As the U.S.-China summit gets under way, the world's two biggest economies are locked in an artificial-intelligence arms race. Leadership on both sides believe future economic and military supremacy rides on mastery of the technology. On the Chinese side, the government is willing to subsidize the AI infrastructure to advance its agenda.

The first area of competition is in AI models. The best models are still made by two U.S. companies, OpenAI and Anthropic. Their current top offerings represent a big step up in capabilities from previous generations and can work on problems for a long time, persistently plugging away until they get the solution. They are particularly good at advanced math, coding, plus both hacking and cyber defense.

But the costs add up quickly when using these top models, and not every task requires them. OpenAI and Anthropic have less expensive choices, and there are other capable options from Meta Platforms, SpaceX, and Alphabet.

But the biggest threat to OpenAI and Anthropic is coming from open-weight models, many of which come from subsidized Chinese companies. This week saw a new Chinese champion take the spotlight: Xiaomi's MiMo 2.6 Pro. According to Artificial Analysis' testing, it delivers about the same intelligence as SpaceX Grok 4.7 at a fraction of the cost when used on Xiaomi's cloud.

Open weight means anyone can install MiMo anywhere under a very permissive license. It's becoming an increasingly popular path for U.S. companies, reducing costs and giving the customer control over its model infrastructure. Meta and Nvidia have also committed to producing advanced open-weight models.

The U.S. sector leaders are starting to feel the pricing pressure. On Tuesday, Anthropic released its latest Opus-class model-which looks like another big leap in capabilities-along with a price cut. The same day OpenAI released the latest version of its Sol-class models, also now cheaper, on top of an August price reduction.

Anthropic and others claim that these Chinese labs can keep up only by training their models on the outputs of Anthropic's Claude, a process called "distillation." In an exhaustive September security report, Anthropic charged that Xiaomi and other major Chinese labs were using thousands of fake personal accounts to relay customer prompts to Claude models, then using the responses as training data for future models.

If distillation remains the primary means of Chinese model improvement, they will remain several months behind the U.S. frontier offerings, without ever passing them.

Because of export controls, China will have a tougher time competing in AI chips. Nvidia is the industry leader, and China's share of Nvidia's data center sales is gradually dwindling to little more than a rounding error. Wall Street analysts expect Nvidia's data center sales to hit $381 billion this year, nearly doubling last year's tally. Some Chinese companies have been able to evade the controls by using Nvidia chips hosted by cloud providers outside China.

There are other AI chips now, including some from China like Huawei's Ascend and Alibaba's Zhenwu. The companies are opaque about their specs, so it's hard to compare them to Nvidia's or others.

But the more important export controls involve chip manufacturing technology. The Chinese homegrown chips are likely made at Semiconductor Manufacturing International Corporation, or SMIC, a partially state-owned chip manufacturer with the best Chinese manufacturing process, one roughly equivalent to what Taiwan Semiconductor Manufacturing first introduced in 2018. Taiwan Semi won't make high end chips for Chinese customers, and neither will the other leading-edge manufacturers, Samsung Electronics and Intel.

SMIC is also hampered by the lack of access to the best lithography machines from ASML, called extreme ultraviolet. Without a Chinese substitute-and there is none on the horizon-SMIC will only be able to make incremental improvements to its chip manufacturing. More recent export controls on the next tier of ASML machines, called immersion deep ultraviolet, mean that SMIC will have difficultly scaling up its manufacturing to match homegrown demand. SMIC has a soft ceiling on both technology and size.

Meanwhile, the non-Chinese manufacturers will keep pushing ahead, making the chips they produce more powerful and energy efficient, and in the coming years the chip gap may grow.