Previewing U.S.-China (Trump/Xi) Summit for AI

The AI Power Podcast

The episode analyzes the geopolitical landscape ahead of a summit between President Trump and Chairman Xi Jinping, focusing specifically on the

Key takeaways

• China utilizes subsidiaries like Averis to bypass NVIDIA Blackwell restrictions by routing technology through Southeast Asia.
• Chinese hyperscalers including ByteDance, Alibaba, and Tencent are renting advanced compute capacity from cloud providers in Malaysia, Thailand, and Japan.

Main topics

• U.S.-China AI Summit
• Semiconductor Export Controls

Notable quotes

China is effectively copying off of America's homework, and you can't get a higher test score than the person you're cheating off of.

Conclusion

The summit represents a critical moment where both nations must naviga

Transcript preview

Speaker 2 (0:10) Welcome back to the AI Power Podcast. I'm Jack Conlon, filling in for Adam Goodwin. In today's episode, we'll focus on the upcoming summit between President Trump and Chairman Xi Jinping and the role that artificial intelligence will play. We'll first give an overview of how each side is doing on the AI front, what their chief concerns are, and how they're feeling heading into the meeting. Then we'll turn to the summit itself, what each side wants, where they're stuck, and what could realistically come out of the summit. And here, as always, to help us make sense of these developments is Greg Allen. Greg, how are you doing? Speaker 1 (0:41) I'm doing great, Jack. It's going to be a heck of a week. There's a ton of stuff to unpack here. So hopefully everybody out here is looking for a China mega episode because that's what you're about to get. Speaker 2 (0:52) And let's get right into it. Before we get to the summit itself, though, I want to understand where each side stands going in. Let's start on the Chinese side with Beijing. Over the last few years, the U.S. has tried to slow China's AI progress, primarily with export controls. And yet Chinese AI keeps advancing. How do you think Xi is feeling heading into these talks? Speaker 1 (1:13) So I think Xi Jinping is probably feeling pretty good heading into these talks. And the basic reason is that he has things that he wants from President Trump, but he doesn't have things that he desperately needs from President Trump. And what I mean by that is, if you think about all the pressure facing the Chinese system, very high youth unemployment, slowed economic growth. Those are not the metrics that Xi Jinping is grading himself most critically on. He's really focused on technological competitiveness and whether or not he's making progress in the key industries. and advancing towards self-sufficiency in those industries. That's the same basket of priorities he's had since he came to power in 2012. And I think everything that he's seen happen in the global economic system and into global geopolitics since then has convinced him that he had it right the first time. And they're actually making important progress on those metrics. So let's think specifically on AI. I think there's basically four ways whereby China has improved its AI posture and its competitiveness vis-a-vis the United States that overall make Xi Jinping feel like he has a pretty strong hand going into these negotiations. So the first is the biggest threat that China faces in AI is those export controls, cutting off access to compute and chip making equipment. But what... they have demonstrated is a strong ability to take advantage of both legal and illegal measures in order to continue gaining access to U.S. compute. That's the first thing. The second thing they're doing is, even though they're still significantly behind, they're also making noteworthy progress on building their own chips and chip making tools. The third thing is that Chinese AI models are still really appealing to not just folks in China, but folks around the world and other countries that matter to U.S. geopolitical calculations. Fourth and finally is just that distillation is working as a rapid catch-up strategy, and the United States does not have a great rebuttal to it for all of its bluster so far. So I think those are the four things that are in the back of Xi Jinping's mind as he walks into the negotiations with President Trump. Speaker 2 (3:38) Let's take those one at a time, starting with the first one. You mentioned illegal and legal access to American chips that continues to fuel Chinese AI push. How is China still getting access to these chips? Speaker 1 (3:51) Well, I think we saw a headline just a couple of weeks ago on The New York Times came out on September 6th that reported that InSpur. a major Chinese data center company, one of the largest makers in the world of computer servers, and therefore a key vector of accessing NVIDIA compute from an integration perspective. They are on the entity list, but they are still able to buy massive amounts of American compute because of a ridiculous restriction. that allowed them to continue buying even when they should not have