Chip War author Chris Miller: China's $295B AI plan is $59B/year — less than one hyperscaler — and Beijing isn't AGI-pilled

Speaker 1: And our next guest is already in the waiting room. We have Chris Miller, the author of Chip War, a fantastic book. If you haven't read it already, he's been on the show before. Welcome back to the show. Chris, how are you doing? What's going on? Doing well. Thanks for having me. Thanks for hopping on. I guess the first place to start since you've been on the show before, we did the whole deep dive on Chip War. I've read the book. It's great. Hopefully, most of the audience has read it. If they haven't, they should go pick it up. But what do you think has changed the most since you wrote Chip War? What what is what have been your biggest updates to your sort of world model, your way of thinking about the Chip War broadly?

Speaker 6: I think the big change is that for the last two decades, if you look at semiconductor spend as a share of GDP, it was roughly flat. Yeah. And then in the last four years, it's roughly doubled. Yeah. And that has been all been driven by AI. And I think, you know, the question is doubling is fine. Can it triple? How much more can it go relative to trend line? That's the big question, I think, hanging over the industry right now. And is that gated by TSMC capacity,

Speaker 1: energy capacity, AI demand? The AI demand seems to be there. Like everyone's inferencing models all day long. The models require more and more chips. A single model these days needs an NVL 72 or a whole rack. But what do you see as sort of like the key inflection points that people should be watching for as the semiconductor industry continues to boom or, you know, tops out at some point?

Speaker 6: Yeah. I I think right now it is it is gated by TSMC and manufacturing capacity more broadly. If But you put yourself in the shoes of those companies, they're looking at AI demand, not sure how strong it's gonna be. Obviously, strong, but it's strong enough to double or tripling if semiconductor spend is a shared GDP. That's a a big jump. And, of course, they're on the hooks for the CapEx. They've got to build the factories. They've to pay for them. And so I think companies in the manufacturing layer are excited but nervous about just how far this is going to run. They have an incentive as a result to be a bit more cautious on building chip plants than AI companies would like. In historical

Speaker 2: infrastructure build outs, let's say like electricity and railroads, was there ever a key supplier that just didn't have any sort of like insane FOMO fear of missing out. Like I'm thinking like railroads, was there a critical supplier that was like, you know what? We're excited about this trend. We're investing in it. We're building enough railroad ties. Yeah. Like, you know, if there's any historical Because it feels like a very it feels like on one hand, obviously, there's frustration Yeah. From other players in the supply chain. But at the same time, it it feels like somewhat like how like, potentially healthy even though I think you can Mhmm. See that bubbles have have, you know, produced great results Yeah. Throughout history.

Speaker 6: Yeah. I think we are in a somewhat unique moment in that there are a series of choke points down the supply chain where you just rely on a single company for most capacity, and that didn't happen with railway ties. Sure. Yeah. It hasn't happened with electricity. You have it with TSMC layer, for sure, but then you look at their suppliers for key materials, for advanced packaging, and there's often choke points there as well. And the trick is that the deeper you get in the supply chain, the further you are from the actual end AI demand and the more skepticism builds up as to how real is this actually. That's why there's been underinvestment

Speaker 1: in the materials and manufacturing relative to what Silicon Valley needs. Yeah. It's the bullwhip effect. Right? A small change in demand amplifies and amplifies and all of a sudden they don't want to get caught at the tail end. And also what I know know Nvidia traded all over the place during the crypto boom when cryptocurrency miners were using GPUs and then Ethereum went proof of stake instead of proof of work and there was a drop off in demand there. But I'm wondering, do you remember if TSMC had the same level of skepticism or or or was even affected by the previous rise and fall of NVIDIA?

Speaker 6: You know, I I think the key is that before the last couple of years, NVIDIA was one of many customers that TSMC Sure. Had and Apple was the most important. Yeah. Whereas now it's kind of flipped. NVIDIA and AI in general is the key growth driver. So in the past, TSMC would have more closely tracked smartphone sales than they would have tracked data center build out. And smartphone sales were on a much more steady pace

Speaker 1: since Apple has their whole machinery of refreshing and planned obsolescence potentially. Yeah. You look at individuals are not like I need like the iPhone's gotten so good, I need five phones. Yeah. Yeah. Yeah. Yeah. There was no exponential take up. But now there's like, okay, I need 10 more, 10 times more tokens, a thousand times more tokens, etcetera, etcetera. Staying on TSMC, I'm curious about the culture there because TSMC is known as this like champion of Taiwan, the most elite institution, the crown jewel of Taiwan. And like why we explained the bullwhip effect a little bit about the reticence to deploy more CapEx. But is that is there also some sort of cultural element or capital markets dynamic? Because when America has a crown jewel, we're like, send it to the moon. Like, oh, oh, there's a there's any excuse to 10 x capacity or 10 x, you know, the the scale of our crown jewel. Like, let's do it. And it seems interesting that you would have a company that has this opportunity and there's like a glimmer of a hope for it being even bigger and yet there is it more of like this protective mindset of like we have this good thing going, let's not risk it or is it just like they can raise prices so they're less there's less downside there?

