Dylan Patel: NVIDIA's New Moat & Why China is "Semiconductor Pilled”

Hello, how are you?

It's very clear we're not sure where AI models are headed in terms of over the next few years, what happens to the architecture. But the thing that I think everyone has sort of agreed on is models are pretty autoregressive, right? Next-token generation is the thing. But beyond that, attention mechanisms change, how it works, everything could change. And so what's interesting is the reason NVIDIA won is because they just took the widest-surface-area bet, and then people kept developing models on that, and that kind of shape worked.

But now the workload is so large that there is room for specialization that will give you 10x increases in certain domains, right? In a general-purpose workload, Groq doesn't work, right? It can't train, it can't inference really, really large models cost-efficiently, right? You can't serve many, many, many users. But what it can do is it can go screamingly fast, right? Same with the Cerebras-OpenAI deal. But that's one workload, right?

Very decode-focused, right? Doing autoregressive tokens in a single stream super fast. Another direction AI models could head: we don't know, are models going to think in one token stream, or is it actually that they're constantly context-switching, right? And they have this humongous, humongous context, and they're generating in multiple parallel streams, right? And so Google and OpenAI have both released mechanisms of this with their Pro models, where the model actually doesn't just have one single chain of thought for reasoning, it has multiple, right?

And then I don't know exactly how they choose which one and what the final answer they deliver to you is—it's an area of research. But there is room for that kind of chip, right? Something that works on a lot of parallel streams of chain of thought. And maybe the latency requirements are not as crazy, right? Maybe you don't want to go blindingly fast. Maybe you're okay with it being, because I can spin up 100 parallel streams of thought or agents or whatever you want to call them.

Maybe I care a lot about cost there. And because it's 100 in parallel instead of one going super, super fast, it's not as deep, right? The tree search or the depth of the inference is not as deep, but it is much wider. There's other parts of inference, the prefill process, creating the KV cache. So NVIDIA has a chip for that, right? That's the CPX. So they've made the CPX, they bought Groq for decode, and then they still have their general-purpose GPU.

So they're kind of trying to cover their bases because unlike the first wave of AI chip companies, where they sort of just made chips and then tried to figure out where it would work, they had a thesis. Groq and Cerebras both, as well as SambaNova, put a lot of memory on the chip. And not necessarily in the case of Cerebras and Groq—no memory off-chip—and in the case of SambaNova, less memory off-chip or slower memory off-chip with higher capacity.

They sort of all made similar bets in that direction. And it didn't work for a while until it kind of did, right? Because there is a workload that now necessitates it. NVIDIA recognizes they're the leader, they're the tentpole. In one respect, they can just run faster than everyone, but it's kind of hard to be 2x better than Google or OpenAI or whoever else's internal chip, right, to justify their 75%-plus margins.

And then they have to be 2x to 4x better to justify their margins, because that's what they're charging above COGS. The question is, what architecture will deliver that? Well, yes, keeping the programmability of their GPUs is great for training and for a lot of workloads. But guess what? I think a lot of people will just be downloading an open-source model, downloading an inference framework, and pressing go, right? A little bit more complicated than that, but that's going to be the consumption method for a lot of enterprises, a lot of startups, a lot of tech companies: they're just going to do that or they're going to rent the GPUs.

Or rent the chips and then download an open-source framework and model and go, right? And NVIDIA recognizes this and there is room for products that aren't general-purpose, right? Creating the KV cache, maybe those workloads could be different chips, right? And the CPX chip they announced, they say it's for the context processing, creating KV cache. It's also really useful for video models, because video models don't care about memory bandwidth. And so why pay for the expensive memory that the general-purpose chip has?

Or why do what Groq is doing, which is tying hundreds or thousands of chips together and not having memory, but keeping the entire model on-chip? The trade-off for that, of course, is you need thousands of chips and you have less compute per chip. And so NVIDIA is trying to capture the whole surface area because, again, you don't know where models are headed and it's hard to say where the research is headed.

I certainly think it's not good from an anti-competitive sense, right? I don't think people should just be able to buy companies without any antitrust process at all. Now, in the case of a large company buying a startup, I'm completely fine with it. The flip side is, we know the deal is happening, right? This happened for a company I was an advisor for. NVIDIA acquired Enfabrica just maybe a few months before they did Groq, and it was a similar style of deal, right?

If someone wanted to strike it down, that's the biggest limbo, right? We've seen this happen in venture, and you probably know more stories of this, but a company trying to get acquired, they get stuck in limbo for like a year.

And then it falls apart.

Yeah, it falls apart. The deal died because of some regulatory BS. And now the company and the founders were focused on getting the deal done instead of making the product better for a year. And now they're behind or they weren't focused on growth as much, right? You only have so much time as a founder. So in that sense, I like the license deals, right?

I think the thing about NVIDIA is they take the Andy Grove mentality more seriously than anyone else, right? Like, okay, fine, Google implemented OKRs because Intel did it, but that's like management stuff, right? Only the Paranoid Survive, right? This is like core to the Bay Area, core to NVIDIA. Jensen is very paranoid about losing, right? These specializations, if you just kept making his mainline chip, would mean people could make point solutions for specific parts of the market that would crush him on cost and performance, and then he can't justify his margin.

That's a threat to NVIDIA's business model as a whole, especially if the best model only changes every three months, or the model you want to roll out. Okay, well then you can have three months to figure out how to make a model work on one chip architecture for that point solution. And it's fine. NVIDIA's software advantage is not that important then. Jensen's super paranoid about losing, and frankly, it's really hard to hire enough talented chip people. When you look across the market, there are only a few companies who have successfully created a chip architecture, software to run the models accurately, right?

Like, because you can look at random APIs of, say, an Alibaba Qwen model, and different people are doing all sorts of tricks like quantizing it, but also many other tricks which then end up making the model quality lower, building a rack-scale solution, networking thousands of chips together, and then deploying an API. And Groq did the whole thing with, frankly, not that many people. So now it's like, okay, well, I'm NVIDIA. I want to make four different chip architectures and actually four different point solutions, maybe the general-purpose, and then one here, one here, one here.

And in addition, my general-purpose thing is actually not just like a GPU chip. It's like GPU chips, CPU chips, networking chips, NVSwitch, NICs. There's many, many chips, and each of those chips has many chiplets. You don't have enough engineering resources, right? And so acquiring Groq is how you get those resources to make more solutions for different parts of the market. And as far as, like, are they threatened? Obviously there's some cool startups out there that are raising a lot currently, or have raised, such as Etched, MatX, Positron, these new-age AI companies.

There's also the prior age of, like, Cerebras is out there still, right? Tenstorrent, et cetera. And so there's a lot of AI chip companies on the startup side, but then there's also Google's TPU, AMD GPUs, Amazon Trainium, who are all really credible competitors. And then Meta's MTIA is somewhat credible. And then Microsoft's Maia is not credible, but maybe it will be one day, right? So you sort of have a lot of competition.

They've got to hold the gates back. And so I think there's risk from all of those companies that I mentioned and, effectively, California slash Seattle, right? Only two places. There's also chips from other parts of the world, right? Obviously China has a number of different AI chip companies that are doing cool things. Anyone would have told you Groq was—their business revenue, their revenue was not, like, stellar, right?

In fact, they missed revenue last year significantly, and yet they got bought, right? Because the value of the IP was there and the value of the team. Anyone else would have been like, well, why the heck would I buy this? Right? Makes no sense. There's definitely a credible threat. Yeah.

I think they do. I think networking is super important. I think the CUDA software moat is very important, but it's also changing rapidly, right? An incredible amount of the software that NVIDIA GPUs run on is not from NVIDIA. It's the developer ecosystem that's open-sourcing it. When you look at, for example, vLLM and SGLang, right? These support AMD GPUs almost as first-class citizens now. And vLLM is getting significant support for TPUs, for Trainium, and there will be other chips coming out from startups that also support vLLM and SGLang.