been able to, at least from a strategic consideration. So the key point here is that Inspur has a subsidiary called Averis that got... $5.6 billion worth of advanced technology as a U.S. subsidiary, where it was able to get around its entity listings and then export that to Southeast Asia. And among all of those technology imports included billions, $3 billion worth of NVIDIA Blackwell chips. So remember... The Trump administration said that their policy was we're going to allow H-200 sales. We're going to ban Blackwell sales. And in the future, we're going to ban Vera Rubin sales. That was the Trump administration theory of the case. But you have to ask yourself, what is the gap here between the policy and the process? What do I mean by that? Let's just say, hypothetically, that we had a law. that said that every American is entitled to the right to vote. Okay? And that is the policy. Every American has the right to vote. But then we have a process. And the process is there's only one place to register to vote in all of the United States. And it's a single DMV in rural Alaska that is only accessible via private jet. Well, then we have a policy that says everybody's allowed to vote, but we have a process that says only people who have easy access to private jets are in practice going to be able to vote. And so now take that analogy and apply it to export controls. We have a policy that Chinese companies cannot get Blackwell chips, but the process that we have for enforcing that has been so. pathetic in so many different ways, you have to ask yourself, is letting folks get away with smuggling, is that actually the secret policy? I mean, that's how bad it is here. Which again, is not to say that the export controls have not had a strategic impact. I think they have, but it goes a long way towards explaining... how we got here where China is still able to do it. Now, keep in mind that Averis example that we were just talking about, the Averis Inspire example, them buying those chips from NVIDIA, that was all legal. It only became... illegal if it made its way into China, which is my assumption that it in fact did as part of the smuggling ring reported by the New York Times. So that's one legal tool that China's been able to exploit very successfully. The second one is renting what it cannot buy, which is advanced computing capacity from cloud computing providers across Southeast Asia. And, you know, like, who is that actually helping? Who is this remote access to NVIDIA chips via cloud computing helping? Well, according to reporting by CNBC, quote, Chinese hyperscalers, including ByteDance, Alibaba, and Tencent, have reportedly accessed compute power from NVIDIA chips remotely via other Asian nations, including Thailand, Malaysia, and Japan. ByteDance was working with Singapore-headquartered Aolani, a cloud provider with NVIDIA chips to access compute in Malaysia, according to a source familiar with the matter. Aolani told CNBC it worked, quote, with a global and diversified customer base spanning customers from North America to Asia. Then it goes on to say, the companies we service do not have ownership. Potential future claim. or physical access to the chips that power our solutions. Any permitted access to our services, infrastructure, or technology is fully compliant with all applicable regulations, the spokesman added. So that's a person basically saying, they didn't say China, but I think between the lines, you can assume they mean China. And they're basically saying like, look, it's legal. If you don't want us to do it, make it illegal. And that's unfortunately where we're at. It's important to note that These Speaker 2 (8:27) companies are not small fish. You have Inspur, which is listed as an elite NVIDIA partner. You have other firms who have been involved in this third-party access like Megaspeed, which was also investigated for chip diversion, but they still provide this entirely legal third-party access through countries like Malaysia and Indonesia that are building up massive compute capacities that China is renting out, again, in a completely legal sense. Speaker 1 (8:52) Yeah, I mean, you put it well, right? Maybe folks who listen to this podcast have never heard of Inspur, but anybody in the data center industry around the world, I guarantee you, has heard of Inspur. And there's a sense in which a company like Inspur is almost too big to fail because they make up such a significant share of global rack integration. They can effectively say like, hey, if you want to entity list us, if you want to try and destroy us, like that's actually going to have consequences for downstream compute availability and demand, not just in China, but around the world. So if you do that. You're not just picking a fight with China, you're picking a fight with all of China's customers. This is the exact same problem that the United States faced when they went after Huawei. These companies are so big that it's just not trivial to punch at them, even if you have a good legal justification for doing so. So one