Speaker 6: You know, I think it is a conservative culture, but I think they're they believe their conservatism has served them well in the past. Yeah. There have been cycles in the past, and and being cautious going into this cycle gives you the opportunity to invest at the bottom of the cycle. And TSMC has done that really well historically. Now I think, you know, right now, they they realize that we're not at the top. They're building new factories as we speak. But I think they're also, you know, looking at the history and saying, we don't wanna overinvest Yeah. Relative to what our customers are gonna need. Yeah. And then and then what

Speaker 1: does the the the domestic chip supply look like from your perspective right now? There's a ton of energy around Intel, and it feels like Intel checked a lot of the boxes for the comeback. Like, the comeback hasn't really arrived yet, but in terms of, the backstop from the U. S. Government, additional investment, the stock price is up and a bunch of buyers and potential customers are falling in line for their most advanced node, it feels like things are on track there. But what else is going on? And is that actually a pressure on TSMC at some point?

Speaker 6: I think to some degree that does create a bit of pressure. I agree that Intel looks like it's in a much stronger position. I think that customer traction is the key thing. And as you say, we haven't we've seen news about customer traction, but not actually volumes yet. And Intel really needs substantial volumes of customer orders. So I think that's one positive sign. I think everyone's watching carefully what Elon's gonna do with TeraFab. Yeah. How that's going to grow. It could be transformative. It could not be. It's obviously a hard industry to to break into. And then, you know, there's Samsung with a a big plant in Texas. They're bringing online and also Tesla's reportedly one of their customers. So you look across the ecosystem and there's certainly a lot more optimism about leading edge manufacturing than it would have been just a couple of years ago. Bull and bear for TeraFab.

Speaker 2: I would love your takes. You know,

Speaker 6: I I think a, never bet against Elon. But b, you know, it seems like he's lining up a coalition of partners on the IP side, on the manufacturing side.

Speaker 1: It'll be interesting to watch how it develops. I'm excited. Yeah. I guess the question is, like, how much of the fab is him versus he's the buyer and there's and it's actually Intel doing everything because I don't think anyone doubts that Intel can like run a fab since that's what they've done for decades. And if you and if you bring on the right partners, seems like it could be possible. Well, interesting. What do you know about

Speaker 2: ASML's current thinking around AI progress and the build out and and how they're thinking about investing against this progress?

Speaker 6: Well, think ASML is in a similar position to TSMC. They want to meet customer demand, they don't want to overbuild. And they're also on the wrong side of the bullwhip. And so they have an incentive to be a bit more conservative than their customers do. But it's been a great year for ASML as, obviously, demand has just been through the roof. That's I think that's going to continue for for some time. I think the other challenge that ASML is is working through is that TSMC, their most important customer, has said they're not going to buy the latest high NA lithography machines at least for a couple of years. That's also a

Speaker 1: medium term challenge for ASML because they're trying to sell their next generation machines. It sort of doesn't matter how AGI pilled you are if your customer doesn't show up to buy the thing that you're making in the supply chain. How have you been how have you been processing the export export controls back and forth? Because it feels like it's been every other day. It's like, okay, we can send anything and then we can sell the nerf stuff and then we can sell nothing and then everything. And it's been going back and forth and that's sort of the nature of this administration. But is there is there a clearer through line that you've detected? Is there something else going on there, or is it as chaotic as I've been reading into it?

Speaker 6: Yeah. I I think, you know, step one is don't read the tweets. Mhmm. And step two is look at congress. Mhmm. Because congress that's where you get the bipartisan support for these export controls. And I think congress has been playing an important role in influencing the administration. You've seen a bunch of influential Republicans, for example, publicly support export controls, trying to put pressure on the the White House. I think the the third thing that's important is, Methos really did change thinking in the White House about Mhmm. How powerful AI is gonna be, and how important it is to have the leading capabilities to yourself. And so I think there's been a shift around the AI debate in the White House just over the last couple of weeks. Yeah.

