Is China Closing In On the AI Race?

The question of whether China has caught up with the United States in artificial intelligence changed character in January 2025, when a Hangzhou company few Western readers had heard of released a reasoning model that performed close to the best American systems but cost a fraction of what they cost to build. DeepSeek did not win the race, and the answer to whether China is overtaking the United States is still no on the measures that matter most, including the frontier of raw capability and the depth of the compute available to train the next generation of models. However, the last twelve months show that the distance between the two countries has narrowed faster than most American forecasters expected, and that the narrowing is driven by a Chinese strategy that treats efficiency and open distribution as substitutes for the hardware China cannot legally buy.

When the company released DeepSeek-R1 in January 2025, its reported training cost of a few million dollars sat against the hundreds of millions that American labs were understood to be spending, and the model was released under an open licence that let anyone download, inspect, and run it. The immediate market reaction was a fall in Nvidia's share price that erased close to 600 billion dollars in a single session, the largest one-day loss for any company on record at that point. The number itself matters less than what unsettled investors: a working demonstration that the relationship between spending and capability, which had underwritten the valuations of the entire American AI supply chain, was weaker than assumed. R1 showed that frontier AI might be achievable with fewer resources than the industry had come to believe.

The picture changes once export controls enter the scene. Since October 2022, the United States has progressively restricted exports of Nvidia's most advanced AI chips to China. That forced Chinese companies to rely on lower-performance chips, including Nvidia's H20 and Huawei's Ascend processors. At first glance, that looks like a straightforward constraint. But researchers at Berlin's Mercator Institute for China Studies argue it also had another effect. By limiting access to American chips, the restrictions gave Chinese manufacturers such as Huawei both a guaranteed domestic market and a stronger incentive to improve their own designs. The controls have made training frontier models slower and more expensive for Chinese labs. But they did not stopped Chinese AI from advancing. If anything, they have accelerated China's effort to build its own AI hardware industry.

Chinese companies also took a different approach. Instead of competing only on scale, companies such as Alibaba and DeepSeek released capable models that developers could freely download and adapt. Those models have since been used far beyond China, including in Europe and the United States. Which shows how influence does not come only from building the biggest model. It also comes from building models people choose to build on.

None of this means the balance has shifted completely. The United States still leads in the infrastructure that supports frontier AI: advanced chips, large-scale compute and private investment. found that U.S. private AI investment reached almost $286 billion in 2025, more than 23 times China's private investment, while American companies continued to produce more top-tier frontier models. DeepSeek showed that highly capable models can be built more efficiently. Sustaining that pace over successive generations, while constrained by access to the most advanced chips, remains a different challenge.