Now, how difficult is it? The reason why CUDA is so important is like, okay, I can do whatever I need to do, right? Programming a GPU. I think most AI chips will not be consumed by people programming anything for it.

They will download an open-source inference engine, they will download an open-source model, and then they will put it on the— and it's really simple to download vLLM and make it work. It's not that hard to set up a server. And NVIDIA's putting out a lot of open-source software like Triton Inference Server and Dynamo and all these things to make it easy, because that is the consumption model ultimately for the majority of AI, right? It might be like, oh, it's my own inference engine, but most servers will not run code besides the inference engine and the model.

It's not like researchers are writing code for GPUs to see if ideas will work and train models and all these things, or just mess around with them to figure out performance or whatever it is, but most of it won't be there. And so CUDA as a moat, CUDA language is fine, right? No one actually writes CUDA, right? They compile and then they just run it on the GPU.

They don't write CUDA. But a lot of this CUDA moat is like, how does PyTorch translate into high-performance GPUs? And that surface area from when people were hardcore writing CUDA kernels to, hey, they're writing PyTorch and then it's compiling down to GPUs versus, oh, I'm just downloading vLLM. It is a curve of, like, not a ton of people that can do CUDA kernels. A whole lot more people can do PyTorch, right?

Random PhDs and random people. It's very simple, right? A crapload of people can do vLLM, download it, run it on a server. Well, if it now supports other chips, what is the CUDA moat? NVIDIA's recognized this, and they've been building software that is not necessarily the CUDA moat. And I can give some examples, right? So the name of the game is fast tokens and lowest-cost tokens, right? And lowest-cost tokens happens by your chip being fast.

But there's also tricks, right? One example, like I mentioned with the CPX versus Groq, is processing your prefill context, right? Super-cheap CPX, right? If I care a lot about speed, then Groq. These are optimizations on the hardware side. There's optimizations on the software side as well, right? And so one example is when I'm doing, for example, if I look at a Claude Code or a Cursor-type application, right? The workload is, like, it takes your repo, it takes the relevant parts of your repo, puts it in the context of the LLM, it prompts, it generates, right?

And if it's in agent mode, it cycles the context a couple of times. It'll collapse, put things off to the side, access different contexts. But what's especially interesting when you think about an agent for software, and you can see this in Codex, Codex actually is not as good as Claude Code, but it can do work on time horizons of like nine, ten hours and do a big refactor better than Claude Code can, even though most of the time Claude Code is better.

And what's interesting about Codex is it'll take your repo, it'll identify parts if you're asking it to refactor it, identify parts, write stuff, make these notes for itself everywhere, collapse the context, switch from this part of the repo to that part of the repo to this part of the repo. But when you think about it, it's like, oh, if this thing is just generating tokens all the time, plus it's switching what my context is constantly, that's really expensive, right?

If you think about what's the cost of inference, I want to say it's like, it's $10 per million tokens of output, $10 for decode, and $3 for prefill. And so if you think about, oh, it just worked for nine hours on one task, one refactor, huge value. But if it changed context a ton of times and your context is like 30K usually, or 50K, or heading to hundreds of thousands, how big your repository is and how much context switching.

Now you're spending all this money on prefill, right? Not the decode tokens. But actually, why am I regenerating the KV cache? I can actually just store the KV cache elsewhere, and then when I need it again, I can pull it and plop it into CPU memory or into GPU memory. And so NVIDIA's got this KV cache manager, and they've been working really hard on making it so they can interface SSDs and stick the KV cache on there and pull it out whenever they want.

So for this kind of workload, and then if you do this and you look at coding as an application, and you look at these coding companies and how much they're paying for prefill versus decode, actually the majority of their cost is prefill tokens, not decode tokens, because the context is just so large and it's switching all the time, even in agent modes. If you can now not have to do the prefill, your costs go down dramatically. But that's a very complicated thing to do from a software perspective.

Companies like Anthropic, Google, OpenAI have already done it, but what about the wide world, right? And so NVIDIA's trying to make the open-source software for this, and that's like CUDA moat, but it's like, actually, no, none of this is CUDA, right? It's like memory management and storage management and when do you call what and how do you transfer it and how do you spread the KV cache across a bunch of different storage nodes and what happens when you read it and the network congestion, just all these things.

Yeah, it's like NVIDIA's wheelhouse, but it's not CUDA. And I think the easy way to say it is, it is the CUDA moat, right?

And so things like this KV cache manager and many other things they're trying to do to reduce the cost of inference is how they build the new CUDA moat. Because again, today it is quite—I mean, AMD's not fully there yet and TPUs are being added right now and Trainium's being added soon as well to vLLM. But all of them will have a very good UX to download a model, run a model on vLLM by the middle of the year, I think, right?

Certainly, AMD is already there by the end of this quarter. We have something that tests this, right? AI. It's open source. All the code is, and the results are. But we run across, I think, $60 million of GPUs, which are donated to us by companies like NVIDIA, AMD, OpenAI, Microsoft, Amazon, Crusoe, CoreWeave, Together AI. All these companies are sponsoring GPUs for us to run this. We're running vLLM and SGLang every night on nine different kinds of GPUs on a variety of different models and different context lengths and all these things, right?

To see the performance. And you can see the performance moving every day, or pretty often, because the software changes all the time. And so the fact that this exists is the key, though, right? It's not that AMD, you can do this on their chips, NVIDIA, you can do this on their chips. It's, oh, when the new model comes out, how fast does it get to peak performance? Because it's a moving target. Or, hey, can I implement this KV cache management thing?

How hard is it? How many engineers do I need? Oh, just one? Great. Or 10, great. If I need 100 people to develop it like Google and so on and so forth did, then that's much harder.

I think AMD will be caught up at times and very behind at other times. Currently, they're super far behind, right? Because Blackwell is just way better than MI355. And then Rubin comes out and they'll be way, way behind, but then AMD's new chip comes out and AMD will be caught up or even slightly ahead on a hardware perspective. Software's behind, right? And you have this leapfrogging, and AMD is a very credible second competitor. I don't think they'll go beyond—I think they'll stay in single-digit market share, single-digit percentage market share.

Single-digit percentage market share is still pretty good.

I mean, NVIDIA's revenue this year is going to be, like—it's a lot.

I think it's actually $4 gazillion.

You sort of have the whole specialization game, right? You have to specialize because you're never going to beat NVIDIA at their own game, right? They're going to have the supply chain on lock. They're going to get to the newest memory technology or process technology or whatever packaging technology, whatever it is, sooner than you. And they're just going to crush you, right, if you play their game. AMD is trying to play NVIDIA's game, but AMD is, like, extremely good at engineering silicon, right?

Everyone else has to, has to, has to try something weird or different, right? And so when you look at Etched or MatX or Positron or Cerebras or Tenstorrent, you've got to look at all these companies, right? There are unique things about what they're doing, and it's not clear if AI models will still be within that realm when that comes out, right? Does—oh, now people use, like, n-grams and other sparse attention techniques. Does that change some of the specializations people are doing?

Or, hey, people are now doing, like, models are now sparse MoEs instead of being dense models. Does that change things? There's so many optimizations and changes on the model side, and you can't predict what's going to happen with the ML research easily. At least you can't. The thing you're optimizing for today has to be a vision of where AI will be in two years. And NVIDIA's fully accepted they don't know where that's going to be. That's why they have a portfolio of chips now, not just one GPU line, right?

It's not just Hopper, Blackwell, Rubin. Now it's going to be—it's not Ampere, Hopper—it's not that line. It's like there is a variety of chips to serve the different markets in different possible scenarios. They think each of them has this vision today, but, oh, it might turn out the general-purpose one sucks, and actually AI models have developed in a way where CPX or Groq-style chips are the best, right? Well, okay, now we have a solution for that market.