final thing is just to think about where is China domestically in computing capacity? And in that regard, we actually got some interesting data. from China's Ministry of Industry and Information Technology, MIIT. So the 15th five-year plan for information and communication sector was just released on September 7th, and it set a target. for 9,800 exaflops of intelligent computing capacity by 2030 and called for 3.8 trillion yuan, which is the equivalent of about $530 U.S. billion in cumulative information infrastructure investment over the 2026 to 2030 period. Now, where is China now and where does that compare? Because 9,800 exaflops. Maybe that just sounds like Chinese to all of you out there listening. But it is like the measure of computational power aggregated. And in July, that same ministry in China reported that China's intelligent compute, so computing capacity that's relevant to AI applications, had reached an aggregate of 2,185 exaflops at the end of June. That's up 177%. over the year before. So it's growing quite rapidly. By contrast, America right now has 40,000 exaflops. So America has more than four times the amount of compute that China expects to have by the year 2030, and more than 20 times the amount of compute that China has today. And, you know, how much is America going to have in 2030? Well, You know, there's various scenarios, but China is more than doubling every year. And so far, America has been doubling or close to doubling the amount of computing infrastructure for AI that's going in the ground thus far. And it's plausible. There are various scenarios put up by Epic AI and others that foresee America having 250,000 exaflops by 2030 or. 500,000 exaflops by 2030. So it's not like the export controls have accomplished nothing. There is this massive compute advantage that I don't think would be there in an alternate universe where these export controls had not happened. At the same time, there really is this gap between how much compute do you need to... Run all the experiments that give you the recipe for creating a frontier level model versus what is the actual amount of compute you need to train such a model once you have that recipe. And the combination of industrial espionage to understand what are all the tricks that are being used by the AI companies, what are the results of their experiments, along with distillation, which is to extract the results of that recipe in the form of useful training data. That means that what took the American companies gigawatts of compute and months of time might actually only take the Chinese companies, which have far less compute. hundreds of megawatts and months of time. Because like the actual training run, and this was disclosed by Dylan Patel of Semi Analysis, but like the actual training run of training in Mythos might be like 100 megawatts or 200 megawatt data center running continuously for a month. And there are Chinese AI labs that have that much compute. What they don't have is the other gigawatts of compute required to run all the experiments. that let them know that that training run is going to work or the gigawatts that serve the hundreds of millions of customers on the inference side. So distillation is really critical in making up that gap, as is industrial espionage. So that to me explains why Xi Jinping still doesn't like these export controls, why he would still really want them gone because of that 500,000 exaflop. versus 9,800 exaflop that might be the gap by 2030. But it also explains why he's not desperate, why he's not coming groveling, begging for compute. He feels like he has a decently strong hand. Speaker 2 (14:17) The second way you mentioned China improving its AI posture is by building up its own domestic chip making and equipment making capacity. Compute self-sufficiency is something Beijing has promoted since at least Made in China 2025, which was released in 2015. And national champions like Huawei have been the standard bearers for advancing this kind of self-sufficiency push over the last decade or so. How is that effort coming along heading into these talks? Speaker 1 (14:41) Well. I think it still remains a priority and something that they're working incredibly hard on. I think one thing that struck me as quite interesting is that we've been talking about these five-year plans that are coming out that are sector by sector. So another five-year plan, the Electronic Information Manufacturing five-year plan, which was issued jointly by MIIT, that ministry we're just talking about, as well as the NDRC. came out on September 15th, and it lists full chain breakthrough in integrated circuits as its number one priority. So what they're saying is of all the things that we have to do in the electronics industry, our top goal. is full chain, meaning full supply chain, breakthrough in integrated circuits. I mean, that's listed above building AI data centers. That's listed above leapfrogging the United States in AI leadership. I mean, what they really care about is, are they self-sufficient in semiconductor