Speaker 1: So related to that, there's reporting in Reuters today. China prepares $295,000,000,000 plan to fund nationwide AI build out. How did you interpret this? Do you see this coming? And then I'm interested to know, like, do they have the supply chain to actually deploy 300,000,000,000 Because the numbers that I've seen back of the envelope for if the chip controls get weakened and NVIDIA can sell the TAM is like $10,000,000,000 which would be great, but not like huge and certainly a small chunk of this. So are they advancing enough with SMIC and SME to indigenize their supply chain and actually spend this money? Do they have other plans? Like, how do you see this playing out?

Speaker 6: You know, I think first those numbers are over five years. So annually, that's less money than Google or any

Speaker 1: of the hyperscalers. So just to put into context. Play the bald eagle sound, baby. We still got it. We still got What do they know? Do they know? What do they know? What do they Chris?

Speaker 4: No, seriously. Seriously,

Speaker 6: like Yeah. Do read into it? They're they're No. I mean, this is the puzzle. Yeah. The puzzle is is why isn't Xi Jinping more AGI built? Sure. You know, if you thought AGI AI was important, you thought chips were an important ingredient, China's been underspending for the last four years on AI. And and I think it's because the Chinese government just doesn't really believe Interesting. That AI is gonna be nearly as important as we do. And if they did believe it, they'd be, a, buying h 200 chips Yeah. And, b, spending more on on data centers. They're not doing it. Is it possible that they just

Speaker 1: that that like, they believe in AGI or something, but they just have longer timelines? Like, it feels like a lot of the underwriting that's happening in The United States is based on, like, eighteen month of coincidentally exactly the VC timeline for for around. It's like what will we do in the next eighteen months? We're AGI in 2027, 2028. Really short timelines here. And then that's also backed up by Dwarkash Patel had a post saying like if AGI comes soon, America wins. If it comes longer, China wins. And it feels like that's almost like I like his analysis, but it's also like a reflection of the of the revealed preferences of the two countries where America is sort of building towards a fast takeoff AGI soon scenario. And China might actually not be like necessarily behind the ball. They might just be optimizing for a 2,035 AGI. And if they ramp up slowly, there's gonna be less gyrations, less of a bullwhip effect, and they can take a longer view. How do you process that?

Speaker 6: Yeah. I I think one explanation is just that The US is a service sector heavy economy, and AI is gonna be more impactful in services first, manufacturing second. Yeah. And so that that could be part of the story. I think, you know, second, US firms just could be more tech forward. You see this in in cloud computing. You know, China has a tiny cloud computing market relative to GDP because its large enterprises just don't wanna move to the cloud. Oh. And so that that could be another analogy. And then I I think third, we we should be open to the prospect that the Chinese government just hasn't gotten the memo yet. Yeah. You know, for better or for worse, we have no deficit of venture capitalists and tech entrepreneurs in government. Yeah. And of course, China is not the center of government. And then that could be another factor. Yeah. You would think they would be excited about it purely from an authoritarian lens? Because like if you have like the all powerful model that that can Yeah. Basically theoretically

Speaker 2: They already have all of every single Well, them c c every single citizens Yeah. The content that they get, the services they get access to, you know, how much how much the retirement benefits they get. All all all these different things like it it it In The US, we think of this authoritarian AI scenario as like a, you know, doomsday scenario. It sounds terrible. But if you're an authoritarian. But if you're an authoritarian you think like, I gotta get I gotta get one of these authoritarian models. Yeah.

Speaker 6: I don't know. Yeah. I I I I think that dynamic is I I think the risk if you're if you're an authoritarian leader is you don't know what the model's gonna output. They're not nearly deterministic enough for your preferences. And that that's the that's the downside for any authoritarian leader. Yeah. There was recently Or or or

Speaker 1: better better better training data. There there there was recently reporting instructed the Kremlin to disconnect or or firewall or air gap all of the CCTV cameras that are, I think I guess, inside the Kremlin or or or like could potentially triangulate him because they were worried about a hack where his location is disclosed to a foreign adversary. And so, yeah, you can see that these tools cut both ways. Have you have you actually followed anything on what's going on in In Russia?

Speaker 2: But before we go there, wanted to ask how closely have you followed the whole Manus saga Yeah. And debacle? Do you know the latest? It feels very unclear how something like that gets, you know, unwound, but

Speaker 1: Yeah. Are they still in China? Yeah. Don't know.

