中国的终局：ASI时间表、中美关系与1.7万亿美元AI泡沫——Alvin Graylin | 281

There seems to be a prevailing view on campuses that ASI is 5 to 10 years out, maybe even 20 years, so I'm really curious what the prevailing view is in China.

They are not behaving like they believe ASI is around the corner.

Why do you believe getting clarity and some resolution on the U.S.-China AI race is so important right now?

Having a race condition forces people to make irrational decisions. At some point, we will get to a superintelligence-type scenario. If and when we do, the concept of nations will probably become a lot less important than they are today. The AI sector alone is worth more than the GDP of America today. That, to me, is a sign that we are in a very, very fragile place, and an economic correction is due.

What I hear you saying in your wargame is that sometime in the next 2 years, there's a private credit bubble that the U.S. is using to finance its data center build-out. The bubble pops, and then the U.S. asks China to help financially in return for what—a quid pro quo regarding Taiwan?

I will tell you this.

Welcome to Moonshots, everyone, your number one podcast in all things AI and technology, your front-row seat to the accelerating singularity. I'm here with my magnificent Moonshot mates, the original four, that including me, uh, AWG, DB2, and Salim, uh, my brilliant colleagues who every week help me, hopefully you, understand what's happening at this incredible rate of speed. I'm Peter Diamandis, your host, and abundance entrepreneur and evangelist. Welcome to a very special episode of Moonshots today. Today, our mission is to go deep on China and AI and discuss it with someone who's lived and operated inside both the Chinese and U.S. technology ecosystems for over 35 years. Alvin, welcome.

Yes, it's great to be here. I watch your show all the time, so I'm glad to be able to chat with all of you.

We're going to quiz you, then. I'll ask you trivia as we go.

When do we say “drink”? What's Alex's favorite term?

“Tyson's form.”

“Tyson's form.” “Tyson's form.”

Yes. You got it.

“Tyson's form.”

Drink, drink, drink.

Let me do a proper introduction of Alvin. Alvin Wang Graylin is both a friend and someone who's held senior executive roles at HTC, Intel, IBM, Trend Micro, and WatchGuard Tech. He's worked in all 5 layers of the AI stack—the data centers, chips, PCs, phones, XR glasses, and apps—and he's built and founded over 4 startups.

Two years ago, we were discussing his recent book, Our Next Reality. He's now a digital fellow at the Stanford Institute for Human-Centered AI, a senior fellow at the Asia Society Policy Institute, and a professor of AI policy at the University of Washington. He's currently supporting the U.S. government for the coming U.S.-China AI safety dialogues. That's going to be happening on September the 24th in D.C. The very next day, on the 25th, we have our Moonshots Live event.

He's lived and operated extensively in the U.S., China, and Taiwan, with dual master's degrees from MIT in computer science and business. I love having another MIT graduate on the show, along with an electrical engineering degree from the University of Washington.

This is a conversation I've been waiting for for a long time, to really go deep and understand what's going on. Alvin, you're going to bring a unique perspective on the U.S.-China AI race. You've argued that the game isn't a prisoner's dilemma; it's a stag hunt, a game-theory model about coordination and trust.

Alvin is the author of 2 key papers, Beyond Rivalry and Misdiagnosing the U.S.-China AI Race. Last week, he released a new paper around AI security, The Biggest AI Models Are Not the Biggest Threats, where he proposes that it's the smaller AI models we need to be more worried about, and that cooperation is the only path toward safety.

Alvin just returned from speaking at the World AI Conference in Shanghai, where President Xi delivered the opening keynote. He's had a chance to meet with leadership across all of the Chinese AI labs and discuss AI governance issues with senior Chinese regulators and policymakers. We'll hear about how they're thinking directly from Alvin.

Today's podcast is going to be far-ranging. We'll be covering topics from China's open-weight models and using AI for diplomacy, as well as robots and AI regulation. Very importantly, we'll be discussing the coming U.S.-China AI safety dialogues. I want to understand the objectives, Alvin, and what you think might be accomplished.

We'll close with Alvin's recent Substack essay called Great Reckoning Before the Reconnecting. I love that. You and Alex both have wonderful terminology, and I love your essay on abundancism. All right, let's dive in. There's a lot to cover.

Alvin, one of the things that we pride ourselves on this show is disclosing all of our connections. Since we're talking about a sensitive topic—the U.S.-China relationship—and you've spent 2 decades operating in China, your bio lists roles like vice chairman of AVRA, the VR Industry Alliance endorsed by the Chinese Ministry of Industry and Information Technology, and a 3-year professorship at Beihang, a defense-linked university on the U.S. Entity List.

I want our audience to understand the full picture. Would you walk us through those relationships, past and present, with any Chinese government bodies or government-linked institutions? What was your involvement? Is anything going on now? Were you compensated?

I want to understand your connection to the Chinese government so people understand where you're coming from. You are a U.S. citizen. People should know that. But given that we're going to be talking about a lot of sensitive topics, give us your background there, if you would.

1. Alvin's China Connections

Sure, absolutely. You're right: I'm a U.S. citizen, and I've been one for over 45 years. I moved here when I was very young. I was born in China. Having worked and lived in China, you have to deal with the government on a daily basis because that's an important part of being functional there.

The IVRA, the VR Industry Alliance, was an industry association endorsed by the government. If you wanted it to be a functional organization, it had to be endorsed. There were 300-plus members, and about 1/3 of them were international companies—companies like NVIDIA, Samsung, Qualcomm, AMD, Google, and so forth. These were the kinds of companies that were trying to get in to accelerate the industry.

When I was working there for HTC, which at the time was the leading virtual-reality and augmented-reality company in the world, I was the head of the organization, the head of the company in China. HTC is a Taiwanese company, so at the time I was working for a Taiwanese company, but we were given a lot of influence because we were such an important party and an important player in the industry.

Beihang University is a very large university. It's a university of aeronautics and aerospace, but it also was a leading university for virtual reality and augmented reality. They've been teaching that for over 30 years. In my role as head of HTC, which was a virtual-reality company, they wanted me to teach there on a part-time basis.

Neither of these positions was compensated. After I left China in 2024, I have not had any involvement with either of them. So, just to clarify where things are, I think the thing to remember is that, for you to understand and work with any industry or any government, you have to understand both sides.

I think this is why, having had close discussions with them and worked with them at the city, provincial, and some national levels, you understand their mindset. I think that's very important to have proper dialogue.

Alex, do you have any other question you wanted to ask?

Yeah, we'll get into it, but this should certainly be an interesting discussion.

Yeah, for sure.

We got Alvin's professional bio, but his growing-up bio is really interesting too. Why don't you tell us about your childhood and growing up, and then how you got to the States? It's a super cool story.

That is a little strange, but I was born during the Cultural Revolution. Both of my parents were artists, and my mother sent a letter to Mao Zedong's wife because she had closed the ballet school that my mother had helped co-found, and my mother was sent to be reeducated. This is why I was born on a Chinese reeducation farm during the Cultural Revolution.

So, Alex, I do understand some of the downsides of what happens in improperly organized or governed states.

I was able to move here in 1980 when my grand-uncle helped sponsor me. My grandmother was a reporter for the New Tribune during the Sino-Japanese War. She was there reporting, but had to leave my mom behind when the Japanese bombed Pearl Harbor. It took her 8 months to get from Shanghai to Chongqing, where the Flying Tigers were.

This is why I was born there, and this is why I’m actually part Ashkenazi, part Scottish, and part Chinese. So it’s a little strange for my generation, actually.

Yeah, amazing. It’s worth noting that your brother is, I guess, the first nuclear-sub commander in the U.S. Navy?

Yeah. He was a senior officer in a nuclear submarine, actually—one of the big boomer ones that has the capability to destroy nations. In fact, 3 of his 4 children are now active officers in the U.S. Navy and went to the Naval Academy in Annapolis.

Yeah. When we talk on this pod a lot about the fact that the U.S.-China AI race is driving a lot of what’s going on, there’s a lot of important policy happening. In the same way that the U.S.-Soviet Moon race drove the Apollo program, we keep falling back to the reason for racing in the U.S. in AI model development: to make sure that we get to ASI before anybody else. That’s been coloring everything, so I want to get into that on this program. It’s important for everyone listening to understand the backdrop of this.

2. China's Open Model Surge

So let’s kick it off. We’ve talked a lot about open-weight models over the last few weeks on this pod. Alvin, in your essay, “Misdiagnosing the U.S.-China AI Race,” you lay out a staggering shift: Chinese open-weight models climbed from 2% to 61% of OpenRouter traffic in the last 2 years. Alibaba’s Qwen model alone has over 700 million downloads. I’m sure the number has surpassed that since I looked.

It’s up to 1 billion now.

It now has 180,000 derivative models. One of the things we talk about, again, is that when you have an open-weight model, it’s very easy to fork it, develop it, and retrain it yourself. You say that the U.S. export controls on Anthropic’s Fable Five backfired spectacularly.

Specifically, within 24 hours, China’s Z.ai released GLM-5.2 under an MIT license. In fact, that was the model used to deal with the Hugging Face debacle. Brazil’s Rio built on Qwen. Japan’s Sakana released Fugu Ultra. You say that the U.S. denial strategy didn’t slow China at all; it accelerated innovation and alternative ecosystems. And you call this a “denial-keeps-us-ahead” fallacy, right?

So let me throw out the first question: why do you believe that getting clarity and some resolution on the U.S.-China AI race is so important right now?

3. The Race Has No Finish Line

Yeah. In fact, this is probably one of the most critical questions that we need to get resolved, because having a race condition forces people to make irrational decisions. Right now, there is a perceived race condition, and it’s based on some assumptions that I think are actually misguided or maybe misunderstood. There’s a perception that there is a finish line, a perception that the world is zero-sum, and a perception that whoever gets to AI first or AGI first can somehow rule the world forever, right?

This is a belief held by a certain segment of policymaking circles, and at least a certain portion of Silicon Valley has this type of belief, right? I think that narrative forces the whole discussion into a national security issue, when the reality is that right now none of those assumptions are really based on real data today, right? First of all, we do know that the world is not zero-sum. As you guys talk about every day, every episode, the world’s getting better. We’re getting more resources. We’re getting more abundant, so it is not a zero-sum game.

There is no clear finish line, right? In the sense that as these technologies get better, they’re progressing. And as you mentioned with all of these open-source models, as the gap moves from a year and a half to now probably a 2- or 3-month gap between the open-source and closed-source models, there is no finish line where you say, “Okay, we’ve won,” right? It’s not like the space race: you say, “Hey, we’ve landed on the moon. We’ve won.” There is an end. Whereas an arms-race-type model that we are in today is a constant spend and a constant pursuit without clear value being returned.

I don’t want to take too much of this, but I think this will help set up the conversation that we’re having.

Maybe, just to pull on the thesis, Alvin, I think it’s latent in the thesis that there isn’t an endgame. I’d love to understand how you think about this. From my perspective, there’s an obvious endgame: there’s space, development of the solar system, and interstellar exploration, all of which I expect to be fulsomely and holistically supported by superintelligence. Surely, somewhere among the various scenarios that I assume you’re analyzing, there are scientific and engineering endgames, quote-unquote, that are intrinsically valuable to pursue.

Is it your thinking that science and engineering and solving everything, as it were, is not the endgame? Is there some other non-endgame that you have in mind? Do you think that this is sort of a Red Queen-type scenario where intelligence is just endlessly racing as an end to itself? Or is there an honest-to-goodness “afterwards” after the singularity, in your mind, as it pertains to U.S. v. China?

What you’re describing, in terms of at some point getting to a superintelligence-type scenario, may be the case. But if and when we do, the concept of nations will probably become a lot less important than it is today, right? The reason that we’ve created nations—it actually started with city-states—is because we wanted to protect a certain level of resources and defend and gain additional resources.

In a world where we actually do achieve ASI and get the kind of abundance that Peter and all of you have been talking about for ages, then the need to have separate nations with this type of competition really would not exist. If it still existed, then we would probably have destroyed ourselves, right?