And so I think that's the challenge with the startups. With that said, I think they're all taking very interesting bets. I think it's much more exciting than the first wave of AI hardware bets: Graphcore, Cerebras, SambaNova, Groq, where they all made the same bet on memory and putting the memory on the chip. They sort of just made a bet and they optimized for a certain kind of model, all similar kinds of model, and it didn't end up working out for a long time, right?

They had to pivot and they had to work on a lot of things, and it took a long time. I think these companies have, like, a really clear vision of what they think models will look like, right? Etched does, MatX does, Positron does. And that's what's really cool about it between the three of them, these new-age—so, I mean, I'm excited for them. I'm very skeptical. I don't know what a venture capitalist views as likely chances of succeeding, but I think all of them are less than 1%, right?

It could, it could. Or it could be any given customer has, like, one workload they care a lot about. Anthropic clearly does not give a crap about video gen, image gen, right? They just don't care. On the flip side, a company like Midjourney cares a lot about image and video gen, right? Image and video gen is very, very—like I mentioned, it's a very—like, it's not very memory-bandwidth-heavy. It loves, loves, loves compute, right? Whereas inference of large language models in the style of, say, for example, coding agents cares a lot about decoding for long streams of time.

And that's very memory-bandwidth-heavy, right? And so that's, like, a simple example, but there's a lot more nuance there in terms of even the size of the matrix multiply units, the tensor cores, the systolic arrays that you use, or the ratios of networking and memory, and what's that memory hierarchy look like, and what are you doing for different kinds of attention, and all these sorts of things. There's a lot of specialization here. And so some people are betting big on different types of specialization.

And I think you could clearly see a world where companies do care about different stuff, right? Like, if, for example, a chip optimized for video and image generation existed today and it was better than NVIDIA, or NVIDIA made it, I think Midjourney would absolutely only use that for inference. I think for training they'd still use the general-purpose thing. And Meta and Google should do that, right? And hey, Meta actually has two lines of AI chips.

Their MTIA, there's a line that's focused on recommendation systems. And then there's a line that's focused on GenAI. The GenAI one is a new line, but that recommendation systems line is still continuing, right? It's not sexy. No one cares because there's no—and ByteDance also has a recommendation system line of chips, and it's not really focused on GenAI, which is fine because this is a $200 billion business or something, which is just deciding what ad to serve me, right?

And what order to put my friends' stories and things like this. So I think it's perfectly fine for there to be specialized AI chips, given the target market is big enough. And you have to have vision to know what that target market is. Unless you're a hyperscaler, then you can just use general purpose until it's clearly there, and then you can make your ASIC, right?

I think it's a variety of things, actually. In some quarters last year, it was even north of 20%, I think, but I don't remember exactly. But anyways, if you look at 2022, China was almost the size of the U.S. in terms of buying server hardware, right? Almost. Not quite, but getting there. And it looked like they were going to be the same size as America in like a year or two after that, right? And if you look at global data center capacity, global cloud capacity, et cetera, et cetera, et cetera, it's American companies and Chinese companies, right, that dominate the world.

American companies obviously doing a lot better here, but both of those dominate the world. And if you look at every industry, right, it's very clear that China wants to insource stuff, right? So in 2015, they made these five-year plans for 2020 and 2025 where they set the percentage of semiconductors they wanted domestically produced. And they've missed the goal both times, which is fine, right? They set really aggressive goals and even, shoot for the moon. Even if you miss, you hit the stars, right?

Yep. And that's sort of what's happened, right? Look, China is not caught up on leading-edge semiconductors, but microcontrollers from China are almost as good as the microcontrollers—are as good and cheaper than the ones from Texas Instruments or STMicro, et cetera, right? Or this random power chip is better than or the same as the one from another company, right? And so they've really built up a semiconductor industry and started insourcing a lot more.

I don't see why China wouldn't be buying 30, 40% of the world's AI chips and the U.S. like 50, 60%, and then the rest of the world. And when I say U.S., I mean U.S.-origin companies. That seems like a more natural state for the world. But there are restrictions and, hey, this is the biggest change in human history, maybe ever, knowledge work and everything that's going to happen there. And then eventually robotics and all these things. Obviously, there's a lot of geopolitical stuff.

And so there are restrictions. NVIDIA has been handicapped from selling their best chips to China. And so that's obviously impacted the sales a lot because, like, why would you do that? And so when you look at who rents the most GPUs in the world, it's three companies, right? So one of them is obviously OpenAI. Second one, actually, they were bigger than OpenAI. They are bigger than OpenAI today. Or no, they were bigger than OpenAI, then OpenAI eclipsed them recently, is ByteDance.

ByteDance rents tons of chips from Oracle and Google and many other cloud companies because they couldn't get the chips they needed in China. They're mostly just serving TikTok, right? Okay, well, they're not allowed to buy them, and that sucks, but they're allowed to rent them. And so, okay, if I'm not allowed to get the best ones, I'm going to rent externally. And if ByteDance is the second biggest renter of GPUs in the world, that's substituting demand that would've been built in China in many cases.

It's instead being built in Malaysia. And Oracle has over a gigawatt of capacity in Malaysia that ByteDance is gonna take, right? So things like this are hundreds of thousands, if not millions of chips, tens of billions of dollars of capacity that would go to China, but it's not. It's going to Malaysia instead, as an example. Another sort of point around this is China's had these five-year plans. And the way these initiatives work from China is there is some top-down ordering, but then they just kind of whip the whole—everyone just kind of gets into it, and it's really cool.

I don't think it's as top-down as many people think. I think the entire country is semiconductor-pilled, right? There are dramas where people fall in love in the fab, or dramas where people fall in love and they're photovoltaic, like solar-cell researchers and engineers. And it's like, this is just the backdrop. And actually, this is super cool for your significant other to be that semiconductor engineer or to be that photovoltaic solar-panel researcher.

As opposed to an influencer, right? I'm sorry, Love Island is—I watched for like 10 minutes because I was forced to. I was like, this is freaking terrible. But—

No, seriously, we're cooked. We're cooked. And so I think when you think about this happening, it's diffused into drama even. And people—there's multiple dramas taking place about the semiconductor industry, and they're romance, comedy, like the entire spectrum, right? Drama, right? Like, it's like, what the heck is going on? Anyways, you have all these provinces, you have all these local cities setting out ordinances and giving out subsidies and all sorts of stuff, right?

It's truly crazy. There's some national-level stuff, like, oh, no taxes on this. Oh, we're going to ban a few things. But as far as I understand, the national government has not banned NVIDIA's H20 or H200, but the local ones have, right? A lot of local ones have said, no, you must use China-manufactured chips. And it's like, who told you that? You're here to uphold this. It's like, doesn't matter, right? I mean, it's cool because then you have this survival of the fittest.

All these provinces and cities are trying to attract different companies with different types of subsidies and grants and industrial parks and all these different things. And then the ones who succeed actually develop an industry and they take over.

I mean, it's actually great. There's this one TikTok—or not TikTok—TikTok and Instagram person, and they sing it. They're like, if you want to buy things in China, make sure you go to the right place. And then they just say the most random shit and name the city. And then you look into it and you're like, wow, this city has the entire supply chain for this. And it's like lampshades, and then it names a city.

It's like, what the fuck? There's a city that specializes in lampshades. And it's like microphone arms, like microphones. Literally, there's a city in China that specializes in—

It's literally everything. Literally everything. There's a city, and it's not like, hey, specifically for camera arms, for example, there's ball bearings in this, and the ball bearings are like—there's multiple manufacturers of ball bearings for camera arms.

It's like, what the hell is going on? And so, like, the semiconductor industry, I think people don't realize, is absurdly specialized. I'm not answering your question. I'm just going on a little bit of a rant because I think people don't understand China semiconductors. It's really sick. Or semiconductors in general. But, like, fascinating. Like, in Japan, they focus on a few different types of chemicals, and they're the best at it. And it's, like, almost a cultural thing, right?

