the team are amazing as well. The best way to reach out would probably be through SemiAnalysis, Twitter, or LinkedIn—whatever works for you.
This was a ton of fun. The industry is developing so quickly that we could probably do one of these every week and find a way to talk about something new. We probably won't do one every week, but we'll have to have you back for a follow-up conversation.
Thanks so much for coming on.
Thanks for having me.
A final disclaimer: Nothing on this podcast should be considered investment advice. Guests or hosts may have positions in any of the stocks mentioned during the podcast. Please do your own work and consult a financial adviser.