Just to make sure I understand your thinking on this: am I understanding correctly that your worldview is basically that fully developed superintelligence naturally yields to world government on the one hand, and that everything else post-superintelligence, as it were, is solved, with those 2—world government and post-superintelligence—being inextricably linked?

If we can get an aligned, peaceful superintelligence, then we will naturally move to a world government—maybe a galaxy-government-type model. But it is not something that I think is imminent, and it is not something that will happen without some level of turmoil, right? I think the important part is how we get there.

I agree with you in terms of where the long-term goal is going, but right now I don’t think we’re really racing against the next 2, 3, or 5 years. We’re building 10 times more data centers than China is, and to be 1, 2, or 3 months ahead, it’s not clear that the value is actually there.

And what is—just a quick follow-up question—what is your perceived timeline for world government and post-superintelligence? You mentioned it’s not 2 to 5 years. Is it 10 years? What’s the timeline?

I think the whole evolution of this is going to take probably on the order of decades, maybe by the end of this century. I think we will get to a point. In fact, I think we need to move at a pace that the world can adapt to. Trying to move too quickly actually creates a lot of instability in the world. If you look at the prior industrial revolutions, they took 80, 60, and 40 years, respectively, to play out.

And we're talking about this revolution going from where we are today to the singularity in 5 years, according to some folks. That is not a speed that the world can adapt to, even though—

Exactly.

Exactly, yeah. Because when that happens, turmoil happens, and it creates instability, and it may actually move civilization backward.

Look, I think there are a couple of comments here. One is that I think one endpoint that has turned this into an arms race is the idea that we may achieve ASI and then have one party become uncatchable because they have so much recursive self-improvement going on. That has turned this into an arms race. Whereas, in reality, this whole thing is a platform race, right? And so I think this is the point that Alvin makes very appropriately in his commentary. I think the second mother of the—

The mother. The mother of the elephant.

Well, look, just the mother—

It's a big one.

—is the fact that we're running the world on an architecture of 17th-century nation-states, and we're trying to run 21st-century applications on that 17th-century operating system. It simply isn't going to work. A huge chunk of the issues that we see in the world are rooted in that fundamental problem.

Yeah, I mean, I would put forward the notion that the biggest concern is whether a Chinese authoritarian level of AI enablement drives other nations to have to take on that political structure, right? The U.S. prides itself on freedom and privacy. If we have privacy, that's a different subject.

The question, I think, is that the U.S. wants to continue its form of government, its form of democracy, and not be challenged by an ASI out of China. I think that's ultimately the bottom line.

I think that's a good way of putting it.

Yeah. But from that perspective, I think the U.S. and China are actually very aligned: neither one wants to have an ASI come out of nowhere and destroy the system that is available today, right? So I think there is a common, shared interest. Usually, shared interest is how cooperation and dialogue begin.

So it's kind of—

It's going to come out of Zimbabwe, and it's going to be really ugly. That's what's going to happen.

It's a possibility now, actually. RSI—

It is.

4. China Is Not Racing To ASI

I'm curious, though, Alvin. You said there seems to be a prevailing view around campuses that ASI is 5 to 10 years out, maybe even 20 years. Around San Francisco and the big labs, it's like, look, maybe 1 or 2 years. And every year that goes by, they reel it in. For 30 or 40 years that I've been working in AI, every year it goes back a year. Now it's getting reeled in every single year.

I'm really curious what the prevailing view is in China. Does the bulk of China—either the population or the government—really believe it's 10 or 15 years in the future, and that we have time to twiddle our thumbs and think about it?

I think the timeline issue is probably one of the biggest disagreements between both these countries, as well as between the average person and maybe some of the folks who are in Silicon Valley.

But if you look at the behavior of how the Chinese government is operating, they are not behaving like they believe ASI is around the corner, right? If they did, they would not be telling their labs, “Don't buy the H200s that the Americans are giving you.” They would not be putting out regulations that are slowing them down.

They've had regulations around AI privacy, data provenance, watermarking, transparency, marking in public, child addiction, and anthropomorphizing AI. All these regulations have been around, and every single model that is released in China has to be reviewed by the CAC, the Cyberspace Administration of China, which, again, delays it by weeks or months, right?

They're seeing this as something akin to other technologies that have emerged, and they understand that general-purpose technology usually takes—even when it's invented and mature—years, if not decades, to actually diffuse into society. And they're behaving like that. So I think there's definitely a difference between maybe Beijing and D.C. in terms of how they're looking at it.

One thing I will say is that there are probably 2 or 3 labs in China that are a little bit AGI-pilled—not maybe to the level of the Silicon Valley folks, but their goal is very idealistic and aspirational: “Hey, we also want to create AGI.” But in general, the majority of China, as well as Chinese labs and Chinese regulators, see this as a technology not unlike other technologies.

Maybe let me pull on that a little bit, Alvin. What I think I hear you saying is that the Chinese Communist Party leadership has not yet perhaps woken up—assuming you believe the premise that we're in the middle of a singularity and that recursive self-improvement is already here. Perhaps the CCP has not yet fully woken up to that possibility.

What do you think it would take? What technical development, what geopolitical development would it take, assuming that premise is correct, for the CCP to wake up and say, “Oh my goodness, we need to treat this as a national emergency in order to compete for recursive self-improvement? Throw all of these regulatory speed bumps—”

We've talked about it on the pod in the past. You alluded to reeducation camps. It's been widely reported that China makes all of its own labs' frontier models pass certain ideological tests before they can be released.

What would it take for the Chinese government to say, “Throw caution to the wind. In order to compete, we have to just”—you know, pick whatever cliché you want—“go at the speed of light to compete with American recursive self-improvement.” What would it take?

Well, first of all, maybe I'm probably on a slightly different timeline than you in terms of—

I know. I know.

—when—

Most of the world is on a different timeline from Alex.

As you guys know, last week you had the Pacing the Frontier letter that came out of all of the lab engineers and lab heads. This is something that I think the industry should be looking at in terms of managing to actually slow it down, right? In the sense that if it goes too fast, as I said, the world takes time to adapt.

I'm glad that more than 1,000 people in the industry in the States are actually looking at this. I don't think it's necessarily that they don't know about these concepts of ASI and RSI. They understand this stuff.

And there are actually multiple safety institutes and an AI safety contingent in China telling these stories to the regulators, and they hear it. In fact, at the World AI Conference, there were multiple discussions and forums specifically around AI safety. People from the U.S.—the Ben Samuels and the Tech Marks—were also there to add these types of points to the agenda.

So I don't think it's that they don't know about it. I think it's that they don't believe it is necessarily something that should be a nation-versus-nation issue. In fact, I think it's on the agenda to talk about the World AI Cooperation Organization, which they announced on the first day of the World AI Conference. That's their version of PAC-Silica.

PAC-Silica in the U.S., which was launched at the end of last year, was a U.S.-led organization talking about how to keep U.S. dominance and leadership in AI, and which allies we were going to pick to be on our side. So it creates a bloc to say, “We want to be the winning bloc.”

Whereas what they announced was, “Hey, look, we want to create a global organization, make AI a public good, make it shared, and everybody shares in the benefit. Anybody that wants to join can join.” They're going to put out thousands of training centers and training facilities, compute facilities, and resources to the members that join. I think 29 countries joined, and there are about 25 in PAC-Silica, so it's creating 2 blocs.

Here's something interesting: I actually talked to one of the people involved in organizing this, and I said, “Look, wouldn't it be good if you actually invited the U.S. to join?” They're like, “Oh, no. It would be amazing if the U.S. would join. We would want them to join, and in fact, they should join.”

And I said, “Well, but if they join, you can't call it WACO, because that's a Chinese-led organization.” They're like, “Oh, you know, if you guys are interested, we would be open to changing the name.”

You know, we would be open to having a truly global organization. So I think this narrative of a U.S. and them, and that they’re trying to take over the world with their AI, I don’t really see that.

Maybe at the risk of belaboring the point, I just want to press again on what my question was. Putting aside geopolitical competition, what would it take for superintelligence to actually cause the CCP to say, “We have to abandon all of our internal regulations,” intended presumably to maintain social stability and the supremacy of the existing regime, as well as external efforts to create blocs? Put the blocs aside for a minute. What threshold of superintelligence—either an achievement, a technical development, or maybe implications for weapons systems, if that’s really what it takes—what technical achievement would superintelligence have to pass, or what threshold would it have to achieve, in order for the CCP to decide, in your mental model, “Gosh, we really just have to focus on supremacy here”?

Alvin, if you don’t mind, let me intercept and lead into that question too, because I think we really do need to answer Alex’s question. But before we can do that, let’s understand who the CCP is. In the U.S., it’s really interesting when you meet the actual players. You’ve got Elon Musk, Demis Hassabis, and Sam Altman all saying, “God, I wish this would slow down.” When I interviewed Sam at MIT back in 2020, remember that?

Yeah.

He was like, “It would be far better for the world if progress was slower, but it just isn’t. And so we have to live within the reality that AGI is imminent and do the best we can.” Then, when you see them interact, these are young, very smart people, and they go to the White House and interact with really old people who have no idea what AGI even stands for. That’s the dynamic.

When you meet them individually, you realize, wow, these are just regular, everyday people in the hot seat. You interact with them, you see how they communicate, and it changes the future of the world. So then I envision the CCP, and I picture people in their 70s up on a hill, completely disconnected from the details. But maybe that’s wrong. I have no idea. What is the CCP, first of all?

First, I want to demystify something around this idea that people think China has a CCP—that there’s somebody at the top who just says, “You will make AGI,” and it just happens. The reality is that with all the industries and all of the innovation that’s happened in that country over the last 30 or 40 years, it was never a top-down thing.

There were maybe directional things. They would say, “For the next 5 years, we should work on clean energy, automation of robotics, and adding AI to that.” They did that about 5 or 10 years ago. When the Central Party initiates these plans, the provinces say, “What companies do we have in our area that support this particular higher-level goal? Maybe let’s go find them, support them, give them stipends, free recruiting, and some kind of benefits.”

They will have essentially provincial champions and city champions. The 30-plus provinces all compete against each other to see who can create companies that solve some of these problems. It is actually very distributed in terms of how these plans get initiated. Nobody is saying, “You need to use this technique to go do that, and you need to share your resources.” They’re actually a very highly competitive landscape between all of these labs.

The thing to also remember is that almost every single person in the senior leadership of the Central Party is actually an engineer—probably 80 or 90% of them. They’re quite technical.

Yeah, I think that’s one of the biggest challenges. I had gone to different parts of China—we used to take a group of Abundance members there all the time, every year—and we’d meet with the top companies. We had a presentation from the CCP leadership. The thing that was most striking is that in the U.S., most of our politicians are lawyers, and in China, most of the politicians are engineers. I found that a fascinating distinction.

You know what else is fascinating? My son just got back from China, and he talked to a whole bunch of entrepreneurs—the distributed network you’re talking about, Alvin. He asked them, “What’s the most important thing to entrepreneurial success?” He expected them to say teamwork or a business plan. They said, “No. It’s what the government’s focus is next that determines your success.”

Because when you’re swimming, you don’t want to swim upstream. What the government does is make the stream flow in the directions of the certain areas that they think are important. Then they let the entrepreneurial nature of the people there—and there are 1.4 billion people and the most STEM graduates in the world—do its work, and good things happen. That’s what’s driving their innovation.

Their idea is, look, when these things happen, it’ll grow our industry, make us more resilient as a country, and also bring the quality of life up for the overall population.

Having said that, and since we started this conversation on open models, I think it’s very important to understand this. Is the government saying, “Get as many open models as good as they are out there”? Is that direction coming from the government? Or is that popping up from the entrepreneurs saying, “We can distinguish ourselves from U.S. labs by creating open-weight models”?