Like, Japanese people are so precise, like with sushi, and it's all about the trade and the craft. And, like, the French food in Japan is better than the French food in France because the Japanese chefs went there and then came back, and they perfected it in Japan because they're so precise. And there's so many different things that Japan is so good at because they're so precise and dedicated to the craft. And it comes out of, like, I don't know, samurai culture or something.

I don't know, right? Like, I don't exactly know how that culture came up. And so when you look at, like—and it's across the world—there's different places where things like this happen, right? Like, oh, the Netherlands makes EUV tools. Cool. I guess so. And you look across the semiconductor industry, there's a famous economic essay called "I, Pencil" or something like that, talking about how the pencil, like a simple pencil, comes from, like, oh, the rubber comes from Indonesia for the eraser, and the graphite comes from this mine here, and the wood comes from these aspen trees in Canada.

And, like, you actually can't make a pencil without aggregating this entire supply chain. The semiconductor industry is way crazier because, like, I would say there's, like, 15 or 20 countries that could shut down the entire semiconductor industry, right? Even, like, Austria could, right? And it's like, what? It's like, well, yeah, there's two different companies there who have, like, 90% share in some random niche stuff. And it's like, okay, cool, I guess Austria can. And, oh yeah, those two companies only have less than a billion of revenue, but they just happen to have linchpin critical things.

And there's linchpin critical things everywhere because the process is so complicated. And so China's been trying to replicate this.

I think there's a lot of things. I think if you were to close your eyes and say—or if you were to cut off every country and say there's no more globalism—China has the most vertical stack in semiconductors today, and they're the best at semiconductors in the world. Because their fabs could still run somewhat on a lot of things because they have built some of these chemical supply chains, right? Like, TSMC, for certain kinds of chemicals, has 100% share from Japan, right?

Or Intel, same thing, right? Or, for certain kinds of tools, 100% share from the Netherlands, or 100% share from this American company or that Austrian company, or this or that, right? Like, there's just all these different places that have 100% share. It might be one company, might be three companies, but geographically or in the same area. And China has built that up, right? Because they've created these Made in China initiatives which just plowed money into it.

And they've got this culture of, like, the diffused—these provinces are like, yeah, I just decided I'm going to fucking focus on—or it might not even be the city, right? It may be, like, someone brought it there and decided, and then people are like, oh wow, you're doing that? Me too. Like, I'm a Patel and I grew up in a motel. And guess what? Almost all the Patels I know grew up in a motel.

And it's because some random Patel immigrated to America and worked at a hotel, motel, and then bought a motel. And then it just started happening, right? Like, these things are serendipitous of sorts. And, like, I view it as the same kind of specialization, right? Chinese cities are starting to do these things. China's missing a lot of things, right? I would say, like, if you say minus-10-year tech, China's complete and no one else is complete, right?

Taiwan is not complete. The fabs would shut down without foreign supply. And you go down, or you go across the stack. But if you go to 10-year tech, maybe more like 20-year tech, you could get a fully vertical supply chain in China, which I do not think any country could do. Like, America could not build a fully vertical fab without stuff from elsewhere, even if it's 20-year-old tech.

Probably not even 40-year-old tech. And so that's interesting. But then the flip side is, like, well, you kind of do need specialization. That's how that chemical gets the purest, best, most engineered—or that slurry of chemicals, or that gas, or that tool. Because every smart person, or a lot of them, in that country grew up around that culture, and the supply chain is there, and everyone kind of knows, and it's a drive away, and sort of, like, this is what makes supply chains work.

Is that there is this specialization, and the best of the best only comes when you have that hyper-specialization. So China doesn't have lithography. Their lithography is, like, 10 years behind, and I think it'll be five years behind in a couple of years, right? They're catching up fast. I don't think they'll be as good as ASML for a long time. Maybe, I don't know, maybe they will be. China—you shouldn't ever underestimate China. But, like, and Chinese engineers are—but, like, for a while, right?

Or, like, I don't think they'll be able to make leading-edge chemicals like many Japanese companies or many American companies, and their tools. And you just go across the supply chain. They're not at the forefront on really anything in the manufacturing supply chain. On the design supply chain, there's some things that they're starting to be at similar par, but, like, cheaper, or, like, a year or two behind but cheaper, and that's fine for a lot of stuff. An example of that is Huawei, right?

Huawei in mobile phones was on par with Apple, like entirely, and they had become Apple and TSMC's biggest customer when they were designing the best thing, and they are number one in telecom, and their tech is just literally better. And so when you think what happens, is China missing anything? It's like they don't have the best investments today in the AI supply chain. They have a complete package and are a couple of years behind, and they'll figure out how to make it cheaper, slash do more, slash catch up, and create a robust industry.

But there's a reason, like, I don't think that Jensen is scared of AMD, really. He's paranoid. I mentioned he's paranoid. I'm sure he's a little bit scared of them, right? Like, I think some of the things that they've done are reactions and competitive dynamics with AMD or Google's TPUs or whatever, right? There's a CoreWeave deal today, and I think that's directly the result of what Google's been doing.

Yeah, NVIDIA invested $2 billion in CoreWeave. But what's more important is that that's sort of just the sticker. What's really relevant is NVIDIA is going to work with CoreWeave to acquire and backstop and all these things: the land, the power, the energy, the transmission, help build the data center, all this capital-side stuff that, because NVIDIA has so much money, they can backstop CoreWeave doing it, because CoreWeave then can be the one who generates demand. Anyways, there's, like—because Google was doing that, and they did that with, like, a couple companies.

Such as Fluidstack and TeraWulf and Cipher. These are some public deals that have been announced. And so Google is doing that with TPUs, and NVIDIA reacted, right?

And so, in the same way, I think NVIDIA's reacted to AMD. And in the same way, I think the thing is NVIDIA is deathly terrified of Huawei.

Because Huawei has caught up to Apple and actually surpassed them as TSMC's biggest customer before they got banned, right? They did just crush Nokia, Sony, Sony Ericsson, et cetera, right? Like, the entire telecom supply chain, they just completely destroyed them. And there's so many other areas. Like, they straight up made a folding phone, right? I have a Samsung folding phone. They have a folding phone that's better than Samsung's folding phone.

And it's like, bro, what? Huawei's really, really cracked. And so, of course, they're terrified of Huawei. Huawei is the most vertical company in the world. No company is more vertically integrated than Huawei, which then leads to huge innovations.

But not just phones, security cameras, actually. I think they have, like—

Exactly. Exactly. I think Huawei is terrifying, right? And so, yes, their chips are not as good today.

Huawei shipped a little bit, but mostly just, like, starter capacity. Like, there's nothing like—no, no. I would say a little bit as in, like, a few servers, not like a billion dollars' worth of stuff, right? The thing is, China's supply chain has to ramp up, right? China's express goal is to have it all internalized, but then a company like Alibaba is like, "I don't want to use Huawei, right? I want to use NVIDIA and just make the best freaking models, right?"

Because that's my business. My business is not using a Huawei thing, but it's like, okay, it's being pushed upon me. There's other companies too, like Cambricon and so on and so forth. And so this sort of supply-chain—companies in China don't want to use Huawei, but they're kind of encouraged, obviously, or pushed: you must. Some local provincial government would be like, "Well, you're doing this much business here, you got to do this," right?

Like, there's all sorts of crazy stuff pushing companies to use Huawei. The challenge is Huawei can't manufacture enough, right? We've done a lot of work on this, and we've just put it out for free instead of to our customers because it's something that's, like, national security: how is Huawei actually building chips? Well, actually, they were using shell companies to get chips from TSMC and using different methods of sneaking HBM, which is memory, from Korea through Taiwan to China, right?