technology? And it makes sense. I mean, if you're China, you're importing more by dollar value of semiconductors every year than you are oil. Even with all the export controls, that's still true. And that is a terrifying thing to potentially lose access to and explains why they're so focused as something that they need to make progress on. And here you're seeing evidence that even though they're still behind, even though they're still not where they want to be, they are still making progress. So one piece of recent news on this topic that struck my eye was the September 4th piece by Mackenzie Hawkins in Bloomberg, where she reported that DeepSeek, the leading Chinese AI model developer on many metrics at least, is planning to install a new data center full of 160,000 Huawei chips and one gigawatt of energy capacity going to be built out in Inner Mongolia. Now, according to the report, these are going to be Ascend 950DT chips, which is the latest generation of Huawei's AI chips. It is generally assumed to be roughly comparable to an NVIDIA Hopper generation of chips. So for NVIDIA, those were the chips that they announced in early 2023. Huawei announced the 950DT in 2025 and appears to actually be making some of them in 2026. When are they going to have 160,000 of them? Not at all clear. The article says that it's going to take more than a year to fulfill this order. That makes sense based on some of the other details it includes, such as Huawei is planning on producing 1.6 million Ascend dies. across its various production network, which mostly includes SMIC's advanced node manufacturing capacity, but also perhaps some of the capacity that Huawei itself is bringing online internally. And 1.6 million Ascend dies is enough for about 600,000 910C chips, which uses two of these dies, and then you have to factor in yield issues. Not a ton of 950DT production expected to take place this year, which means they might not be fulfilling this order for a very long period of time. In fact, just six months ago, it was a breakthrough when one company brought online a compute cluster that used 10,000 Huawei chips. And then in June, somebody was reporting the first ever model trained on 50,000. quote, domestic Chinese chips, although we don't really know if those were even made in China or merely just designed in China. But all this is to say that even though China is way behind in the compute capacity story, they're still making progress. They're still better this year than they were the year before, and they're still better than many, many countries around the world. I think the last point of comparison would just be SpaceX AI. When they built the first iteration of the Colossus supercluster in the second half of 2024, that was 100,000 NVIDIA Hopper chips. So that means sometime in the 27, 2028 timeframe, DeepSeek might have a data center equivalent on domestic chips to what SpaceX had in 2024. Still a pretty significant gap, especially when you take into account. That by that time, companies like SpaceX AI are going to be deploying Vera Rubin chips, or maybe even what's the next generation of NVIDIA chips, which are just way better on a per chip basis than what Huawei is coming out with. And Speaker 2 (19:27) Greg, in the AI value chain, chips are just part of the picture. You wrote a sub stack in August about how China is making a significant push to produce semiconductor manufacturing equipment domestically in order to reduce their reliance on companies that are more aligned with the U.S. in both East Asia and Europe. How is China doing on that front? Speaker 1 (19:46) Well, so that Substack piece was really about lithography specifically, which is arguably the most important piece of equipment in a chip manufacturing operation. Although, you know, which is more important, your brain or your heart? You need both of them to live. But Chinese suppliers have been pushing to domesticate semiconductor manufacturing equipment for a long time. They had explicit metric targets in Made in China 2025, that policy we talked about from 2015. In general, there is a group of companies that are making progress on this, denoted as just advancing their own technological sophistication and capacity of production. So in July, when we were writing that Substack piece. Shanghai Ashenega, which I'm sure I'm pronouncing that wrong, electronic technology group reportedly began volume production, immersion DUV machines. And we're talking like single digit numbers this year, double digit numbers next year. In some sense, you could say that puts them two generations behind ASML. So ASML right now is making what's called high NA EUV machines, like extreme ultraviolet lithography. And then since 2010. ASML has been making extreme ultraviolet lithography. The most advanced type of equipment before that was this immersion deep ultraviolet lithography technology, which was already mature in ASML in the early 2000s. But Huawei... Even though it is best known here in the United States as a telecom company, even though they're also pretty well known