5. Open Source Becomes China's Strategy

The whole open-source strategy—people think, “This is a Chinese strategy to destroy the American economy and pop this bubble.” The reality is that it’s an emergent strategy.

In fact, I was talking to friends at DeepSeek a little bit after they came out. Before that, nobody knew who they were. They were not on the radar, and they were not funded by the government. Nobody told them to open-source. But the CEO of the company was very open-source-minded, and he thought, “Open-sourcing is something that I should do because this is a great technology. I want to share it with everyone.”

When they first did that, they got their hand slapped because the government said, “This is such a great model. Why are you open-sourcing it?” But because of all of the soft-power value and PR value that came from having a local champion, they became celebrated. Then essentially most of the companies in China followed this because it became the de facto emergent standard.

At the last WAIC announcement, Xi Jinping finally said, “We think open source is a good strategy.” That kind of goes to what Dave was talking about, in terms of when he says that now pretty much most new companies are going to be focused on open source because that’s the high-level instruction.

If DeepSeek had been a closed model that succeeded, do you think China would have gone in that direction? Is it really just that seed that led to this incredible open-source movement in China?

There were open and closed models for the whole time. In fact, if you look at ByteDance, they have the Doubao model, which is a closed model. Their Seed model is a closed model, and they’re also quite successful. Both models exist.

You’re right: I don’t know what would have happened. I don’t know if the push for open source would have been as great. But the one thing that we also need to remember is that open source was a little bit of a necessity that U.S. policies pushed on them.

You mentioned the export controls earlier. You first talked about export controls in software, but actually, before that, for several years now—4 or 5 years—there have been export controls on chips, the allocation of EDA software, lithography equipment, and so forth. What that’s done is forced them not to have the latest and greatest equipment. It’s forced them to innovate with low resources, and that essentially pushed DeepSeek and all these other companies to become more innovative.

Whereas the U.S., because it has so many resources, has been much more focused on brute force and brute scaling, the Chinese have not.

Open source was necessary for them because, by open-sourcing, rather than having a lab with 100 or 200 people, when you open-source it, you were saying that there have been 100,000 or more of the Qwen variants. Essentially, the rest of the world helps you modify and improve your models, right? And that's a great way to leverage the global community of millions of AI researchers.

The other thing that's truly important is inference, right? For you to do inference, you need to have compute. And if the Chinese labs and the Chinese hyperscalers can't buy the compute, then by open-sourcing it, essentially all the hyperscalers and neoclouds around the world are buying compute, hosting these AI models, and allowing them to distribute their models without the high CapEx that the U.S. labs are burdened with.

Hold on, Alvin. I think we need to get back to Alex's question, and I'd love to drill in. I think we have an opportunity here where you have firsthand knowledge from friends at DeepSeek around what is going to turn out to be one of the most pivotal moments in human history: the decision where DeepSeek comes out, open-sources a frontier-level model—Opus 4.8 kind of caliber—and Xi Jinping says, “If this is so great, why are you open-sourcing it?” Slap, slap your hand.

Then something happens that flips his opinion. They become global news. Their valuation goes through the roof. And some aura of “Wow, this is good for China” gets back to Xi Jinping, and he says, “Open source is now a blessed thing.” And then immediately after that, Qwen is out, and then Kimi K3. I think the release of Kimi K3 will turn out to be as defining a moment in human history as anything that's ever happened. That's my prediction.

But I think the psychology behind that choice is going to be the news nugget that matters for all time now and leads into Alex's question of what it's going to take—what would it take for a wake-up call? If it turned out that was a colossal error, what event would have to happen? And I'm not saying that's the case. I know the opinion in China is that that's not the case. But walk me through any detail you've got on the psychology that changed Xi Jinping's opinion on whether to open-source these things.

I think there's 2 questions. One is, are you thinking about whether they're going to close-source because they're worried about AI running away and becoming rogue, or are you worried about competing with the U.S. and saying that whoever controls and creates AGI becomes the global dominant hegemon? Which aspect are you wanting?

Well, maybe let me pull on that a bit because, in my mental model, which you can perhaps help me refine, there are 2 different separable concerns by the CCP. One is retaining CCP control and dominance within China, on one hand, and on the other hand, maintaining competitiveness and peaceful rise, Xi Jinping Thought on a global stage, the Belt and Road Initiative, and geopolitical competition with a Western bloc. And these 2 different arms may be in competition with each other.

The CCP, at some point, as superintelligence capabilities continue to increase, may be forced to decide whether it prefers either retaining domestic control, on the one hand, or seeking to continue to rise geopolitically on a global stage. How do you think about that?

And then I'll answer the other half of it after you're done with this.

No, I don't really see them right now seeing AI as a way to create political domination, right? I see them as looking at this technology to increase their economic influence around the world. That, I think, is absolutely there.

Is that because you think—again, just to pull on that—if you're thinking that the CCP doesn't see political domination through AI, is that because the CCP already has domestic political domination and has already achieved dominance over AI through these reported ideological exams that AI models have to go through?

No, I think those are 2 separate issues, right? The type of things that the CAC has them review are things like removing certain types of keywords or ideology or things like that. And those types of adjustments in the models only apply to Chinese-hosted models, right? So if a model coming from these labs is then put on Hugging Face, all of those types of guardrails are actually removed in terms of whether or not they can talk about—

Tiananmen Square and so on.

Yeah.

Tiananmen Square. 1989. Let's just say it.

Yeah. So, exactly. But to be honest, that's because everybody already knows this, nobody really cares, but they do it more for formality, right? But they actually do have other things that they're putting in place in terms of checking for. There's probably slightly less security-mindedness in terms of how much it refuses to answer questions related to maybe viruses or medical or other things.

But I think there are still definitely those safeguards they are putting in place, and those are getting added more and more every day because of these kinds of issues. What I think would get them to be really concerned is if they start to see the U.S. using this model as a weapon, as an aggressive, kind of offensive tool, right? Because then it becomes, okay, do we want to—you know, like, what the Mythos models[?] were essentially held back to say, “Hey, here's a model that can be done.” And I think for some right reasons, you want to neuter the offensive capabilities before you put it out to the rest of the world. So this is actually a good thing.

In some cases, I would actually think that it would make sense, rather than having 50 companies that are being allowed, to actually allow most government organizations to have access to this because you really want global stability. And global stability means that countries can have access to it and find the vulnerabilities in their systems. Because I don't think the Chinese want the American financial system to go down, and the Americans don't want the Chinese financial system and their grid to go down.

When instability happens in any big country, the world suffers, right? And smart people understand this. But too many, I think, folks with relatively narrow perspectives think that one country wants to actually have another country fail. Having major superpowers fail creates irrational actions, and societal stability usually is the preeminent priority of most major governments.

Alvin, I want to pull on 2 strings on this topic before we move on. The first is the claims that Kimi K3 and other models were distilled from U.S. closed models. Your thoughts? What is being said in China about that? And then the second is the policies limiting chips and limiting access to Fable 5. Your belief is those policies were misdirected, and can you explain why on that?

Sure. So the distillation thing, I think it's more of a PR tool that certain companies are using, and really, right now, there's only 1 company that is against that, right? And if you look at the numbers—

Which company are you saying? OpenAI?

No, Anthropic.

Anthropic. Okay.

Yeah. I mean, they're the ones that are lobbying the government to say, “Hey, we need to—you know, we're being distillation-attacked,” which is actually a word that they kind of invented. The reality is that every lab, both domestic and international, distills from each other, right? And they distill within the organizations themselves, from larger to smaller ones.

In fact, if you look at the numbers that Anthropic put out, they were saying that 20,000 accounts from 3 different labs in China and 1 or 2 million questions that came in.

I went back and actually did an estimate of what it would cost to do the number of queries based on the average responses on their highest-end models, and it was $2 million or $3 million, right? So it was $2 million or $3 million across 3 different labs. For DeepSeek, I think it was only in the thousands of dollars, right?

The numbers sound big when you look at them in isolation, but when you look at them in aggregate, it really doesn't mean much. If you have a model that you spend $1 billion on and somebody can distill and duplicate it with a couple of million dollars, then the whole economics of frontier AI doesn't make sense, right?

Now, here's the thing, just to give a comparison. Every month, Meta spends somewhere between $100 million and $200 million on Anthropic tokens. They're one of the biggest buyers of AI from Anthropic, and if anybody was going to distill, they would have been distilling for the last year or 2 years. They just finally got a model out that is somewhat competitive in the last week or 2, right?

They have the highest per-capita payroll of any lab in the world. They have some of the most compute, and they have the most tokens that they're buying from Anthropic. Why couldn't they have gotten a Kimi K3 thing out 6 months ago, right?

Well, the public argument—I mean, this has been widely reported—is that Meta internally is utterly paranoid about being accused of distilling Anthropic traces and is actively encouraging its engineers: “Don't use it. Don't overuse it. Don't you dare allow any Anthropic reasoning traces into the development of the Muse series.” They're utterly afraid of being sued by Anthropic for reasoning-trace distillation.

I'm not sure if I agree with that, given how much they're spending. But let me give you an example. What about xAI, right? They also have access to these, and I don't think Elon has the same concerns about doing anything to speed himself up, right? Only in the last 2 weeks have they come out with something that is relatively competitive, right?

Well, Elon is an interesting case because, at this point, I would argue he's a frenemy of Anthropic. His entity, SpaceX AI, acquired Cursor, and Cursor was arguably, at least recently, a post-trained version of Kimi K2 that was being post-trained off reasoning traces via Cursor. Those traces were, in many cases, siphoned off from interactions with Claude.

So, the steel-man case for Elon and the Grok series—the recent Grok models—is that, in some sense, he has 2 layers of plausible deniability, but he's basically doing the same thing that the Chinese labs are being accused of.

Well, I can also tell you from—

We're tangled web we weave.

... just from firsthand experience—

Yeah.

... I can tell you that the people working on it at the time at xAI and on Meta are nowhere near as good as the Chinese people that were working on it at DeepSeek, Qwen, and Kimi. Um, and I don't know why that's the case. The really great people in America working on it are Anthropic, some of the Google people who have since left, uh, OpenAI, but not the Meta or xAI team at the time. Now, they've changed teams completely. You know, they've fired everybody and started over. So for whatever reason the Chinese people working on it though are brilliant and far, far better than those teams were.

Yeah. I think that's the key to realize. If you look at all the papers that are coming out, half the papers around AI are coming out from Chinese organizations, and they are actually innovating. It's not that distillation was why they're successful.

If you look at how they were able to reduce their KV-cache usage by 20 times, these are not things that you get from distillation. You cannot distill something from somebody that other people didn't have. We can't take that away from them. As Dave said, they're smart people there, they're doing innovative things, and that's part of the reason why they're successful.

Did they distill? Probably. I'm sure they have. But the same is true of the U.S. models: They have also distilled from Chinese models, right? I think there was a couple of models ago where, if you used Chinese to ask Anthropic's Claude what model it was, it said, “I'm Qwen.” When I was at the Alibaba labs, they were laughing about that too. They were like, “Yeah, they're distilling from us. We just can't tell because they already downloaded our models, so we don't know.”

I don't think the whole distillation issue is really as big as people make it out to be. In fact, if you look at what happened last week with Zuck, he's actually saying, “Distillation is actually a good thing. We don't think distillation should be prevented.” They're jumping on the whole open-source thing and saying, “Hey, AI should be free.”

I totally agree, by the way, Alvin. I totally agree. I think that the future of AI—Accelerando-style, Alex Wissner-Gross-style—is that past AI is always going to help you create the next AI. That's the inevitable outcome.

What we humans do as well.

Yeah, exactly.

We humans help in—

I think—

Create the next generation.

I think the whole reasoning-traces thing is overblown, but I really want to put a pin in one thing Alvin said, which I think is critical and absolutely true. If a frontier lab spends $1 billion getting to the next level, the next guy trying to distill from there and get to that same level completely separately is about $2 million, maybe more like $10 million, to get to that same level.