Like, all sorts of crazy stuff we've reported on, and people—it's like a whack-a-mole, right? They shut it down, or tools get shipped to China and they shouldn't be for making leading-edge chips, but they actually are. And all these sorts of things are happening because they can't make everything. And if they want to make the leading-edge stuff, they do need to rely on the foreign supply chain quite a bit in terms of the upstream supply chain, right? Memory, logic chips, tools for fabs, chemicals for fabs, et cetera.

Huawei cannot satisfy the market because there's not enough advanced leading-edge capacity in memory, logic, and all these other things domestically in China. And they're trying to build it as fast as they can, but that means there's just not enough to satisfy the market. And so NVIDIA has a market. I think they'll figure out how to sell chips to China. And Jensen's in China, I think, like, right now, or was yesterday, and so he's clearly wheeling and dealing to try and get his chips into China because I think NVIDIA's argument is, if we sell them chips, then there won't be as much of a domestic market. The feedback loop for software and everything else won't be there. That was sort of really challenging, right? Most of the open-source software for AI has a lot of Chinese contributors, right? VLM, PyTorch, SGLang, and all of these other libraries and things. And it goes to low-level software especially, right?

A lot of the best open-source stuff is actually just from a Chinese company who decided to open-source it. And same with models, right? And so it's like, okay, well, if they can't use NVIDIA chips anymore, then this open-source stuff won't be designed for NVIDIA chips; it'll be designed for Huawei chips. And now, does that weaken the CUDA moat? And now, not only is China domestic, now they have a feedback loop internally, and then they can externalize across the rest of the world, right?

So this is the argument NVIDIA makes. I'm not sure if I am. I think my AI timelines are so fast. I'm not that fast, not in terms of AGI, but hey, AI is $100 billion of revenue across the industry. I think the industry could hit $100 billion ARR by the end of this year, like $45, $50 for OpenAI, like $35, $40 for Anthropic. And then Vertex, DeepMind's models at Google, Gemini, right?

And then Vertex API for Anthropic models and Bedrock APIs and Azure AI Foundry APIs. I think $100 billion by the end of this year. That's a lot. And then what's the economic value of that $100 billion? Now, how much of that is in China, right? China's number is probably 10x lower, right? Because they just haven't been able to pervasively push AI, right? ChatGPT has a billion users, roughly. And then you add on Gemini, and Meta claims they have 500 million users.

I don't know. I think people just accidentally click generative sticker or something. But anyways, there's a lot of usage of AI in the West already, and it's going to climb. It's going to keep climbing, and you kind of have to get used to it. And so the question is, what's the economic benefit to the world? Right. And at the end of the day, this is an economic war, right? If the U.S. and the West win in AI and control more powerful AI systems that have this feedback loop that improve economic growth and weapons systems and whatever else, engineering of grids and cyberattacks and all these sorts of things, they have this advantage over China, then China will not rise to be the global hegemony.

But without AI, China definitely will rise to be the global hegemony. They're just gonna outrun America. And so the question is, that's, I think, the other view, right? And how fast are super powerful AI systems versus China building a domestic ecosystem for chips and models and everything that is a few years behind? What's actually the value, right? That's sort of, like, around restrictions and regulations.

I think TSMC is manufacturing wafers, and they're building real wafers, and there's real fabs, and there's some other fabs that have been announced, and they're doing well. And there's a bunch of different kinds of plants, like a Korean company making a random gas plant in Texas for their chips, right? For chips. And all these sorts of things are happening. I think the CHIPS Act did really well with its $50 billion. It's just, I don't think people understand the scale of the semiconductor industry.

It is the most complicated supply chain in the world, right? It's much bigger than, say, manufacturing airplanes. It's much bigger than really anything else, right? If you look at the top 10 companies of the world, I think eight of them design semiconductors. Right now, obviously, Google designs semiconductors, but it's like, oh, wait, no, their cost of search would be like 10x higher if they didn't have TPUs. And TPUs are super optimized for search, right?

Or you go down the list, right? Meta serves recommendation systems with their chips, right? You go down the list, everyone is making their own chips. Apple devices would be materially worse if they didn't have their own chips.

Right? And you just go down the list. It's the most complicated supply chain, and they're spending something on the order of, like, $150 billion, roughly, in subsidies a year to the chip industry. We are doing $50 billion over, like, a decade.

There's a difference in scale here, right? The collective total amount of CapEx that has been spent in Taiwan is like $500 billion-plus, right? Across the industry, across all the companies that are making semiconductors in Taiwan. And Taiwan doesn't have a domestic industry. How is $50 billion of subsidies going to change America's needle, right? It does move it a little bit, right? I want to be clear, the CHIPS Act is awesome. I don't understand why EVs or solar was given this massive, massive trillion-dollar package.

Semiconductors were only given $50 billion. Semiconductors need a lot bigger package to actually incentivize onshoring. I think what's happened so far has proven that it's working well. TSMC is literally making chips for NVIDIA and Apple and AMD and others in Arizona today. And I think that's really great.

Well, the CHIPS Act only passed because automotive prices went up, because car manufacturers are the worst because they do just-in-time inventory, right? Or not worst, but it is just a thing, right? Just-in-time inventory systems. COVID happens, sales plummet, fabs that were making random power ICs or random microcontrollers for engines got repurposed to the boom from COVID, which was data centers and PCs and smartphones. So that stuff was booming. And then when people were like, oh, wait, actually, I have some money.

I stayed at home. I didn't go out. I didn't drink. I have some cash, right? Let me buy a car. They went out and bought cars, and cars started skyrocketing in prices. Oh, yeah. Can you sell me that microcontroller for the engine again? It's like, no, I'm making a slightly different microcontroller that works for, let's say, a keyboard or a mouse, right?

Or whatever. And it's like, they actually did just leave me flat-footed, and they weren't, like, a partner through COVID, right? Versus, you just left me. Screw you, Ford or whoever, Toyota or automotive OEM, that supply chain. And so the CHIPS Act only got passed because that happened. And people were like, oh my God, the semiconductors are why cars can't be made. If that didn't happen, we wouldn't even have the CHIPS Act.

It's silly. Whereas that's what was pitched to all the senators. I know people who were running around Capitol Hill just pushing that narrative and story, and that's why it finally got passed. In reality, it was all for advanced, leading-edge chips, right? Nothing that goes in a car, right? And so it's like this funny thing.

I'm very optimistic.

I thought we just needed a bigger CHIPS Act, but look, Trump's kind of gotten TSMC to promise to invest a fuckload more. And they're moving on it, right? They're actually just building it. It's like, "I'm going to tariff the shit out of you unless you build a fab." And it's like, "We'll build a fab." And they're building it right now. The timelines for fabs just take forever because, again, it's the most complicated thing in the world.

The cleanest place in the world is not a hospital or a biotech lab or whatever. It's a semiconductor fab. And the most expensive tools in the world are not any of these medical tools or whatever. It's semiconductor tools. Or it's not a rocket. It's a semiconductor tool, right? I remember when I was a kid, I was like, "I want to be a rocket scientist." And then I was like, "Oh, I want to be a surgeon."

And I'm like, wait, chips are like rocket surgery, but even cooler, right? Anyways, there are fabs being built in America. They won't take America to self-sufficiency. I don't think that's a relevant goal, right? Globalism is generally just good. Hot take, in terms of economics.

Dude, you're going to get me canceled. I tweeted about ICE. It was a complete joke, but so many people got mad at me because I'm too much of a joker. These are serious things. Yeah.

Yes. Anyways, I think we are building fabs, and I think it's going to move. And now even Elon's talking about building fabs because he sees the shortages in the world, right? There's a lot of semiconductor-related shortages for building out AI. And so I don't think it's hopeless. I'm very optimistic that we're going to do more and more and more. And maybe this administration threatens tariffs and they get the deals, and the next administration comes back with the carrot.

