When Chinese President Xi Jinping sits down with U.S. President Donald Trump in Washington, D.C., he will do so with China in a markedly different position from when the two men last met face-to-face in the United States nearly a decade ago. Nowhere is that shift more apparent than in AI. China has emerged as an AI heavyweight, with homegrown companies pushing Chinese models toward the global frontier. 

Alex Lamb knows both sides of that race. A distinguished AI researcher who trained under Turing Award (a top annual computer science award) winner Yoshua Bengio and went on to work at Microsoft, Amazon and Google’s DeepMind, Lamb joined Beijing-based Tsinghua University’s newly established College of AI in 2025.

In a recent interview with Beijing Review reporter Peng Jiawei, Lamb discussed the shifting balance between the two countries’ AI sectors, the different ecosystems behind them, and why an intensifying tech race need not rule out possibilities for cooperation. Edited excerpts of their conversation follow:

Beijing Review: Having worked in AI in both China and the U.S., how have you seen the AI race between the two countries evolve?

Alex Lamb: The timing was actually quite interesting. I accepted an offer to come to Tsinghua University just around the time the fighter jet Chengdu J-36 was unveiled and AI assistance platform DeepSeek came out.

It was purely a coincidence, but there was suddenly this cluster of major technological breakthroughs coming out of China, and I think that caught a lot of people by surprise. It surprised me, too. You could see the reaction in the stock market.

Before that, GPT-4 was still widely regarded as the leading model, the frontier of AI felt much more closed, and the computational advantage of the leading U.S. labs seemed substantial. So when DeepSeek-V3 came out, I think everyone was shocked.

Since then, I feel like there has been a bit of a holding pattern, where the best Western AI models would be leading, and at times it can look as though they are pulling further ahead. Then a Chinese lab comes out with a major release that closes the gap again.

More recently, for example, you had GLM-5.3 [Chinese Z.ai’s flagship 2026 coding and cybersecurity-focused AI model—Ed.], one of the most capable open-weight models for coding. It was the first Chinese model to demonstrate really strong agentic coding capabilities--an area where Western AI labs have been particularly strong.

One area where Chinese AI labs have not yet shown as much capability is frontier-level mathematics and physics research. So I think people are watching closely to see whether the next generation of Chinese models can close that gap, and how the race goes from there.

What are the biggest differences between the AI ecosystems in China and the U.S.?

There are indeed some major differences. For whatever reason, the U.S. approach is much more concentrated. The leading labs tend to be quite secretive and closed off. It can be very difficult for outsiders to get any access to companies such as OpenAI or Anthropic.

Chinese labs feel more open in some respects. Leading AI companies often have close relationships with universities and academic researchers. One of the co-founders of Z.ai is Tang Jie, who is also a professor at Tsinghua. People still see him riding a rental bike around campus, which is a little funny given that he is now a billionaire.

There is also a tendency in China to fund a larger number of companies. Individual companies may be somewhat smaller, and their strategies may sometimes be more conservative, but there is a broader ecosystem of experimentation. In the U.S., by contrast, there is more of a tendency to concentrate resources in a few leading organizations.

Also, the picture is quite different when it comes to AI software and hardware.

On the software side, China has an enormous pool of high-quality engineers and researchers. Most of the world's good AI engineers are in China.

The bigger constraint has been hardware, because advanced AI chips depend on an extremely deep stack. [An AI stack is a collection of technologies, frameworks and infrastructure components that facilitate using artificial intelligence (AI) systems—Ed.]

The U.S. has been able to work with other countries to restrict the export of some of the most advanced semiconductor technologies to China. One of the fundamental technologies involved is extreme ultraviolet (EUV) lithography, a technology used for manufacturing integrated circuits. China is investing heavily in developing its own capabilities in this area.

There is another piece of technology introduced by Huawei called logic folding, or tau scaling. [Logic folding is a 3D chip-stacking architecture that vertically folds circuits to shorten signal distances and boost transistor density without relying on advanced EUV lithography—Ed.] This is one of the ways developers are trying to reproduce the fundamentals of semiconductor manufacturing.

We often talk about the AI race between China and the U.S., but is there still room for cooperation?

There are many reasons why AI development in the U.S. and China did not necessarily have to become this competitive. The two countries could have been quite complementary, because some of China’s advantages are very different from those of the U.S.

The U.S. has two major structural advantages. One is language. Globally, many people working in technology prefer to work primarily in English, and English is the official language of the U.S. That helps attract engineers and scientists from around the world. The second is the sheer scale of its technology companies, which means they would naturally play a very large role in the AI ecosystem.

China’s advantages are different. It is very strong in manufacturing and infrastructure, and it also has a large pool of talented engineers.

I could imagine, hypothetically, a relationship in which AI models were developed at companies such as OpenAI or Anthropic, while some of the inference data centers were built in China. I think both sides could end up wealthier if that kind of arrangement were developed.

In an ideal world, I think it would be very good not to have boundaries, and instead to work together toward some form of global governance.

From an AI-safety perspective, if AI generally becomes more intelligent than humans--which seems fairly plausible within the next couple of years—then that raises important questions about how its development should be jointly managed.

There is a quote often attributed to John von Neumann: The trouble with saying that a machine cannot do something is that, once you can define exactly what it cannot do, someone can try to build a machine that does precisely that. [Von Neumann is a Hungarian-born American mathematician who pioneered game theory and, along with English mathematician Alan Turing and Father of Information Theory Claude Shannon, was one of the conceptual inventors of the stored-program digital computer—Ed.]

That is why I think it is important to establish some guardrails around AI development, ideally through cooperation rather than an increasingly competitive race between countries.

Copyedited by Elsbeth van Paridon

Comments to pengjiawei@cicgamericas.com