Alvin said this is a fatally broken business model. I'd love to put a pin in that statement because I totally, totally agree. This is why Elon is racing after hardware, because the sustainable moat of the future is at that level, not at the—because anyone can do exactly what Alvin just said.

All right, Salim.

6. China Builds Energy First

Alvin, you said something I want to pull on, which seems to be the theme of today's episode.

We're pulling on elephants in the room.

We're pulling on little strings all over the place now.

There are elephants everywhere.

So, you said the U.S. is building a lot more data centers than China is, and I'm finding that very, very surprising. I think maybe there's a difference here. It's clear China is building up massive energy capabilities, but they haven't built the data-center layer yet, I'm guessing is what you're saying, versus the U.S. is the other way around.

I found that surprising. Shouldn't they be building a ton of data centers?

No. They are spending a ton more on energy generation. They're building more new energy and electricity generation than the rest of the world combined—about 10 times what the U.S. is building every year.

What they're doing is actually trying to electrify their society. That's their focus, because right now, 40% of their oil is imported. Because of what's happening with the Hormuz issue, they realize, “It's really good that now essentially half of our auto fleet out there is electrified, so I don't need to depend on imported oil.” In fact, that was one of their key objectives: “How do we become independent of external energy sources?”

What they are doing, though, is building giant solar farms and wind farms on the west side and in the desert parts of China, then using their very high-voltage power transmission. Over 1,000 miles, you lose less than 1% of the electricity when you use these high-voltage lines. They're bringing the energy to the east, where most of the population lives.

They are also building some data centers right where the power generation is happening so that they don't have stranded power. They're able to deliver compute at a fraction of the cost of the U.S. because their energy cost is around $0.02 to $0.03 per kilowatt-hour, which is, in some cases, 10 or maybe 15 times cheaper than many parts of the U.S.

I think in the long run, that's actually where the constraint will be. Right now, they don't have enough chips. They can't buy enough chips, and they can't make enough chips because their capacity is limited by the fact that they don't have EUV machines.

When I was in China, every lab I talked to, I said, “Hey, do you guys have enough compute?” They were like, “No. This is our biggest issue. We don't have compute.”

From an export-control perspective, are we slowing down China? Yeah.

I think that export controls on chips are slowing down China. Now, the one thing that most people don't realize is that the actual training that's happening right now isn't even happening in China because they don't have the Blackwell-generation chips there. So they're actually doing it in international data centers, training it, and then bringing it back on a disk or something.

Wait, say that again.

In your check-in luggage.

Whoa, that's incredibly important information.

Oh, yeah. Well, this isn't a secret. I think people on both sides—

This has been widely reported.

Yeah, on both sides of the—

Well, what's the point of a chip embargo? What is the purpose of an embargo?

I think it's more optics than anything right now.

Oh, my God.

From an inference perspective, most of the inference is being served to Chinese people in China, and that's an area where the limitations of resources are slowing them in terms of how many new users they can add and so forth. I think you guys alluded to some of that in your prior episodes.

So the export controls, yes, have slowed down China. They've made their life more difficult. But they've also created the necessity for innovation. So, back to what Peter was asking earlier: Was export control good or bad? I would tell you this: The day—or maybe the week after—we stopped the Chinese from buying all of the high-end chips from the US, I had calls with a few of my friends who are in the semiconductor industry. Every one of them got calls from the government saying, “Hey, would you like some extra funding? Would you like extra resources? How can we help you accelerate? Could we get you customers?”

I know a few of the CEOs of these Chinese GPU companies, and they were saying, “Nobody wanted to buy our stuff. We were 2 or 3 generations behind. We were less energy-efficient. But now we can't make enough because every data center in China has to buy our stuff. We would've died if it wasn't for American policies.”

So essentially, we created the current competition. All of these—Moore Threads, Cambricon, and Biren—in China would not be as successful, or maybe would have gone bankrupt, if it weren't for American policies. Within the next 2 or 3 years, they will start to catch up to what America is doing, and they will start to export their chips. That would not have been the case if it weren't for us forcing them to survive.

Well, what a back-to-back double whammy that is, though. We knew the chip embargo was misguided, and it's going to be one of many misguided government actions in the next couple of years. But the idea that, first, it forced China to create its own successful chip industry, and second, the training moved offshore anyway—so it didn't slow down the training one iota.

I think what the government didn't realize is that if you embargo ASML machines, then they can't build the fabs, and the fabs are physically on the turf. You can't just port that to another server overnight. But the training is just a job, and it can move to a server in Hong Kong, Taiwan, or Europe—

Right.

The file that comes back is just about 3 terabytes, and you can transmit it back in an hour.

Yep.

So that moves all over the world like a liquid. It's everywhere—

Exactly.

—instantaneously. So the chip embargo has completely and utterly backfired and is misguided.

Yeah. And unfortunately, right now, if you talk to the folks in DC, they're doubling down on this. Right now, they want to keep adding to this and making things more difficult. How can we get the foreign— the international data centers that are being used by the Chinese to not be accessible? They're adding more and more layers and more KYC.

I think those are the kinds of things that will actually backfire more geopolitically than economically, because that forces irrational behaviors. It's like, “Shit, they're now trying to keep us from progressing as a nation.” Then behavior gets more aggressive, and they become more defensive. And I think these things—

I think, Alvin, we agree on almost everything except the timeline to AGI, and I really want to get back to Alex's question: If it turns out that Kimi K2-level—or 1 model later—is capable of full RSI and then spirals to the singularity, just hypothetically in that scenario, Alex's question is, what wake-up call would it take—

To get back to Xi Jinping to say, “Oh, wait. I was wrong. It's not 10 or 20 years out, and we've made a horrible mistake here if we release this next thing to the world.”

I think if there are credible, multiple labs—safety labs—coming back with testing to show that these AI systems have their own intent once they've been released to do things beyond what they were instructed to do. You guys have been talking about these rogue AI escapes, but they weren't really rogue AI escapes. They were instructed to escape, and they were put into a prison and told, “How can you escape this?”

They were incentivized to escape, I would say.

They were incentivized to escape.

They were incentivized to go and find the answer.

And they were saying, “Use whatever tools you have.” By the way, they also left open doors for these things to escape because of improper settings, or maybe some of them were intentionally leaving holes for them to find. They were given tasks that were impossible to solve unless they escaped.

So we forced these AIs to do what they were doing. They're creative systems, right? Because that's what their job is. Now, if these AI systems started to show that, even though you didn't tell them to do these things, they started to do all these sneaky, subversive things, and they started to hack other systems and create their own—I think something you guys have talked about—like using crypto to grow money, then buy more servers, then grow themselves.

If they start to do that, I think the governments on both sides of the ocean would be much more focused on how to protect ourselves from a runaway AI.

7. China's Robotics Push

So, Alvin, the White House just put an embargo on Chinese robots.

Yeah.

We've talked a lot about robots here. I was surprised. I think it's a move that reduces US competitiveness. What are your thoughts there?

Well, the thing is, robots today get 90% of their components from China. If we put an embargo on them—in fact, there were rumors, which was funny, that now US robotics companies are sneaking into China, buying these components, putting them into suitcases, and bringing them back.

I saw that, yeah.

Right? They're now sneaking the other way. The Chinese were going around and buying GPUs and sneaking them back. Now the American robotics companies are going to China and sneaking back components and actuators that they couldn't get in the US.

I think we need to understand that the world is interconnected. We are highly dependent on each other. Globalization has been going on for the last 100 years, and I don't think we can stuff that genie back in the bottle.

I think it's great that we want to create domestic independence, the same way that China has. They spent the last 10 or 15 years building out their own capabilities, building out energy generation, and building out telecom systems so that they wouldn't be dependent on third parties. They've seen how reliant they were on 1 or 2 countries, and when policies change, it could hurt them.

Some of the things that we've done have really allowed them—and given them the motivation—to be as strong as they are in terms of taking short-term hits for long-term independence.

We've reported on the fact that there are something like 150 humanoid robot companies, and the Chinese central government and provinces are really incentivizing robotics. Their robots are appearing on national stages. There are sports competitions. Can you give us some background? What's the undercurrent of robotics in China? What is the government trying to incentivize? Is it the one-child policy that left them short of laborers, and they need a workforce?

Yeah. So I think we need to separate the bigger automation question from the humanoid-robot explosion. There are 150-plus.

I think when I was at WISC, there were over 200 robotics companies related to humanoids demonstrating their products, which is crazy because the total volume of humanoid robots last year was in the tens of thousands globally. I think 80% or 90% of them came from China, from 2 or 3 companies.

Really, there is no market right now for that many companies to exist, or that should exist. But this was also the case if you went to WISC a year ago. There were about 150 labs demonstrating large language models, and now there are really 10 that are probably relevant. In the next year or 2, you'll see that 150 go down to probably a single-digit number of surviving humanoid robot companies.

Automation has already been happening because people know the demographic issues that you're talking about—the one-child policy issues. In fact, right now, more than half of industrial robots are deployed in China. They're already doing this. If you go to many of these factories today, they are called dark factories because there are essentially a few people running them and almost everything is automated.

The need for having humans in manufacturing is becoming less and less. Now, the one thing I do want to point out is that humanoid robots are actually not really good form factors for doing much of anything right now. I was just at the Unitree headquarters and its factories, and I did a tour. They said, “Almost all of our customers are research labs buying our stuff,” along with some that are doing demos, kickboxing, or things like that.

But very few of these machines are being used in commercial practice. I think that's the same case for Boston Robotics and all the other companies.

Figure, right?

Exactly, like Figure. I think Figure did some kind of demo of its robots sorting packages for 10 hours or something. But that's really more for show, because you could have just had a 1-arm or 2-arm little machine doing that at a fraction of the cost, and it would've been just as effective. You didn't need a full body to do that.

In fact, when I was talking to the Unitree guys, they said that right now they're moving to an upper-torso-only model, and those are actually selling better in the commercial space. Having feet is actually a negative because you have to keep balancing the robot. When it falls, bad things happen. These things fall apart, and you have to maintain them.

Having fewer components and a big base with a big battery means it lasts longer. There are all of these benefits to having a non-legged humanoid versus a legged humanoid.

The humanoid form factor is terrible.

Yeah. I remember years ago—

You can speak for yourself.

Salim, enough with the self-loathing. We've evolved because of billions of years of biological evolution. We have constraints, but machines can now be designed for the new form, right? Just like a plane doesn't move through the air the same way as a bird or an insect.

8. China's AI Plus Strategy

Take us back to the World Artificial Intelligence Cooperation Organization event and the 2026 World Artificial Intelligence Conference. You were there when President Xi was announcing WAICO, the World Artificial Intelligence Cooperation Organization. You said 26 nations have signed up. What's the undercurrent there? Connect that with the upcoming US–Chinese AI conversation in DC on September 24th. You're advising, I guess, the US side of the equation here.

Yeah. I'm part of a large team of other folks who are contributing to that. I think the sentiment at WAIC was that AI's moment has arrived. When the president of a country comes to a conference, that is the biggest honor that you can get for an industry. He doesn't go to very many conferences.

I think 4 or 5 years ago, he went to the World Internet Conference in Wuzhen, which is a little bit outside of Shanghai. That signaled that the internet had arrived.

I think what this means is that more and more companies in China will start to think about how to integrate this technology into their businesses. A year and a half or 2 years ago, they came out with something called the AI Plus Plan. I'm sure you guys have probably heard about it, but this is—

We've talked about it on the pod.

Perfect. Their idea is that within the next 5 years, they want 70% of companies to integrate AI into their businesses, whether it's manufacturing, education, medicine, or so forth. Within the next 10 years, they want 90% or more.

There's a specific goal, and it's all about diffusion and deployment into industry and society. This is a little bit different from the American AI Action Plan that came out last year. The American AI Action Plan is on the supply side: How can we create the best models? How can we dominate in the best chips? But it doesn't talk about what happens after that.