If it is the Democrats, whatever happens, I don't know. I was at a comedy club on Sunday night, and he's like, "Oh, I use ChatGPT." And there were a couple of people who booed, and he's like, "Yeah, I'm one of those guys, I know." And it's like, wow, people hate AI.

Or New Jersey power prices are up, right?

Is it because of a data center? Well, New Jersey, the governor's election—I think literally there was an election that changed recently in New Jersey because power prices were up, and people blamed a Microsoft-Nebius data center in New Jersey for that reason. But in reality, that data center has nothing to do with power prices going up. It's Superstorm Sandy, five years ago, or whatever, how many years ago, knocking down the state's electrical infrastructure, and then all these improvements.

And then those improvements have to be paid by someone. And it turns out the consumer has to pay for them with higher power prices, right? And so there's a lot going on in that regard, right? That kind of is sad. And people hate AI, and they're blaming AI on it. And artists hate AI. And you see all this deepfake stuff, and I think it'll be the hot-button issue, especially as we're really getting into—I think last year Google spent $3 billion on Waymo, and we're waiting for their guide for this year.

$3 billion on Waymo taxis. But their Waymos went from like $300K to like $100K or $90K, the new Waymo car. And they're going to spend more than $3 billion because they've just launched in like four cities now.

Right? Or five cities. And they're testing it a lot. And the same with robotaxis. People are going to hate AI for that reason. People are going to hate AI because of the slop on the internet. People are going to hate AI because of the perceived job replacement. People are going to hate AI for all these reasons. And so, yeah, it's going to be a hot-button political issue, don't you think?

I'm obviously a maxi. I think we're going to need a lot of infra. And I think I'm literally paid to analyze the supply chain and do consulting. That's what my company does. So obviously, I'm very biased. I think we're pretty good at calling when things go down, though, before a part of the supply chain, whatever. But anyways, again, going back to the economics of it, it's north of $100 billion of revenue exiting this year for AI from a base of sub-$1 billion.

From a base—because ads and stuff is already a multi-hundred-billion-dollar AI industry, right? Go back to 2023, it was less than a billion. In 2024, I don't know exactly what number, maybe let's call it 10. And '25 was maybe 30, 40. It'll be north of 100 easily. If you're talking about $100 billion of revenue, let's say at a 50% gross margin, so that's $50 billion of gross profit and $50 billion of COGS. That $50 billion of COGS needs to run on infra, which costs roughly, if you're talking about five-year depreciation, call it $250 billion, right?

For $100 billion of revenue. Okay, what is the actual spend on AI? And for this year, it's going to be like—I mean, it depends on what layer. If you're talking about energy, those are longer-lived assets and all these other things, right? Data centers are longer-lived assets. The chips are not as much. People are putting CapEx down, and the hyperscalers' CapEx is going to be like $500 billion this year or something like this. And then besides them, there's also a lot more CapEx elsewhere.

And so, is it a bubble? I mean, theoretically, it's twice as much as it should be, but it's also like, well, no, there's an R&D component to this. And the excess spend that wasn't revenue-generating last year is what led to models being so good this year, and led to everyone who can using Claude Code and that changing their life. This is, like, it's not a bubble, right? I don't think it's a bubble yet.

I think if AI model progress stops, that's the main thing, right? The moment model progress stops, all the spending is for naught. But so far, we've had consistent improvement. As you put in more compute, you get more performance and better models. Yeah.

Or, yeah, of CapEx, right?

Ultimately, the CapEx that Microsoft spent in 2024 for OpenAI is what results in 2025 for OpenAI or CoreWeave or whoever, is what results in their models being so good this year. Same with Anthropic and Amazon, Google and their models now being so good now is that CapEx. And actually, they still haven't paid for those chips yet because those chips still have a useful life for another few years. Right. I think model progress is very clear. The moment that stops happening, if we hit a wall, there's no new research directions, then it's cooked.

Yeah. Right.

Yeah, but like 2% of GitHub commits today are Claude Code.

As in, committed by Claude Code. You can disable that where it's not automatically committed, but 2% of GitHub commits today are Claude Code. $2 trillion of software wages paid in the world. If it was 2%, then you're like, wait a second, this is an insane amount. AI is under-earning the value that it's producing in the world, right?

By a significant margin already today.

Yeah. One of my roommates, I was asking him because he's always been a really low-level, good programmer, and he started—I was like, he's like, he had this holiday obsession, right? I mean, he was using Claude Code for work already, right? Whatever. But he had this holiday obsession. We got into playing Age of Empires II, myself, my roommate, a handful of people from OpenAI, GDM, Anthropic. We just would do LAN parties of AoE II over the holidays a bit.

Not like Christmas, but a little bit before, a little bit after, because most of us went home for Christmas. But we'd do these LANs. My roommate got so obsessed with the game that during Christmas week, because he didn't go home, he just stayed in San Francisco, he just worked on an RTS game and he built an entire RTS game. And I think, I kid you not, I think he used like $10,000 of Claude in one week and built an entire RTS from scratch, about age—but instead of being a standard RTS where it's like, oh, Age of Empires where you advance through ages, or StarCraft, it is an RTS where it's China versus the U.S. and you're in the AI space, and you go from the start of the information age all the way through to AGI and robots and humanoids and spacefaring civil—it's crazy.

He built it in a week and he didn't type a single line of code, right? He only dictated to the model. And he told me, yeah, we have an indicator internally at Anthropic where you see how many people actually write code now. There's only a few holdouts left.

Yeah, but also the economics of, like, say you spend—let's say you build a gigawatt, you put down roughly $50 billion across the data center, the chips, the networking, blah, blah, blah, blah, blah, right? Let's say it has a five-year useful life. So it's $10 billion a year. Is it a bubble if the first year you didn't make any money at zero? The second year it's zero, and then third, fourth, fifth year you're at 50% gross margins. And so you make 20, 20, 20.

Now you've made $60 billion off of this $50 billion investment. It's not the best return on invested capital, but it did pay for itself.

Is that a bubble? Well, that's what's happening today, is that people are spending all this money on infra and there's no return for a lot of it, right? A lot of it is just doing research and trying to get adoption, and it's for users, and what does that mean?

The timing. That's the timing, though. But that $50 billion CapEx was spent in year one.

It would if the utilities were willing to let it. But I think the utilities are so slow and dumb that they don't want to. Not destroy, but expand the grid. I think the U.S. could have a way better grid, but we just don't want to. No one's made the effort or taken the initiative. There's not enough power. America hasn't really built power for 50 years, right? It's converted from coal to gas and things like this, but really has not built wholesale new power on a large scale.

And there have been a lot of times where the industry blew up, right? Independent power producers, IPPs, have blown up multiple times in the 2010s when Korean and Japanese investors flooded the market because they saw such a good return there. Or before, in the early 2000s, power was growing a little bit for a little bit, and so people overbuilt on power. So the power industry has been burned a couple of times, so no one really builds power. And then you've got data centers now all of a sudden coming online and going from 2% to 10% of the U.S. grid in just a handful of years.

And so you've got this humongous, humongous change in the industry. We don't have the labor, right? I think ultimately that's the biggest problem, is the equipment and the labor. And equipment is basically, again, labor and time. It takes time to build a factory so you can build the things. I think the equipment side of things will be solved more reasonably. And one example was gas, right? People initially thought, oh, you can only use the two vendors, right? Siemens or GE Vernova for gas turbines.

They have the best ones, the most efficient ones. It's like, okay, also Mitsubishi exists, and they're ramping up production fast. Oh, Doosan in Korea exists, and they're ramping up production fast. Oh, actually, I can just take Cummins engines, right? If you've ever ridden in a pickup truck or diesel trucks, everyone loves Cummins, right? You see the Ram on the street and it has the Cummins badge. It's like, that's an aura symbol for a certain kind of redneck from South Georgia, which I have a little bit of.