Speaker 6: Well, this is I think the the really interesting part is that, you know, if you have a broader campaign of pulling passports from leading entrepreneurs, that is a extraordinarily chilling sign for the rest of the the industry. I think there's a real you know, if I was sitting in in Beijing, I'd be really nervous about being too tough on my entrepreneurs because those are the people you need to invest and build companies. Yeah. And that's a pretty risky move, I I would think, for Beijing in terms of building their own domestic ecosystem. I mean, Liang Wenfeng, the founder of DeepSeek,

Speaker 1: has the most control, I think even higher level of control than Elon at SpaceX in terms of founder ownership and concentration because he had High Flyer, the hedge fund, and then he funded DeepSeek. So he did the first couple rounds and provided like 80% ownership. I think it's like 80% ownership and I mean, it's already a multibillion dollar company. Could be one of the biggest companies in China if AI labs are any indication in America, if there's any correlation there. And so he's gonna he should get a second passport maybe or something. I don't know.

Speaker 2: Anyway.

Speaker 1: But Russia Russia. Russia. Trying to we've been trying to figure out how many like GPUs they have. Well, ten years ago Vladimir Putin was giving a talk and he said whoever wins AI wins the future and he lay he was AGI pilled before way before it was cool. Then he got into a crazy war and he and it seems like he lost all of his best talent to Nebius and what what's the other one? Clickhouse Yeah. And a bunch of other Yandex spinouts and stuff. And so I'm just wondering like has Russia come up at all in your research of, like, where chips are going, where data centers are getting built? I just don't hear them in the conversation at all and it's surprising to me because they have nuclear weapons.

Speaker 6: Yeah. No. I I think I I think you're absolutely right. They're not a big player, and the reason is because Nebius is outside of Russia. That's an example of a team that was one of the leading groups of tech entrepreneurs in Russia, at least until 2022, and now they spun outside of Russia and are building a great company in the West. And that's a microcosm of the entire tech ecosystem. The best estimate is that hundreds of thousands of Russians left the country after 2022 disproportionately well educated Wow. Disproportionately in the tech sector. And so, yeah, there there have been some reports of GPU smuggling into Russia, but it's, I think, small scale because there's just not the demand for building big data centers. They're gaming. They're they're getting those GPUs. They're gaming.

Speaker 1: It's like they they only smuggled ten eighty TIs? You can't run advanced LLMs on those. Yeah, they're gaming. Talk to me about the reception in China of American GPUs, Western GPUs. There's like interesting smuggling dynamics, interesting debates over NVIDIA export controls. But then we see these articles pop up every once in a while saying like Beijing doesn't want them or Beijing is going to pressure local companies to use the Huawei Ascend chip or go more indigenous. And I'm just interested because we label China as like authoritarian but obviously no system is perfect one way or another, obviously. And so, do they actually have the ability to restrict this? How much do they actually want? Do they really care because it feels like the pool of compute is pretty fungible and I don't know how are they worried about like security risks or dependency? Like, would you rank their worries?

Speaker 6: Yeah. Yeah. I mean, what we know is there is smuggling at a reasonably large scale. Mhmm. There's also offshore compute access. If you're a Chinese hyperscaler, you can get access to compute in Malaysia Yeah. Or wherever else. Mhmm. But it seems like the Chinese government is is really fixated on pushing firms to buy Huawei. Mhmm. And it's surprising, I that think, they haven't let Chinese companies buy the h 2 hundreds Mhmm. Despite that it's now been half a year in which the Trump administration has been allowing sales. That goes back to what's the Chinese government trying to do. If you were AGI pilled, you would buy the h two hundreds. Yeah. But if you're not, then you're focused on domestic manufacturing and supporting the Huawei ecosystem. If you're worried about data security and think there are backdoors, which I think the Chinese might be worried about, then maybe say ban the foreign ships. Yeah. There'll be a bit of smuggling, but better to build up your domestic ecosystem first. And that does seem to be the revealed preference of the Chinese leadership.

Speaker 1: Why do you think what are the keys to China's broad support of AI locally? AI pulls so terribly in The United States. It pulls terribly in the EU as well. But China is almost flipped. It's like 80% support artificial intelligence. Only in America, it's like 20%. What's going on?