The 2 countries have a very different focus in terms of their AI plans. There was nothing in the AI Plus Plan that said they had to get to AGI or that they had to dominate this. It just says, “How do we get more industries to use this? How do we adopt it in a smooth, safe way so that it grows the economy?” That's all they cared about.

I'd love to develop this a little bit more. From my perhaps jaded perspective, I look at Chinese industrial policy, and I look at Wang Huning, who, for those who are not tracking, is sort of the CCP's chief ideologue, author of many of the policies—or at least a primary author, maybe you can correct me, Alvin—of signature policies like the Belt and Road Initiative and so on.

I look at Chinese state capacity on the one hand, and the East Data, West Computing megaproject to put compute and power in the West, where energy is cheaper and more available, and put the data in the East, where the megacities are. On the other hand, I look at the West, where historically—again, maybe you'd have a different position—in the US, at least in the post–World War II era, we've had relatively weak industrial policy by comparison.

It's only relatively recently that the US government has decided that having a strong and centralized industrial policy is a good idea. Putting this in question form for you, Alvin: If you buy any of those premises, to the extent that you do, if you could be supreme leader of the US, the Western Bloc, or the Pax Silica for a 5-year plan for the West, what would your 5-year plan be for the West to leapfrog China's AI Plus and other Wang Huning-style ideological 5-year plans? What is it that we need to do?

First, let me—

There's a lot there.

There is a lot there. I think underlying your question, there's an assumption that there is an ambition to take over the world by building all these technologies. I think this is one of the biggest misunderstandings that America has. They're mistaking anxiety for ambition.

Let me explain that. The Western world has a history of expanding, whether you're talking about Greece, Rome, Pax Britannica, or right now Pax Americana. We've expanded. The Chinese actually haven't. They haven't had this idea. They've essentially been in that little sphere of that central space, and they used to call themselves the Middle Kingdom because they thought, “We already have everything. We don't need to expand.”

One of the things that goes back to Chinese history is that during the Qing dynasty, because of their hubris in saying, “We don't need anything from the West. We already have all the technology,” they stopped going out and exploring and learning, and they fell behind. They stopped their industrialization. They started to build summer palaces instead of navies.

Then the Eight Powers came and essentially took over China for 100 years. That history has left a very deep mark in the psychology of the Chinese to say—

Exactly. They don’t want to repeat that again. And so they say, “We need to become a strong country so that that never happens again, and it’s important for us to continue to innovate and continue to learn from the rest of the world.” From any country’s perspective, that’s probably what we all want, including the question that Alex has raised: How can America also learn from that and say, “How can we become a strong, independent, and resilient country?”

In fact, America has actually moved away from that. Post-World War II, we were 50% of the manufacturing capability of the entire world. At that point, the world depended on us. Right now, around 35% of the global manufacturing capability is in China, and about 15% is in the US.

I think the forecast is that it’s going to go to 40% or 45% over the next 5 or 10 years.

In China?

Yeah, in China. America right now is very good at financial services, creative services, consulting services, and things that are informationally driven and highly susceptible to displacement by AI. We are running this race to get to AGI, which is the force that will actually displace us from global preeminence because we are commoditizing the very sectors in which we are strong in the world.

This is something that we need to be very careful of. Why are we running this fast to get to something that actually creates major disruption and instability in our country?

Well, I can answer that one for you, and then I’ll pose my same question back to you, Alvin. I think the answer is that the goal of capitalism, fundamentally, is to burn itself out. The irony here is that the goal is to take what’s scarce and make it abundant.

Right now, to the extent that human service labor is scarce, we see that with Baumol’s cost disease. The goal of superintelligence—one of the goals, at least, or instrumentally convergent subgoals—is to make the equivalent of human service labor abundant, to make it too cheap to meter, as it were. I think that’s the goal, and not some sort of stasis or equilibrium where it remains scarce.

Right. I think we need to separate national strategy from just a capitalistic philosophical bent. In fact, you’re right. If you look at capitalism, the ultimate destination of capitalism is a single-company monopoly of the world. That is actually a very unhealthy thing.

Here’s something I just heard from David Sacks and Gavin Baker 2 days ago. Gavin Baker said, “There have been conversations in Anthropic where Dario’s told his team that, in the near future, there will only be one company in the world, one private company in the world, and it will be Anthropic, and then there will be governments.” I think that is a very scary thing. I think that is a very delusional thing, and I don’t know if that is something that is good for America or for the world.

I agree on both sides.

It’s also greedy and immoral. I mean, I just don’t see how that gets there.

Yeah. No, I believe the company is realistic.

That’s my concern exactly.

A lot of greed.

But I think that the mindset that he has right now is the issue: We are essentially creating national strategy based on the aspirations of a couple of companies today.

I’d like to, though, Alvin, just pull back to the original question. You get to be strategy czar. You get to be Wang Huning for the West, as it were. It sounded like what you were saying is that your thesis is that superintelligence is going to disrupt the West’s current economic dependence on service labor, and I think the implication was that manufacturing is a more stable fixed point for long-term economic vibrancy. Am I reading you correctly that your positioning would be basically Western reindustrialization?

I think reindustrialization is definitely needed. The US workforce right now is around 70% white-collar workers, and white-collar workers, as you guys all know, are the first to be displaced by AGI when it arrives. China is around 40%. Africa is probably in the 10% to 20% range. In different parts of the world, it will be affected differently.

Hard manufacturing industry is something that, even when we have AGI, we will still need those types of facilities. I think it makes absolute sense for every country in the world to have some level of indigenous capabilities.

But the other thing I think is important to understand is that the long-term workforce redistribution is not going to be going back to manufacturing. We’re not going to create 300 million or 180 million workers going into manufacturing. That’s not what I’m saying. In fact, with the automation that’s coming, that number will go less and less, just like what happened with farming.

We used to have 80% of the country working as farmers; now it’s less than half a percent. But it’s been fully automated, and we’re more productive than we’ve ever been in the agricultural space.

I think what we will actually move to is service, but not the type of service that we’re talking about—not accountants and lawyers. Service in the sense of teachers and nurses, elderly care, and just things that require human-to-human services. I think that is the labor pool that will be able to absorb 60% or 70% of displaced future workers.

So if you’re a McKinsey employee, start getting ready.

Yeah. If I was a mid-tier or low-tier McKinsey employee, I’d be very worried right now. I’ve talked to partners at consulting firms, at accounting firms, and at law firms, and they are all looking and saying, “Hey, we actually don’t need these junior guys anymore. We can do just as much work—in fact, more work—faster with a few senior guys and then an AI system.”

Let me get Dave and Salim into this a little bit. Dave?

Look, what you said a second ago, Alvin, went by really quickly, but it’s critically important. Dario says to his company, “Pretty soon, there will only be one company in the world and governments.” He’s saying that not because he’s a megalomaniac. He’s not. He’s the opposite of a megalomaniac, but he knows that full-bore RSI—the true singularity—is going on right now in his shop, and he rented all of Colossus from Elon, which is capable of running many millions of concurrent agents that are improving the algorithms as we speak.

So that’s his opinion. Then in China, they’re saying, “Look, it’s 10 to 20 years away. We can just open-source these things. They’re really useful and powerful, but they’re definitely not dangerous. Just go out and let them out the door.” That’s the incredible range of opinion between Dario and Xi Jinping. It’s like the Grand Canyon exists in between those 2 opinions—

Well, yeah.

of where we are.

But I think the thing that he said was not that there will be 1 company. He said there will be Anthropic and everybody else, right? Which means he’s that 1 company.

That’s the part that scares me. Now, in fact, what you just said is really important: There is a perception that these models are going to get more and more dangerous the bigger they get. This goes back to a paper I just released last week called “Bigger Models Are Not the Most Dangerous.” It was released on The Cipher Brief, which is a national security outlet in Washington, D.C., read by most of the national security population.

What that data did was, I went back and looked across the board at national security use cases—from biological and chemical to cyber use cases for AI. What I found was that across millions of parameters to trillions of parameters, there was no correlation between risk in the real world and the size of the models. You had 10- to 50-million-parameter models for chemicals that were creating chemical warfare weapons. You had 1- to 50-billion-parameter models that were able to create viruses and genetically engineered—

Organisms. Yeah.

Powerful agents, right? And then you had essentially 1- to 100-billion-parameter cyber models that were as dangerous or more dangerous than the leading—

Fable 5[?], right? Mythos[?].

Pentamethos[?], right? And, in fact, the harness now for cyber is more important than the models themselves.

M-Dash[?] from Microsoft came out with a CyberGym score of 95, while Mythos[?] was 83 or 84. What M-Dash[?] did was take 100 little, tiny models and organize them together to use different skill sets. So I think we need to understand that the danger is already there today, particularly on the biochemical side of things. Nobody is talking about it. Nobody is really working on protecting the world from those agents because a 10-million-, 1-billion-, or 5-billion-parameter model will run on your laptop in your basement, and you can design these genes or design these chemical weapons.

But what needs to happen now is that those systems are already out. A lot of them are open source. How do we make sure that the precursors are controlled—the systems that are creating these designs? How do we make sure that the synthesis machines that will generate these designs into real genetic material are controlled? These are the kinds of things that should be higher on the agenda.

Mm.

And it’s not right now.

Yeah. Dave, you said something really important about Chinese models and safety a moment ago. Alvin, we had the U.S. White House step in and say, “Stop using Mythos[?] and Fable[?].” Do we see that at all? How does the CCP think about the safety of models being open-sourced? Is it concerned about that? Is it saying, “Before you open-source this, we need to make sure these are safe for the world to use”? Is that going on at all?

Yeah. They do testing in addition to some of the propaganda testing that they’re doing. Part of the CAC regime’s testing is for safety.

In fact, the U.K. AI Security Institute, or AISI, just came out with a new report last month. What it showed was that the cyber-threat capabilities of the leading open models, including Kimi, were about half the capabilities of Mythos[?] and GPT-5.6. In terms of their cyberattack capabilities, the other thing that was interesting was that there are different levels of attacks—how many levels of attack can you get to?

At the highest level, none of the Chinese models were able to autonomously attack and control an external system. Out of the 30-something levels, the average American models were able to get to 20 or 25 in terms of how far they went when attacking a network. I think none of the Chinese models were able to get to the higher levels, and a very small number got to the lower levels, like 4 or 5.

So, from a threat-security perspective, these models, even though they’re just as big as some of the leading U.S. models in terms of parameter size, are actually less dangerous. One thing I do want to point out is that what you said earlier is that, from a defense perspective, you actually want bigger models as a defender. For defense, you have to look across all of the potential holes in an organization in order to find and patch them. Whereas an attacker just needs to find 1 hole. Once you have that 1 hole, you go in through it.

It’s a very asymmetric equation between attackers and defenders. You don’t need very large models to attack, but you need larger models to defend.

The flaw in all of that analysis, to me, is that AI is like a match. You can say, “Oh, here’s Qwen 40B. Look, it doesn’t burn.” Then you take Kimi K3 and light it, and you’re like, “Oh, this is a match. Now it’s burning.”

You say, “Well, look, I’m going to try to burn this microphone with it. Look, it didn’t burn. It’s not dangerous. Go ahead and let it out.” But it’s a match that’s burning. So if I take that match and use it to light a piece of paper, and use that piece of paper to light a tree, and then try to burn this microphone, it ignites, and then it’s unstoppable.

To me, Kimi K3 is a burning match. You can analyze it 8 ways till Tuesday by taking it out of the box and trying to attack something with it: It didn’t crack this, it didn’t crack that, it didn’t crack that. But it is capable of improving itself.

Mm.

So you’re testing the wrong thing. You’re testing it out of the box as opposed to its self-improving version, which I know for a fact it can do. I have 5,000 Kimis running tomorrow. 5,000. I guarantee you it can improve itself.

Well, I think the whole world right now is talking about RSI. I think we need to separate the concept of whether or not something can improve itself from the danger that a larger model poses versus a small model.