Anyways, I don't have a car. I don't have a truck. I have, though. But anyways, there's all these engines. People are figuring out how to make the equipment. Solar sucks; it's too intermittent. Wind sucks; it's too intermittent. Nuclear sucks; it takes forever to build. Coal sucks; it's way too dirty. How do you make power for data centers besides gas? And, okay, the grid's not willing to put the gas on your site, right? That's what Elon did.

Now everyone's doing it, right?

So there's this annoying thing where everyone's like, oh, AI is using all the water.

AI and data centers are going to, like, use up all the water, and now we don't have any water. And it's like, that's so silly. Water is a distribution problem, not a we-don't-have-enough problem, right? You look at California, it's like California has shitloads of water, but people decide to make oat milk, which consumes like 1,000x the water of, like, anything else. Like regular milk, even. And cows obviously consume a lot of water.

But anyways, data centers consume very little water, actually, right? So the U.S. grid will get to, like, 10% of power by, like, '28, '27 is data centers. For water consumption, it's not even going to crack 1% by the end of the decade.

And so the comparison we made is because it was a bit of a shitpost, but it was serious research. Basically, we were doing serious research because we keep getting this question and debunking it, and we would do it seriously. But then I was like, no, no, no, this is too complicated. Let's make it very simple. So I was like, guys, why don't we just compare it to hamburgers, right? Because I've heard that argument from some vegetarian people before, or some Hindus. I'm Hindu myself, although I do eat beef sometimes. So we made this comparison to hamburgers, right?

Hamburgers require a shitload of water, because cows require a ton of water. And what takes a lot of water with cows, it's not the cow itself; it's all the feed you're feeding them.

Because no one grass-feeds their cows and just lets the rain take care of the grass. They either irrigate the grass or, most likely, they do mass industrial farming of corn, soybean, alfalfa, et cetera, which uses shitloads of water. Or almond milk uses tons and tons of water. Produce is, like, the main user of water. I think the metric was the entirety of Elon Musk's Colossus data center, right? Five In-N-Outs. Because you do the calculation on what's the average revenue per In-N-Out and how many hamburgers does that translate to, right?

If everyone's ordering, like, a combo, right? Okay, let's ignore the drink. Let's ignore the fries. Let's just talk about the hamburger. Let's ignore the bread, which does have grain. Let's just do the meat and the cheese. And all of a sudden, all this water—there's so much water, right? Like, a single query, like all of your AI usage from ChatGPT of the average user, is like a hamburger, right? Like, it's like, okay, this is nothing, right?

Because these things at the data centers are mostly closed loops and, sure, they evaporate some water for cooling reasons, but by doing evaporative cooling, they're using less power, right? And that's actually better for the environment than not using evaporative cooling. There's all these reasons why this myth or hoax of AI using all the water is just nonsense, right? Like, Meta's data center in Louisiana is getting protested because the water is—it's going to be the largest data center in the world.

It's going to be like four or five gigawatts, at least announced so far. We're tracking some other ones that may be as big or bigger. But Meta is getting protested because the local population around that area is like, oh, the water's dirty. It's because of this Meta data center. And, like, there's these trucks, all these big trucks, on these back roads that used to be completely empty. They're just mad and annoyed about that. But at the end of the day, what actually made the water dirty is that that's an area where you go fracking.

Fracking is absurdly worse.

And almost all of that gas is being shipped to an LNG terminal and being shipped to Asia, like Japan or Taiwan or China or Korea, and some Europe as well. Actually, all of this water is dirty because of regulations for fracking. I support fracking, by the way, but that's an insane take too, maybe. But water usage is not a relevant argument.

I think IPPs will do well. I think IPPs can secure contracts at premiums to what they've previously been able to for new power plants that are either dedicated or grid-connected but come with a pairing of a grid load, right? For example, utilities won't let you just do data centers now, but if you come with a pair, right, you're like, hey, I'm going to build this massive data center, but we're also going to have this massive power-generating asset, right? Say, whatever it is, right?

Not connected to the grid at all. Like some data centers, partially like Colossus from Elon, the original one, or part of Abilene, Texas, OpenAI, right? Like Crusoe. There's a lot of room for power producers to get outsized returns. I'm not necessarily bullish on nuclear. Existing nuclear, fine. Yeah, it can find a higher buyer, a higher-priced buyer, but the majority of it will be gas. But you can do renewables backed by gas and then just turn off the gas, and it costs more, but whatever, right?

Or you can do wind backed by gas.

Takes too long.

No one can build nuclear fast. Even China takes, like, five years to build nuclear, right? It's complicated. It's unsafe, right? I love nuclear. I wish it would work. It's just not relevant in the timescale that AI's power is going crazy. But yeah, there's a lot of interesting stuff. I had a client buy a coal plant, and we were advising them on the transaction. They just showed up and they're like, yeah, we want to buy power assets.

We believe in this power story. It's like, okay, great. So yeah, here's all of the power plants that we know of. You can get some of it from EIA. Which of these? And then we worked through the economics and we looked at new data centers being built in the region and all this. And then they decided to buy a coal plant and they restarted it, and they're making tons of money now because now a certain hyperscaler wants to buy the entire pipeline of power and put a load near it, right?

Instead of just being a grid-connected asset. So it's like a super awesome investment. So power is going to do great.

Generally, yes. Right? Hyperscalers are paying for transmission grid upgrades, which people will benefit from, right? Or investors are obviously going to benefit. Like, plumbers' wages are skyrocketing. So there's a lot of trades that are doing really well too. I think that's definitely also part of it. Yeah.

I think it's completely fine. And I think people are freaking out and making narratives where there really shouldn't be one. It's like, well, okay, Google doesn't have enough data center capacity and they need people to build data centers, but no one can build a data center because they don't have the capital. In many cases, they don't have capital, right? Or no one will give them a loan because they don't trust some random fucking company. But then Google's like, well, no, we've diligenced them.

We think they can build it. We'll even guarantee we'll buy the thing or start using it once they build it. Just having a customer alone spoken for was enough, right? In the case of CoreWeave, they were actually able to, no backstop, right? They were able to just say, hey, look, here's our Microsoft contract for this many GPUs. I want to put in that data center, that data center, that data center. Here's the contract for renting those GPUs.

I want to hire these people. I want to do this. They don't have any money, but then they were able to have it work out because they were able to get people to lend to them. I think CoreWeave did that and there was no circular financing, but that was when the scale of investment was single-digit billions or less than a billion. Right now, the scale of investment is hundreds of billions.

And so the question is like, oh, well, if I want data center capacity, how do I get data center capacity? I just go to everyone who's going to build it, looks smart, is smart enough to do it, but can't afford to do it, and tell them, I'll take it. And in fact, I won't just take it. I'll go to your debtor and be like, I'll guarantee you. Because obviously you're a new company. I vetted you, but the debtor hasn't.

And so they don't want me to just be able to walk away. Because in the Microsoft-CoreWeave deals, Microsoft could have walked away if CoreWeave fucked it up, right?

There's no—I mean, yeah, there's always cancellation or whatever possibilities. And so there's just a further form of guarantee on a lot of these backstops. As far as Oracle getting the money and then OpenAI getting money and NVIDIA paying and it's this whole circular—it's kind of nonsense because NVIDIA's getting equity in OpenAI. They're basically saying, hey, every gigawatt you buy, we'll also buy some equity.

Right. Okay, well, cool. Now NVIDIA owns an asset which they think is valuable: OpenAI, right? OpenAI is turning around and is trying to rent those—use the equity they bought. What are they—what was their use of equity? People's cash pay isn't that great, right? Mostly, just 99-plus percent of their spend at the company is probably just compute.

So it's like, okay, well then I raise this money, I'm going to do the whole thing that I explained earlier, right? Year one and two, I lose money. Year three, four, five, I hope to make money on it, right? And OpenAI has been doing that, right? So I'm gonna go out there. I've raised $50 billion. I've raised $10 billion. I'm gonna rent a cluster for five years for $65 billion.