Speaker 6: You know, I think part of it might just be the media landscape. You know, in in in China, if there are topics that the government doesn't want discussed, they don't get discussed. That's part a. You know, b is that the the entire tech sector is much less controversial among the Chinese populace than in The US. We've had a decade of debates about Mark Zuckerberg and social media. That's been present in China to some degree, but, much less toxic, than in The US. And then I I, you know, I think third, US AI firms have been uniquely bad with publicity. Mhmm. Promising to destroy jobs, promising to threaten the future existence of humanity. That's been the marketing of AI over the last couple of years and, you know, when you hear that, it's not surprising people don't like it. Yeah. Yeah. That makes a lot of sense. Is anything happening in Japan that's relevant right now? I think the big story in Japan is the the number of companies that have been discovered to have extraordinary materials capabilities that sell into the chip supply chain. Mhmm. So my favorite example is one of the toilet makers in Japan

Speaker 2: has an advanced ceramic capability that's used in the machines that make chips. I know and it's so funny because when the headlines started breaking around that Mhmm. This was probably two three months ago, a lot of people were like, wow, top signal,

Speaker 1: like Yeah. The toilet company is getting into Yeah. It was like an all birds Yeah.

Speaker 2: It looked like just like a pump but no, I actually have, you know, an industrial process that's very valuable and So fun. Very real. Yeah.

Speaker 1: I mean, yeah, great lineage there in I mean they were doing humanoid robots. The Honda Azimow or something was in like 2002 and it feels like the Japanese economy just sort of lost a step and just didn't keep doubling down. They were just too early and so they ceded a lot of that territory to China and America.

Speaker 6: But I'm I'm rooting for a comeback because I like that Honda robot. It was cute. Didn't feel like Terminator at felt like c three p o. I don't know. That's right. That's right. I think the Japanese robotics firms, I I think they're they're so focused on factory automation. Mhmm. And that's where their expertise really lies that they're struggling to get their heads around what does AI plus robotics actually

Speaker 2: mean. Sure. Sure. Interesting.

Speaker 1: Last question for me. Is NVIDIA a car? What's going on with their moat? How are you thinking about NVIDIA in the supply chain? How important is their role? How defensible is their position? The debate was Jensen saying NVIDIA is not a car. It is not something you can just step into and get where you're going. It's a special product from a special company.

Speaker 6: So I think we're into year two at least of the ASIC debate. Yeah. It's clear ASICs will have some market share, but it also seems clear that NVIDIA is gonna be the dominant player by far. I thought it was interesting to watch NVIDIA acquire Grok technology six months ago showing that they're gonna be able to diversify into inference focused hardware, building their empire of sorts, and CUDA still remains a very powerful part of the moat. So, you know, is the market share gonna decline from 90% to something lower than 90%? You know, that that's what's happening. But that does that mean NVIDIA is, not gonna be the dominant player? It certainly seems like it's well on track to be the key provider of GPUs for long time to come.

Speaker 1: When's the next book coming? I was going to ask. And and tell us the title. Is he but obviously, you can't announce anything. But if you were to create the next chip war, would it be called Chip War two or would it be called Battery War or Robotic Supply Chain War or, you know, Energy War? Like, is there a new war that will supersede the the chip war, become the most relevant topic or are we just doubling down on chips?

Speaker 2: You're you're kind of using a little bit of fear based marketing yourself, Yeah. War. Yeah. A

Speaker 1: Why not just beautiful trading and collaboration and yeah, there'll be prices. Chip trade, I don't You know, yeah, chip trade doesn't sell Chip any industry. Chip industry dynamic. Yeah. Chip industry deep dive the book. That's what yeah. Fear based marketing or you caught him redheaded. It's working on me. It's working on me. I want I want the next one. The next book, the next hot topic, the next thing that you'll be digging into. What are you thinking?

Speaker 6: You know, I I think I'm spending all my time on chips right now. Okay. I don't know if there's gonna be a sequel, Chip War two. Chip War two, Revenge of the Oh, that sounds so good. Revenge of the Chips, that's where we are. Yeah. But sequels are never as good as the original, so I think probably not a sequel on that topic. Maybe Chippeas.

Speaker 1: You've done Chippeas. Chippeas. Hope so. What is the path to Chippeas? That's what I want to hear. A lot of fear mongering, lot of worries about war. We need a lot of ways to get out of quagmires, I think. That's generally what I'm rooting for. Anyway, thank you so much for taking the time to come chat with us. This is a lot of fun. Yeah. Really appreciate it. Glad to have back on soon. Go get the book, Chip Ward. Available everywhere books are sold. We'll talk to Have you great Goodbye.