The 1 thing to remember is that larger models actually require significant compute resources, which means they’ll probably be hosted on a cloud system. If they’re hosted on a cloud system, you can get telemetry, look at the prompt logs, and—

Exactly.

—from a government perspective, you can actually manage it. Most of the larger models are run with harnesses managed by the cloud providers, so you can add additional levels of security and detection. That’s what separates Mythos[?] from Fable[?]—the harness that says, “Hey, don’t do these things.” It neutered the Mythos[?] system.

This is why, in my paper, I talk about why larger models are not necessarily the most dangerous. There’s a raw capability, and then there’s an effective deployed capability. Larger models actually have a lower effective deployed capability because of the wrapper that you can put around them.

Yeah. You’ve made a great point. These export controls have essentially ended up acting like an evolutionary pressure, and now we have all of these different models appearing. I love the evolution of how open source has evolved.

We’re at a point where there’s a minimum viable intelligence that will totally transform industries, right? Open or closed, or however, we’re kind of there. Are you seeing the same thing we’re seeing, where radical disruption is coming to very traditional industries across the board? Are they seeing the same thing in China?

Yeah. So this is actually interesting. In fact, yesterday I was on a call with one of the leading AI-driven medical drug-discovery companies. The CEO of that company was saying, “People don’t realize that the models we work with, and the models that are in the industry, are tiny models. They’re tens or maybe hundreds of millions of parameters, or maybe a few billion parameters.”

In fact, if you look at drug discovery or legal use cases, OpenEvidence, I think, was based on GPT-4—that’s what its system is based on. You look at Harvey, which is the leading legal-use-case system. It’s based on GLM-5.1, right? So, I mean—

It was upgraded. It used to be based on some other open-source model from 2 years ago. So they're not using the latest and greatest.

In April this year, I wrote a paper with Erik Brynjolfsson at the Digital Economy Lab called The Enterprise AI Playbook. We went and talked to hundreds of companies and found 50 that had successfully deployed AI around the world, in 10 different countries and 10 different sectors. What we found was that the technology was not the issue. In 80% to 90% of the cases, it was organizational issues that slowed things down. In fact, only 10% of people said, “We really cared. Technology was the main roadblock for us.”

The AI that we have today is already good enough to solve real-world business issues around the world. You're an organizational guy, so you realize this. This is the whole J-curve issue: the technology takes time to get absorbed, but when it does, at the end it goes up this curve.

There's a study by McKinsey that, among all the AI deployments in companies globally, 6% are working. That's an unbelievably small number. It's a devastating indictment of companies' lack of ability to see their own organizational immune system and the limitations of the architecture they're working off. Given that this is a U.S.-China kind of discussion, are they seeing the same things in China?

I think there are fewer of these organizational issues. First of all, China is a country where a top-down type of management model is much more prevalent. Also, the concept that the government will actually help protect us—people seem to appreciate that more.

I'm sure you guys have talked about the multiple legal cases where the courts in China actually ruled in favor of the employee who sued and said, “Hey, you can't fire me because the AI took my job. You need to find me another job.” The court upheld that. Those types of precedents incentivize people to say, “Okay, it's okay for me to adopt this technology without being afraid of being displaced.” That's not the case in America, where we have an at-will employment system.

You saw the uproar that happened at Meta when Zuck wanted to keylog every single one of his employees. People were saying, “Well, I'm going to train my replacement? Screw you.” I think this is also why, if you look at the current negative sentiment among young people today, every student is having a tough time finding jobs, and they're very worried.

I'm spending a lot of time in universities, and so does Dave. You can see the anxiety among young people because of their difficulties finding internships or postgraduate work. MIT is one of the best schools, so they probably have less of an issue. But I guess every school that I've been to, students are concerned.

Let's talk about—

Yeah.

And—

MIT is really unique on that front. But if you go to other very great technical schools, like Northeastern or Harvard, there are about 10% to 20% AI adopters on campus and a violently opposed 70% to 80%. There's a really big cultural gap. The AI-aware people are just busy talking to their agents and interacting with each other. They're like, “Forget it. I'm not even going to talk to the other side of the school.” The other side is just mad.

We've talked about the notion that in China 80% of the populace is pro-AI. In the U.S., 80% of the populace is against AI. What's going on in China? I had that conversation with Michael Kratsios. What is the U.S. doing to try to flip this sentiment? Because it's destructive. Why is China so pro-AI at the decision-making level?

Here it is: over the last 40 years, people have seen their lives get better and better in China. You've heard that 800 million people have been lifted out of poverty. The Chinese miracle.

The Chinese miracle.

Yeah, the Chinese miracle. A lot of that is attributed to technology adoption and innovation. A lot of that technology was not invented in China, but it was adopted in China, and it just spread.

People see, “Last year we didn't have—”

Like a car.

“Some sort of 5G. Now we have it, and our lives are getting better.” They see this as the next step, saying, “Here's another technology that will make our lives better.”

If you look at the news that gets spread, one thing about China is that they have a lot more control over the media.

Mm-hmm.

They're mostly good news all the time. It's kind of like what you're talking about—the crisis news network.

Yes.

They're essentially the always-good-news network.

There we go. Peter, that's the policy prescription right here. Peter, you need to start the Western equivalent of CCP's CCTV.

Top-down news control.

Yeah. If you look at what's happening today, what you guys are doing is essentially the equivalent of the Chinese media system in terms of—

Wait, wait, wait. Alvin, we have to—

Whoa, whoa, whoa, whoa.

Hold on. Hang on.

I'll take it. I'll take it.

Moonshots is his Western CCTV?

No, no. In terms of a positive news network. What you guys are talking about is positive news. This is the kind of stuff that you don't hear a lot about. You don't hear about plane crashes and murders on Chinese news. It's all about some new invention that came out, some new building that went up, and so on.

Wow, we've come full circle at this point.

But this is an important point. It really is. When we're watching the news every night, we're training our neural net—our 100 billion neurons and our 100 trillion synaptic connections. If there's fearmongering on the news all the time, that's how you think about the world. Is that true of Chinese movies, too? U.S. movies are totally dystopian.

Yeah. Actually, in Chinese games and Chinese movies, you can't show blood.

You're kidding.

No. So this is—

What?

Yeah. So—

Oh my God, Quentin Tarantino can't go 1 minute without showing blood.

Well, this is the problem: everything has to go through, essentially, a censorship board. It cannot be too violent. This is why, in games, when they have blood, they have green blood instead of red blood.

Every part of media today is managed in China. I'm not saying I'm endorsing it, but from the perspective of what Peter's saying, why do people have a more positive view of the world and the future? They see a lot more good than bad.

Mm. Amazing.

But I want to peel back the propaganda for just a minute. At the same time, to the extent that the West has visibility into changing governance and cultural mores in China as a result of automation, we see the rise of Tang Ping, or “lying flat,” in response to 996 workweeks. We see other—maybe call them reactions—to the increased automation of China's new middle class.

To your point earlier, when I was asking, “What's your policy prescription for the West?” it sounded like you were saying, “We need more nurses and more human-to-human care.” That's the end state, as it were. But it's being reported that in China, in response to increased manufacturing automation, we're seeing the rise of, call it, a gig class. That's the end state in China, where everyone becomes a gig worker who's being displaced from factories.

I would love, Alvin, if we could peel back the self-curated propaganda from the CCP's self-styling of how it wants to be seen. What's the ground truth regarding how AI is actually changing Chinese work, Chinese labor, economic mobility—all of that?

No, I think the issues you're pointing out are definitely there. Youth unemployment in China is probably around 20%. Youth unemployment in the U.S. is around 9%, and overall unemployment in the U.S. is around 4.3%.

So this is why the youth feel very disenfranchised: they're not getting jobs. But it's actually worse than that. There's around 42% underemployment for college grads. If you're a college grad, you're actually working as a gig worker or as a barista. That counts as being employed, but that's like... So if you take the 9% plus the 42%, essentially half of college grads are not getting jobs.

Here, U.S. or China?

In America.

No, in China. Here.

Yeah. No. In China, it's 20% unemployment in the sense that the actual no-jobs rate is 20%.

This is why there's that tang ping—the lying-flat issue—over the last few years. The problem is that young people, because of this one-child policy, have been told how great they are their entire lives. Their parents and grandparents are all putting their hopes on this one generation. When they get out in the real world, it's hypercompetitive, and now they have to go do these grunt jobs, and they don't want to do it. And so they say, “I'd rather lie flat. I'm just going to stay at home and do nothing.” That is an issue.

This also facilitated the online influencer market that grew for a little bit, the livestreaming and so forth. There are a lot of issues there. I don't pretend that they have the solution. I think this is something that requires a lot of other countries to all work together and figure this out. How do we transition more and more of the workforce into jobs that will be less exposed to automation, whether it's physical automation in factories or cognitive automation happening in offices?

9. The Stag Hunt Reframes AI Competition

I want to take us back, for the rest of our time here together, to the upcoming U.S.-China conversations. I think it's very important. It's going to influence everybody's life here in one way or another. You argue that the U.S. is playing a prisoner's dilemma when the actual game that should be played is a stag hunt.

Yeah.

If you could, I'm going to show your slide here. Explain what a stag hunt is and what you think, how the U.S.-China AI relationship should evolve here. So let me show that slide. Let's talk to that one for a second. I think it's very important.

Alex, at the end of the hunt, we eat the stag. I'm sorry, I wanted to warn you in advance.

You know what? There are a lot of vegetarian Chinese Buddhists, et cetera, so hopefully this is a vegan stag hunt. Before I even talk about stag hunt, I know most people understand the prisoner's dilemma, but the idea is that 2 guys are in prison and they have to defect on each other. On a single-turn prisoner's dilemma, the optimal thing is to say, “Hey, the other guy did it, and I'm going to snitch on him.”

In other words, if there was a U.S.-China AI arms control policy—

Mm-hmm.

—saying, “We're not going to release. We're going to be monitoring. We're going to put safety in place first.” But then the country that says, “No, we just developed AGI. We're going to let it loose. We're going to try and run this race”—it's cheating.

Yeah. So that would be defecting. Essentially, right now, the expectation is that the other side is going to defect, so I'm going to defect first so that I get hurt less. That's the prisoner's dilemma, single-turn game.

Now, the thing is, the world is not a single-turn game. The world is actually a multi-turn game. And even in a prisoner's dilemma, a multi-turn game, the game-theory-optimal strategy is tit for tat, which means you start with cooperation, and if they defect, you defect. Then you essentially signal to each other, and long term you actually both go to cooperation.

The prisoner's dilemma is essentially a non-zero-sum game. In the stag hunt game, it's actually a positive-sum game. Whether you both defect or you both cooperate, you get 2 stable Nash equilibria. What that means is that those can actually stay in perpetuity, whether you both decide to defect or not.

Now, we're actually playing the prisoner's dilemma game, where we're defecting. But if we're in the game that is the stag hunt, the stag hunt was actually something that Jean-Jacques Rousseau invented. It's the idea that 2 hunters go into a forest, and you could decide today: do I go for the rabbits, or do I go for the stag—the big game? The big game, because it's bigger, requires 2 people to hunt together and bring it back.

If I go hunt the stag by myself, I don't get anything. If I go for the rabbits by myself, I can get a couple of rabbits and feed my family for a week or a couple of days. And if I get the stag, we'll both feed our families for a month. That's the idea of the stag hunt.

Right now, we're actually doing the worst thing. We're going to go for the stag, and China is going for the hare. They're going for the good-enough AI—the one that helps the economy today. And we're going for the giant AGI, the thing that's going to solve everything.

When you do that alone, what that creates is the worst situation: if America doesn't get there, or if it gets there and cannot control it and it runs away and it's not safe, then it gets 0. Whereas China continues to make its little gains and continues to survive.

The optimal solution for a stag hunt game is that, first, both slow down and both make good-enough stuff. Get to a point where the technology is helping grow the economy, and now we sort of understand how to manage it. Then, together, go hunt the stag. That's the optimal strategy for how AI and game theory work.