And I've rented that contract, and now I only have enough to pay for the first year, to be clear. But you trust me, Oracle. You think I'm gonna grow and you think I'll be able to pay for it. Oracle's like, yeah. Or if you're not, I think I'll be able to sell it to someone else. So they're like, okay, cool. I'm gonna spend $50 billion this year to build that data center. And this is for a gigawatt.

And so is it circular that for every amount of GPUs they consume, NVIDIA gives an investment, that investment is turned around to pay for the first year of the rent of the cluster or second year, and then the first two years go? It's sort of like, it's fine. It is a little bit funky, but I don't think it's a big deal.

One thing is simple stuff like GitHub commits. Other things are like, what's the amount of usage? How much are people using? All these sorts of things. I think there's so many different alternative data sources for tracking AI model progress. AI tokenomics, token economics, tokenomics. And so that's, like, an entire practice for us.

Yeah, I don't believe in crypto people. I've always hated them.

Yeah, yeah. And Jensen's used it now, so I've convinced him to use the word. He's used it at sovereigns. And so I think we've won.

I've said it to him. We've written it in articles. It's an entire practice of consulting that I started in, like, 2023, was token economics. And we've been trying to build out these, but basically I think the main thing is, like, people who don't code can use Claude Code now, right? I think people don't understand that. Like, even if you don't code, you've never had any training in software development, you've never had a job as a software developer, you can code.

Let's take an example of what one of the analysts at my company did, right? Comes from an engineering background, but on semiconductor systems, right? Like, worked on mechanical systems, worked on these sorts of things. And they coded this thing, which was they wanted to do an analysis of area of clean rooms, right? Clean rooms are the building that the fab has all the tools in, the most complicated kind of building in the world, has all sorts of chemical systems, and all this area, and revenue of the company who builds these systems, right?

And so it was like, okay, we have this fab dataset, pointed it at it. It was like, hey, here's this fab dataset. What's the square footage of all of them? And we have this thing that we built, which just pulls with Claude Code separately, which, for data centers and fabs and everything else, just calculates the area of something from a satellite image, right? Very simple. So we have the square footage of all these things, pointed it at that. Here's the company. Okay, go find the filing.

So it dug through all these filings, it pulled the data, right? Okay, great. Now told it to compare these two, make a chart. Great. Oh, wait, there's this, like, weird inflection. Oh, that's because they bought a company five years ago. Can you do a pro forma of this analysis without those financials of that company they acquired? Okay, great. And then we were able to figure out an investment case for our clients, as well as some other interesting details, from someone who's never really coded, just using Claude Code and it doing this all.

And this is, like, not even their— and it wrote the note, and they didn't even work on this full-time for, like, three hours, right? They just told the model and would go work on other things, and told the model and worked on other things. It just did this. People don't understand that the skill sets that, like, I think, like, if you go talk to an analyst, right, a very junior analyst at any company, right, whether it's venture or especially growth venture or public markets or private equity, their job is, like, finding data, cleaning it, making charts.

It's like, this is Claude Code. You don't need junior analysts. Just like a lot of companies have stopped hiring L4 engineers because it's useless. Why would I hire an L4 engineer? I just tell Claude to do it. This has happened, and this is a really big shift, I guess. Low-level knowledge work just doesn't matter, right? Why would I use Excel when I can just tell Claude to manipulate CSVs?

Why would I use Word when Claude will just generate the markdown and I can copy and paste the markdown directly into our WordPress? And then that WordPress is fully formatted now, and it's like, oh my God, what's the point of Word, right? And what's the point of doing all sorts of stuff? Opus 4.5, it's coming somewhat soon in the March-ish timeframe, maybe February, March-ish. But yeah, because OpenAI has a better RL stack than Anthropic today. It's just their pre-trained models suck compared to Anthropic's pre-training.

Right. And so if they catch up a lot on pre-training and keep their better RL stack, they would actually have a model that's much better, right? Flip side, Google has a better pre-trained model than Anthropic or OpenAI, but their RL stack sucks. So if they catch up on RL, these models are gonna get ridiculous. And then Anthropic is obviously advancing as well, right? And then you look across the ecosystem, everyone's advancing really fast. Progress, these moments are happening, right?

ChatGPT was a moment, Ghibli was a moment. Those were more consumer. Those were less like—I mean, ChatGPT is everyone using it for work too, but I think Claude Code is like a new moment, right? Opus 4.5 on Claude Code is a new moment where the way you work has forever changed. And so now we're trying to force everyone in my company—there's 54 people here—I think half of them have coded. The other half, we're trying to force them to use Claude Code.

And it could be like, oh, well, actually, you come from a semiconductor consulting background. Oh, you come from semiconductor engineering of packages. Oh, you worked in a fab, right? These kinds of people, they're using Claude Code now, right?

And their productivity's being boosted. And it's like, Claude Cowork is new. It sucks compared to Claude Code, but it'll get there, right? He said he coded it entirely in Claude Code. You know that, right? Or it was on your pod, right?

So I've heard that, and I think maybe that might've been from your pod. Original disclosure.

Oh, okay, okay.

I think it's like a brand-new age, and there's so much low-hanging fruit. As Sholto said on the episode when he was here, there's so much low-hanging fruit. Yeah, I mean, for the models progressing. And then I think model progress will translate to revenue. Adoption is difficult, but actually, the UX of Claude Code sucks. But give it six months, the models will be good enough that the UX can be like talking to it.

And you don't even have to have CLI integration, right? It's something even easier. Or Claude for Excel was released recently, and it's not bad. Building models and all these sorts of things are just gonna be like, tell someone, right? Like, why tell a junior analyst, right, when you can just do it yourself? I think it's a whole new world. And it's the $2 trillion of software work, but also of wages. But it's also, we have north of 2%.

2% is Claude Code, and then there's Codex and Cursor and all these other guys. So probably like 5% of code committed today is AI-generated, if not higher, marked as AI-generated. What's going to happen when normal workers who do spreadsheets and office processing start automating their workflows? I think it's a whole new world.

I'm straight, but I've been accused of being homosexual, which is perfectly fine, for how much I praise this man. Because think about it, right? He's like 6'4". He's really good-looking. He has an Australian accent, sounds amazing. You've heard his episode. I have an annoying voice, probably. His voice sounds amazing. He's absurdly good at coding. He was an Olympian-level fencer. He picks up any sport, he's really good at it, right?

Because he's athletic. It's like, holy crap, you're a specimen. Yeah, yeah, yeah. This is an eclipsed incentive for sure.

What's a podcaster's podcaster mean?

Yeah. Yeah. When he's preparing, it's like he's so locked in and he prepares so hard for interviews. It's great.

And then he might only say like 100 words on the episode, but he's prepared so hard. And then I think people just realized, oh wow, he's not just like, oh, he just has good guests. No, no, no. He's preparing really hard, but you can't tell if you're not realizing that. And then once he started writing more, people were like, oh wow, he's actually really, really smart. Like, yeah, because he's studying like crazy.

Like, it's like, oh, I'm interviewing an AI researcher who worked on this. I'm gonna try and train a fricking model, right? It's like, that's the level of commitment he goes to when he records this stuff.

With Sholto, it's like the Age of Empires game, because we got super into it for a bit. We talked only about that and his RTS that he made. With Dwarkesh, it's, I mean, it's all sorts. It's like normal roommate stuff. It's like, how's your dating life? Oh, okay. You went on a date? It didn't go well? Okay. Well, okay. Yeah. Like, that's me.

That's me. My dates don't go well. No, just kidding. Or like, it's like, oh, you want to have dinner? We can invite a few friends. Like, yeah, great. Or like, it's all sorts of normal stuff too. Obviously, we also do talk a lot about tech, right? Like, this is our lives, and tech is the most fun thing.

Thank you so much.