Slowing China’s progress in artificial intelligence (AI) has been a top priority for Washington for the last three years. To achieve that goal, the Biden administration has escalated controls on the sale of advanced chips and chipmaking equipment to China, including a fresh salvo of restrictions earlier this week. 

Policymakers may be flummoxed to learn, then, that Chinese companies aren’t just keeping up in the AI race: some believe they could overtake American industry leaders as soon as next year.

The latest breakthroughs came late last month, when two Chinese AI companies released new models that perform as well, if not better than their American peers. Developed by tech giant Alibaba and High Flyer Capital Management, a Chinese quantitative hedge fund, the technologies compete directly with OpenAI’s latest o1 model, which can “reason” through problems — a process some researchers have described as a new paradigm. 

These achievements by Chinese firms underscore how formidable a competitor the country remains in the global AI race. Buoyed by a wealth of engineering talent and intense domestic competition — plus ample chip supply for now — Chinese AI firms are unlikely to fall back in that race as easily as some in Washington may hope. 

Influential figures in the AI community are taking note. On Monday, Clement Delangue, chief executive of HuggingFace, a popular platform that offers tools and data to AI developers, predicted on LinkedIn that China would “start to lead the AI race in 2025.”

Several AI model benchmarks suggest that, on performance, Chinese models are already in an effective dead heat with offerings from the U.S. and Europe. High Flyer says that its latest model, called DeepSeek-R1, performs on par with OpenAI’s o1 on two widely used performance benchmarks. Alibaba claims its latest model outperforms o1 on those tests. 

To be sure, these tests are imperfect. “We don’t want to read too much into a difference of a few points on these different benchmark comparisons,” says Jeffrey Ding, an assistant professor of political science at George Washington University and author of the ChinAI newsletter, which follows Chinese AI. “But the bottom line is, there’s a wave of Chinese open-source AI models that are as competitive as their Western counterparts,” he says. 

Besides its vast pool of engineering talent, China’s lower energy costs, particularly in far flung regions into which Chinese cloud infrastructure providers have poured investment, have helped it close the AI gap in the last year. Training AI models is highly energy intensive as it requires developers to run up to hundreds of thousands of chips over long periods.

Technological innovation has also helped to drive down the cost of training and using AI models — the process known as inference. Earlier this year, DeepSeek’s developers worked out a way to greatly improve inference efficiency, allowing the firm to offer its model at a much lower price than its competitors while still remaining profitable, its founder has claimed. That kicked off a brutal price war among Chinese AI providers that has brought the cost of using Chinese AI models down to as little as one-hundredth of OpenAI’s price. (AI companies charge users based on the number of words returned per query.)

“In China, where the focus is on accelerating AI adoption, user growth and developer engagement are crucial,” says Adina Yakefu, a researcher at HuggingFace. “A larger user base fosters a more vibrant ecosystem, unlocking diverse application scenarios that, in turn, drive continuous model improvement.”

…the thinking last year was that the Chinese government was influencing these models and forcing them to behave in a specific way. But now that their capabilities have improved, more people are using them and realizing they can be used for many different applications.

Omar Sanseviero, an expert on the open source AI ecosystem

Figures from HuggingFace illustrate just how quickly the user base has grown. Alibaba’s Qwen was the most downloaded open source model this year. Its platform also boasts more than 75,000 derivative models, making it the most adapted open-source AI model globally. (A derivative model is created when a developer fine tunes an existing model for their own purposes. Models like Qwen or Meta’s Llama enable this because they are open source, meaning developers can download and adapt them independently. Closed-source models, such as those offered by OpenAI or Anthropic, don’t enable this kind of adaptation.) 

It’s difficult to gauge just how many developers outside of China have embraced Chinese AI models. A majority of Qwen’s derivatives were likely created by domestic developers, for example. “Compared to other well-known open-source models like Llama, or Mistral [a French AI company] etc., Chinese open-source models such as Qwen and DeepSeek still lack the same level of global recognition,” notes Yakefu. 

But observers say the reputation of Chinese models overseas is changing. 

“Many people in Europe and the U.S. were initially quite skeptical towards these models,” says Omar Sanseviero, an expert on the open source AI ecosystem. “People were sensitive about the censorship — the thinking last year was that the Chinese government was influencing these models and forcing them to behave in a specific way. But now that their capabilities have improved, more people are using them and realizing they can be used for many different applications.”

One application that DeepSeek has excelled at, and where censorship is less of an encumbrance, is coding. In one test that pitted different AI models against each other and had humans vote on which produced better code, High Flyer’s DeepSeek V2.5 model emerged as a top performer alongside Anthropic’s Claude Sonnet 3.5. 

For policymakers in Washington, the performance and popularity of Chinese AI models raises the question of whether its expansive export controls are hitting their mark. 

“I think it is naive to expect an immediate impact from those controls on training AI models,” says Lennart Heim, a researcher at RAND’s Technology and Security Policy Center, a Washington, D.C. think tank. “We’ve had export controls since 2022, but the thresholds weren’t quite right. So really it’s been since 2023, meaning we’ve only had one year of chip export controls.” 

Heim notes that Chinese companies were able to purchase tens of thousands of chips used to train their AI models from U.S. chipmakers like Nvidia before the U.S.’s tightened export controls took effect. But as the compute demands of training ever more advanced AI models soar, Chinese companies may find it harder to make do with the chips they have in hand. 

“They have a generation of data centers with tens of thousands of chips, but as far as we know they don’t have data centers with hundreds of thousands of chips, which is what [AI companies] are building right now in the U.S,” says Heim.

In an interview with the Chinese outlet 36kr in July, Liang Wenfeng, CEO of High-Flyer, the hedge fund behind DeepSeek, echoed those concerns: “Money has never been the problem for us; bans on shipments of advanced chips are the problem.”

Experts say Chinese companies can’t afford to get complacent after their recent gains as developers benefit from the widening range of AI models on offer globally. 

“One of the main value propositions of open source for developers is that they can easily switch between AI models,” says Sanseviero. “Being able to switch is power. Lots of the tooling for open source is done in a generic enough way such that changing the model is just one line of code.”

---