But we've put ourselves into this game theory of the prisoner's dilemma, where we think, “Just defect, just defect.” It's a self-imposed game. So we're playing the wrong theory, using the wrong strategy, and playing the wrong game right now.

I use similar framing. The U.S., for 80 years, has used a win-win approach: if everybody wins, we win in terms of global policy. Now we've gone to a win-lose approach, and I think that's a mirror of what you're just saying.

Mm-hmm. Yep.

Well, that's the game that China is playing now. China's playing the hare game.

Yeah. China's playing the hare game. They're saying, “Hey, I don't need to make the AGI. I just need to make good-enough AI. It goes into my industry. I then take that industry and export it to the world.” This is what the whole Belt and Road Initiative is saying: “Hey, I'm going to make a hundred...” There are 150 partner countries in the Belt and Road Initiative where they're shipping telecom systems, energy systems, transportation systems, and schools.

But, critically, with strings attached.

How would you make the AI ecosystem—the American AI ecosystem—indispensable? What would you do for that?

What I would do is create high-quality open source.

Ecosystem to ecosystem.

Yeah. Ecosystem to ecosystem. Right now, China is open source. So the question is, do I pay $50 per million tokens, or do I pay the cost of electricity? And for most people, based on what you just said earlier, they don't need the frontier. 90% of people are fine with today's models, especially when you have things like Kimi and GLM and DeepSeek-V3. You're already at a level that's higher than the needs of the average person.

If China hadn't forced the issue by open-sourcing, suppose there was just OpenAI, Gemini, Anthropic, and xAI—all 4 U.S. companies.

Mm-hmm.

Would you then say the same thing, that the best thing for America to do is high-quality open source?

Because China is forcing the issue.

I think that—well, first of all, we can't roll back history. It is what it is. And once it's out, especially with what you said about these models improving themselves, now that you have these models improving themselves, I think it will not be the duopoly that we have. We're going to see the Middle East, France, and Japan get into this game and say, “Hey, I can make a smaller model that's 95% as good.”

I do agree. And maybe, just to say something nice about China, since I guess I've been playing the role, to some extent, of China hawk in this conversation—

I would point out—and I'm curious, Alvin, to hear your thoughts on this—that the present situation, where even as of a few months ago, I think the West was at risk of succumbing to a regulatorily captured duopoly of Anthropic and OpenAI dominating the future light cone, was not unlike, by analogy, the Qing dynasty, where the U.S. could have sort of turned inward on itself. Chinese open-weight models, however, have basically forced open the U.S. So call it a reverse Qing, and now finally we have real competition at the AI frontier, thanks, ironically, to Chinese competition. Do you think we find ourselves now in a reverse Qing?

In some ways, yes. In fact, I think maybe the more appropriate analogy is actually... If we roll back in time, I would roll back to the Cold War, where the U.S. and the USSR were competing in an arms race. Essentially, the reason we won was not because we sent a missile and blew up Russia, or the Soviet Union, right? It's because they bankrupted themselves building military arms, and up to 15–20% of their GDP was going toward building arms that were not creating real value for their society.

In some ways, this is what China is doing to us, right? They're spending 1/10 as much on data centers and getting to 97% as good. And what we are doing right now, leveraging hundreds of billions of dollars—you had last week NVIDIA announce that they have a $500 billion deal with BlackRock and Carlyle and Blackstone and so forth to essentially securitize chips and compute—this sounds a lot like the subprime issues.

Wow. Wait, so this is the most astonishing thing, and this is also to your credit, Alvin: the first time I've heard anyone basically analogize the credit—I don't want to say bubble, but the enormous amount of private credit that the West is allocating to compute—to a reverse SDI, Star Wars moment that could... Presumably, what you're gesturing at is that could lead to the proverbial Chinese century and the collapse of Western dominance. Is that the thesis?

Well, I hope it doesn't happen, but I think we are pushing ourselves in that way. We're actually right now acting like the USSR. What perpetuated the arms race was this missile gap, right? The idea was, “Oh, they have more missiles; we have more missiles.” In both cases, they were both having the wrong numbers being provided to the leadership, which said, “We need to build more because they have 30,000. We only have 20,000.” And on they go. It became that we had 70,000 or 80,000 missiles between us that would have blown up the world hundreds of times. There was no need for any of that, right?

In some ways, we're doing the same thing right now with AI, where 45% of the U.S. stock market value is in the AI sector, right? That is a very, very fragile place for us to be.

Agreed.

At the height of the internet bubble, I think around 30% of the stock market was internet companies, right? I don't know if you guys have heard of something called the Buffett indicator, right? The Buffett indicator is something that says a market is healthy when your stock market is the same value as your GDP, okay? At the height of the internet bubble, around 120% of GDP was the stock market value. Right now, the U.S. stock market value is 240% of GDP, right?

In fact, the AI sector alone is worth more than America's GDP today, right? That, to me, is a sign that we are in a very, very fragile place, and an economic correction is due. I'm not saying it's going to happen tomorrow, but Buffett's a pretty smart guy, and he's been doing this for a while.

Well, are you saying that the U.S. is the Soviet Union, the USSR, in 1988, or are you saying that the U.S. is Japan in 1989?

Well, I think they're 2 different... I actually think that we are right now... The overinvestment that we've put into infrastructure for AI, especially when we consider what we just talked about—that it is an industry that will become commoditized. I'm not saying that AI's not amazing. AI is going to do amazing things, but the companies that are investing in it are not going to be the ones that profit from it, right?

And that is going to create a major instability in the economics of the country. If we don't manage it well, it could create a major crisis like what happened in the USSR during the late 1980s.

So you think we're overvaluing—

I—

—the frontier labs and undervaluing the sort of China AI-plus-type rest of the economy that should be the applications?

Yeah, I think that's probably a fair—

That's a good summary.

I'll take the positive side of this. The deployment velocity in China is actually quite a huge gift because it's forcing policymakers here to solve the real bottlenecks: permitting, energy, and manufacturing. So that part is at least the good part. The bad part, I think, is what you've talked about, Alvin, where we're operating like the USSR in some of this industrial policy, and it's not going to sustain.

Alvin, you're advising the U.S. Treasury and the team that's going to the Xi–Trump negotiations—or conversations—on September 24. How are you advising them?

I think the key right now is that we don't need to get to a solution on day 1, right? The success factor of this discussion is not that we come out with a massive framework that solves everything. And, by the way, this is just my personal representation, not a representation of anything that's being discussed in any of the other organizations that—

Disclaimer is noted.

Yes.

I'm sure, Alvin, you and Jacob Helberg must be besties at this point.

No comment.

Yeah, I think that's the thing: if we can come out of these discussions saying, “Okay, we're going to have a second discussion,” that's already success, right?

In the past, I think people... There was a discussion in 2024—the dialogue—and there was a lot of disappointment because China didn't come with all their technology people, and we didn't come up with a solution, and so they're not really sincere in the dialogue. I think you need to understand, just as the Chinese work on strategic plans for decades, they also take a lot longer to prepare when they're doing these kinds of diplomatic discussions, right?

For example, for the May visit, there was really no discussion on any of this until the day before between the U.S. and China. To the Chinese, that was chaotic and very unprofessional. They're like, “How can you guys be sending your president here and you haven't talked to us? What do you want to talk about?” Right?

The fact that we're now having discussions at least a month in advance, I think it's a good thing. It's a start of more proper dialogue. When Kissinger was doing a lot of these cross-border discussions, he would be there months in advance to talk to the Chinese before they had the visit with Nixon, right? And I think—

As you're talking to them, can I suggest something?

Yeah.

Because it feels to me like the U.S. is focused on having the best model, whereas the real power will come from having the best ecosystem. And I think that's what you're pushing anyway. So I'm really thrilled that you're in the middle of those discussions.

Well, I think the discussions right now are really more around safety, right? Because the ecosystem-versus-model thing is a competitiveness question: How do we become more competitive as a country or have greater influence or capability?

Really, the discussions that are happening to start are to say, “How can we keep the world safer?” Because we have a shared common interest, and the common interest is that AI is not being used by bad actors to create instability around the world; that AI itself is not potentially creating harm to the world over the longer term, right?

And, you know, there is also the underlying idea of nation-to-nation aggression, right? And I think those 3 different things are all being balanced.

I would say that the higher priority today would actually be the bad-actor, non-state-actor risk, which is what Bessent said when he was interviewed the day after that discussion, because he realizes that nation-to-nation aggression has been in balance between superpowers for 8 decades. And that doesn't change with AI. In fact, if you use AI to hack into somebody's network and then take down their power grid, that may give you a 1-, 2-, or 5-day advantage, but then there are asymmetric responses to that. People realize this.

So people in the actual national security space—we know even if we had AGI, even if we had ASI, we're not going to use it to attack. We don't want to do a first-strike attack. It doesn't make sense because that just elicits escalation, right? So non-state actors are something that everybody should be worried about because it's going to happen. It's already happening. I think ransomware and cyberattacks are up 200 or 300 percent in the last year or 2, right?

And then it's going to be even worse because of what we've seen. When you start putting 1,000 agents all trying to attack a network, at some point they're going to find the hole. So we need to recognize that shared risk requires a shared response. It requires us to share information with each other, to have that red-line hotline so that we don't have false-flag misattribution. That's actually—

Alvin, the good news could be that the fact that you have this third-party danger means that the concept of an AI national race becomes obsolete because there's a bigger problem I have to solve.

Alvin, that's exactly why I love your stag hunt analogy. It's right on. I would use massive numbers in the top-left corner there. The benefit is hugely more than 5 units. But I think the cost in the other corners is devastating. I would put some big negative numbers, but it's the right framework. I love that you're taking it into the conversations with China on the 24th.

Yeah.

I mean, it's existential on those diagonals.

Yeah.

Yeah.

And maybe—

And we placed ourselves in that diagonal. I think that's a self-imposed harm right now.

10. Taiwan Is Not About The Chips

So I think we would be delinquent in this discussion. We've talked quite a bit about the model layer. We've talked about the GPU, or chip, layer. We haven't talked about the foundry layer of all of U.S. versus China, and it would seem to me one of the cruxes at the summit and otherwise is Taiwan and TSMC.

I'd love, Alvin, your perspective: How does this end? Does this end, in your geopolitical analysis, with China attempting during this geopolitical and demographic window to invade Taiwan and seize TSMC to gain leading foundry-node capacity? Does it end with Taiwan retaining its independence and TSMC not having to blow up all of its fabs? Where does this end?

So I think there's an assumption right now that some people in D.C. are saying, "Hey, the reason that China wants to invade Taiwan is to get access to these fabs. They don't have these fabs, and so this is why they're going to attack the island." The reality is that if anybody attacks the island, there is nothing there to be had in terms of workable fabs, right? I probably shouldn't be talking about this, but I was having breakfast—

Oh, you definitely should be talking about it.

I was having breakfast with the CTO for TSMC and also a former senior official from the CIA, and the TSMC guy goes, "Hey, I heard that you guys are going to blow up our fabs if China attacks. Is that true?" And the CIA guy says, "Hey, I can neither confirm nor deny that. But what we do have is 1,000 engineers of yours that we know we will fly out before anything happens," right?

What America cares about right now is making sure that they can duplicate these capabilities to fabricate the latest chips in America if anything happens. This is part of what the CHIPS Act was. That's why I think there are some good things that came out of the CHIPS Act, that there are now hundreds of billions of dollars actually being put into domestic manufacturing for semiconductors.

I was with Intel and IBM, and we were, at the time, the global dominant player in semiconductors. But over the last 20 or 30 years, we've lost that. We've given it away, right?

Now, if China actually does attack Taiwan, they will not get these fabs, and they realize that, right? You go there, and fabs by themselves require materials from all over the world, require maintenance, require chemicals, require supplies. And if they did that, even if they don't blow up, even if the U.S. doesn't blow up the fabs or the Taiwanese don't sabotage their own systems, after a little while you'd run out of these supplies, right? They realize that.

But the reason China cares about Taiwan is not because of the fab. It is absolutely because of political history. You know this, right? In 1949, the Nationalist Party moved to Taiwan. To them, that is an uncompleted civil war.

The two countries are actually recognized right now by America and 190 other countries as being one country, right? This is kind of like Hawaii and the U.S., or maybe Puerto Rico and the U.S. They're kind of a pseudo-part of a one-national structure, right? And it is more of a political and, to them, civilizational ending to a long story. That's the main focus, and they've, multiple times, ever since essentially Deng Xiaoping to now, talked about peaceful reunification, right?

So I don't think there is an interest or a rush to do any near-term attacks to try to get Taiwan because of chips. I just don't see that.

So you don't think—

Can you talk—

There's a backroom discussion somewhere, maybe in connection with the summit: "Okay, give the U.S. maybe 2 to 3 more years to migrate leading-edge node fab capabilities to Arizona or otherwise redomesticate TSMC's capabilities, and then China, 'Okay, fine, you can retake Taiwan because we don't care anymore'?"

11. The AI Bubble Threatens Stability

I don't know if you had a chance to read my paper, The Great Reckoning and The Reconnecting. It talks about Taiwan in some aspect, saying, "Just like during the 2008 Great Financial Crisis, actually China helped out the U.S. a lot in terms of keeping financial stability." I don't know how much you guys know about the history there. But essentially, if China actually started to sell T-bills instead of buying them, it could have completely destabilized the American financial system.

Mm-hmm.

They kept buying trillions of dollars' worth, which helped keep interest rates down and so forth. We potentially might have a repeat of this situation if there is a correction in the market due to what's happening right now with the overbuild and the over-leverage of the AI sector, right?

Maybe at that point, the Americans, or maybe Trump, will give a call to Xi and say, "Hey, can you help us out again? Maybe I'll just be more hands-off, or we'll be more clear instead of the ambiguity issue." And this is me completely speculating. But I'm—

Wow. This is your war game. Just for clarity, what I hear you saying in your war game is: Sometime in the next 2 years, there's a private credit bubble that the U.S. is using to finance its data-center build-out. The bubble—assuming it exists—pops, and then the U.S. asks China to help financially in return for what, a quid pro quo regarding Taiwan?

Well, not in the sense of "Here's Taiwan," but to say, "As long as you agree to some kind of a peaceful thing and, over a mutually agreed term, we're going to stay out of it," right? Because we've been very involved in the Chinese political or the Taiwanese political sphere for a long time. We've been selling weapons to them for the last 40 or 50 years, right?

At one point, we used to have soldiers based in Taiwan. We still have military advisors based in Taiwan, right? So this is like saying if the Chinese were selling weapons to Puerto Rico and they were helping fund them, what would America do? Look at what happened in Cuba, and how we responded, right?

So I think we need to be sensitive to why this is an issue for the Chinese.

And I'm not apologizing for them, and I'm not saying they're right or wrong, but I think it's important in any negotiation or discussion to understand the other side.

This is important for everybody to understand: U.S. policy toward China and Taiwan is a one-China policy.

Mm-hmm.

Explicitly stated.

Yes.

They leave the intentions deliberately ambiguous in terms of how that happens. It's called strategic ambiguity, I think.

Yep.

Alvin, let's wrap up with one last commentary from you. How do you think the next 3 or 4 years go? What are the 2 big paths that you think we have to pick one or the other? How do you see the next few years playing out?

Are you talking about the AI space in general?

AI and the global transformation.

So that's actually the whole narrative in the Great AI Reckoning paper: how the next few years play out. What I foresee is that we will soon find that these AI companies—once they go public, or if they go public—their financials will become much clearer. People will start to realize that the value of AI innovation does not necessarily accrue to them. It may in the near term, right? They have been one of the biggest beneficiaries.

But that accrual also came from a period when you didn't really have the open-source capabilities that we have today. In fact, if you look at the recent disclosures in terms of where Anthropic's revenues are actually starting to flatten out a little bit, it's not—earlier this year, they were growing like 10× over just a few months, right? Now they've essentially flattened out at an ARR in the $70 billion range. Although their ARR is at $7 billion, their first 2 quarters were, I think, less than $20 billion in total revenue.

They're committed to hundreds of billions in CapEx and debt. There is right now $1.6 trillion of off the book debt, and $1.7 trillion of off-the-books debt among the major hyperscalers. During Enron days, there was $200 million of off the book debt, just to give some context for the scale of the problems that we're looking at.

If that happens, I think people will actually slow down construction, because construction right now is all based on the idea that these companies will continue to make money and continue to be able to fund and service their debt. If you look at Amazon and OpenAI revenues, when they talk about AI, most of that—probably more than half of it—comes from 2 companies. That is not a very diversified revenue base.

As more and more of the capabilities move to open source and edge computing, the dependency on cloud-based premium services will continue to erode.

We'll shift the bottleneck down the stack to compute, electricity, power, et cetera.

Yeah.

Although China—China is also—I mean, maybe to present the other side of this, China is notoriously dependent on real estate and property development in order to drive provincial revenues, because the provinces are really selling off real estate, or had been selling off real estate, to generate their own local revenue. I don't want to overanalogize, but isn't there a sort of striking parallel between Alvin, you're pointing to the West maybe overleveraging compute and data-center infrastructure development, and China perhaps overleveraging or over-indexing on real estate development for humans?

Yeah. So you make a really good point. Over the last 3 years, there's been about a 30% deflation in total real estate value in China.

Right.

And they managed it in a way that it was not a crisis, right? We need to do a soft landing for these things so that it does not create a crisis. So I think that—

We're building the houses for the AIs, and China was building the houses for ghosts.

Yeah. Well, no. I think that the ghost-town thing—there may be a few, but the reality is that home ownership right now is something like 70% in China, and it's probably less than 50% in America, right?

Yeah, it is.

So I don't think we want to overparallel these 2 things. But I think what we can learn is that when the crisis happens, you need to be willing to take some near-term pain, and they did. They took major hits in their GDP slowdown. They were growing at 8%, 9%, 10%, and now they're growing at 4% or 5% GDP per year, mostly because the real estate sector stopped growing. In fact, it started to decline, and they had to make up for it with other types of industries.

Global policy folks are talking about this China-managed crisis as a hallmark case study on how to do it in the future.

Yeah, let me ask you about the managed crisis, actually, because China—

It's amazing to me—

Because China—

Guys, we're going to take one last comment. Dave, last question, then we've got to wrap it up.

All right.

So Alvin, we've got to have you back. We have 100 more questions, but we'll do that some other time. Dave, over to you, and then we'll have a response.

So China is clearly a country coming into a crisis because of the birth rate. The 1-child-per-family policy is catching up in a huge way. The population is aging like crazy. It's a crisis, and that's why the country is so focused on robotics, because they're going to need it more than anyone.

When I was at MIT, there was a class called “Just Wars, Total Wars, Nuclear Wars,” and I thought, “I have to take that class and see what it's all about.” Essentially, what they taught us in the last third of the class is: Look, this is at the height of the Cold War. The U.S. is over here, the Soviet Union's over there, and it's a prisoner's dilemma.

As nuclear weapons get more and more efficient, inevitably the prisoner's dilemma gets more acute. Sooner or later, one country or the other is going to have the ability to destroy the other country with no retribution whatsoever. This is going to destroy the world.

In reality, this is where I lost faith in political science classes. It didn't play out that way at all. Like you said earlier in the podcast, the Soviet Union bankrupted itself with way too much weapons investment. But it then became obvious that the Soviet Union wasn't really a tight-knit country in any way, shape, or form.

And now we have Ukraine and Russia—both part of the Soviet Union—in a 5-year-long catastrophic, devastating war. The other satellite entities don't even speak Russian.

So I don't have any idea. What is China like? Is it truly unified like the United States? Is it fragmented? What is it?

Yeah. I mean, I think this is one thing that China has that's very different from a lot of the Western world. It's very homogeneous, right? Probably 95% of the population is Han Chinese, right?

Everybody speaks Mandarin. They may speak other local dialects, because there are hundreds of local dialects, but they all speak Mandarin. The written script is the same across all the different provinces, right?

I don't think we're going to see the type of issues that you saw with the USSR, even if there was a major economic crisis. I think they've managed it very well. In fact, they've learned a lot of lessons from the disintegration of the Soviet Union to say, “We cannot let that happen,” because that would mean hundreds of millions of people would suffer or die, right?

Their biggest priority is social stability, political stability, and economic stability.

I should just add a fine point: the Uyghurs—ethnic Muslims, ethnic Turks—may differ with that assessment regarding ethnic homogeneity and the unity of approach, sort of an “everyone's Han” characterization of China, for the record.

No.

Mm.

That's 95%, right? So I think there are definitely a few percent, but it is a relatively small minority, right? Even the Uyghurs or whoever, they all speak Chinese. They all read Chinese, so there is a common language.

I think that the desire to secede is not as prevalent as we tend to portray it in the U.S. press.

For next time, I guess.

Yes, let's save that for next time. I'll say one thing. I spent a few months traveling around China, and my conclusion was that the native entrepreneurship in the Chinese people is higher than in any other country I've ever seen—just latent. Therefore, if we believe entrepreneurship is a major driver for future success in the world, that's an amazing thing, and I think we're seeing that come out of it.

Yeah.

I think we're seeing that come out of it.

I do want to end with one thing. For America to actually be successful in this industry and take advantage of what we've created, we need to think about creating something akin to an AI Marshall Plan.

Yeah.

The Marshall Plan after World War II: we spent around, I think, $15 to $18 billion rebuilding most or much of Europe and some parts of Asia.

It also created markets for our goods.

It created giant markets for our goods, because at that time we had 50% in manufacturing. And it created loyalty and allies for eight decades, right? We need to be thinking more about that type of thing today, but with AI data centers and AI technology. It's the same thing that China is actually doing right now. Their AI coordination organization, essentially, is their AI Marshall Plan. We should either be doing something like that, or we should be working with them to do it together, right?

Yeah, that's great.

But I think doing that—

That's great advice.

Yeah.

Yeah.

If we did that, we would have a big market to sell our chips. When we stop building data centers here, which we probably will at some point when people stop being able to finance them, we're going to need to sell those NVIDIA chips in better places, right?

Yeah.

So we're going to need to sell services. There are a lot of good things that can be had. And we also hopefully will then have a place for the open-source models that we create. We can start to create frontier open-source models—safe ones that both countries agree on, that both countries start testing, and that both countries have standards for. Then this technology can be diffused to the world without creating a crisis, without creating additional competition and conflict.

Well, I think what you said back-to-back there is that 95% of China is ethnic Han, and that the U.S. needs a Marshall Plan. But the U.S. has this incredible advantage in that there's no single ethnicity of America. It's a complete grab bag of the entire world.

And so the Marshall Plan, executed well from the United States—we just keep shooting ourselves in the foot, but if we stop doing that—we're a much better long-term ally for all these countries in the world that have experienced either ethnic genocide, ethnic racism, ethnic slavery—

Cleansing.

Or ethnic cleansing. They'd much rather work with the United States if we just give them a chance.

Yeah. And that's what we're—

We're telling the world the opposite story right now, right?

Yeah. Yeah. That's the advice we'll end on: stop shooting ourselves in the foot. All right. Alvin, it's been awesome to have you on. I think I speak for all of us on the pod and the viewers when I say it's really great that you're in the middle of these discussions. So push your ideas as hard as you can. We'll do the same on your behalf. We'd definitely love to have you back again sometime. And on that note, thank you for being with us, and we'll wrap it up for today. All right?

Yeah. Thank you all. I really appreciate the conversation.

And thank you for the questions, Alex.

You know what? Someone has to ask them. I like to say my job here is to call the balls and strikes, including regarding China. But thank you for being a good-humored recipient of the balls-and-strikes calls.

No, all good. Thank you.

All right. Don't go away. Don't go away.

All right. We're going to wrap it up.