As US-Chinese AI model gap narrows, what next for Washington?

For years, Washington has fixated on the gap between US and Chinese artificial intelligence (AI) models as AI has increasingly taken centre stage in the strategic competition between the world’s two leading powers.

The release of OpenAI’s ChatGPT in late 2022 established a clear early lead for the US, with the first batch of export controls on advanced semiconductor chips introduced in October that year, putting American tech ahead by three years or more, some experts estimated at the time.

But years soon turned to months as China leaned into algorithmic improvements and an AI strategy of “fast following”, most notably with the release of DeepSeek R1 early last year.

In just the last few days, that lead has seemingly shrunk from months to weeks for the first time, following the arrival of new Chinese models from Moonshot AI and Alibaba Group Holding – to the surprise of many in Washington. Alibaba owns the South China Morning Post.

Many in Washington now describe US AI policy as being at a crossroads, with three difficult questions in need of answering: how tightly to control access to frontier US models, whether to restrict open-weight Chinese models at home, and how to drive adoption of US AI technologies globally.

“I don’t think anybody truly knows what’s going on,” said Ryan Fedasiuk, a fellow at the Washington think tank American Enterprise Institute (AEI) focusing on US-China AI competition.

China’s AI breakthroughs trigger US national security concerns

China’s latest AI breakthroughs come at a particularly sensitive time for Washington’s AI establishment.

Last month, the Trump administration abruptly forced leading company Anthropic to suspend access to its latest models, Fable 5 and Mythos 5, over national security concerns.

Some AI experts saw the move as damaging to global perceptions of the US’ reliability as an AI provider – a view that was further underscored when Chinese President Xi Jinping delivered a landmark address on Friday at the World AI Conference (WAIC) in Shanghai, where 28 countries, mostly from the Global South, signed on to China’s vision of “inclusive” AI “for all”.

At stake is not just global leadership in AI capabilities but also AI governance more broadly, as experts warn that AI is now powerful enough to pose significant risks to critical infrastructure and key financial systems.

Central to China’s pitch and the core of its challenge to US AI policy is the country’s continued commitment – so far – to releasing powerful open-weight models that are almost on par with their closed US counterparts.

The perceived national security risks of these US models – kicked off by the “unprecedented” cyberattack capabilities of Anthropic’s Mythos – were what pushed the Trump administration to adopt a more “hands-on” approach to regulating US AI companies in recent months.

Some US national security analysts had assumed that China would do the same and abandon its open-source AI strategy once its models crossed similar national security thresholds – as Moonshot’s Kimi K3 has now done, according to Ben Hayum, a research assistant for the technology and national security programme at the Centre for a New American Security (CNAS).

But the latest Chinese models, Kimi K3 and Alibaba’s Qwen3.8, continue to be released as open-weight models, meaning they are available for free online and can be widely downloaded and customised for many different use cases.

This, in turn, has sparked heated debate about whether the US has ceded global AI leadership to Beijing by erecting barriers around its AI frontier while China has kept its open, particularly after President Xi himself endorsed open-source AI in his WAIC speech.

In recent weeks, users of leading models from OpenAI and Anthropic have increasingly voiced their frustrations with the latest model safeguards introduced by the US companies.

While the companies say the measures are meant to block harmful requests, such as those that could assist in mounting cyberattacks, some users have reported being blocked from using the models for routine tasks such as coding and AI research.

“Kimi K3 just fixed 15 critical security bugs that Codex and Fable refused because of ‘cyber guardrails’,” former White House AI tsar David Sacks wrote in a social media post on Sunday. “There’s no reason to limit American models on tasks that Chinese models handle without issue.”

While it made sense previously for the US to try to control frontier AI as a “scarce commodity”, the proliferation of advanced Chinese open-weight models means there is “a good chance that approach fails”, said AEI’s Fedasiuk.

“An effective US strategy must hedge against a world where AI is cheap and abundant and uncontrollable in its proliferation,” he said.

Instead, the US should now adopt a more “forward-looking” strategy of accelerating the buildout of AI infrastructure globally in countries like the United Arab Emirates and Saudi Arabia to deploy US AI models, Fedasiuk said, as the US and China now shift to an “all-out infrastructure competition” rather than a competition over who has the best AI model.

The view is based on the assessment that while US export controls have not stopped China’s AI industry from developing frontier AI models competitive with US ones, they do slow its rate of expansion by limiting the compute it can use to serve its users.

On Sunday, Moonshot said it had to suspend new user subscriptions just days after releasing its powerful new model due to a compute shortage, making it the latest Chinese tech company to report difficulties in deploying its powerful models.

Another key uncertainty is how Washington will regulate these powerful Chinese open-weight models domestically, which have steadily grown in popularity in the US over the past year.

Many in Washington’s national security establishment support introducing restrictions on Chinese models, arguing that they pose significant supply chain and security risks.

Meanwhile, Sacks and other former administration officials such as Sriram Krishnan, former White House senior policy adviser for AI, have called for a more hands-off approach to allow for the proliferation of American open-weight AI instead, highlighting the emergence of Kimi K3 as an example of how the US “lose[s] the AI race”.

The deep divisions in US AI policy circles have come to the fore in recent days, with Axios reporting on Monday that Kimi K3 had emboldened parts of the Trump administration to push for broad bans on Chinese open-weight models, while other “pro-competition” officials remain opposed.

“Some US restrictions on adoption of Chinese models could make sense given the risk of ideological bias and sleeper agents,” said CNAS’ Hayum, highlighting new research that suggests that AI models could have “secret loyalties” where they are covertly advancing the interests of “actors of concern” such as China.

Trump administration ‘supports open source models’, says Bessent

Treasury Secretary Scott Bessent said on Tuesday that the Trump administration “supports open source models” while threatening to sanction Chinese companies found to have illicitly trained their models on the outputs of US ones, a practice commonly referred to as “adversarial distillation”.

He also hinted during his appearance on Fox Business that US companies may eventually be required to disclose to their customers their use of Chinese models.

Wang Yaqiu, a fellow at the Penn Project on the Future of US-China Relations, said that Washington is likely to shift towards highlighting the ideological differences between the US and Chinese models, as they become harder to differentiate on capability alone.

But even then, there are question marks about whether the Trump administration is committed to doing so.

The administration previously tasked the Centre for AI Standards and Innovation (CAISI), a technical body that operates under the National Institute of Standards and Technology, with evaluating Chinese models on their “alignment with Communist Party talking points and censorship”, but the body has stopped publishing such findings in its latest reports.

On Monday, a spokesperson for the US Department of Commerce confirmed that CAISI’s director, Chris Fall, had left his role after just three months on the job, raising questions about the future of the AI testing body.

According to CNAS’ Hayum, a central risk for US AI national security policy at this critical juncture in US-China AI competition is continued indecision: “If [the Trump administration] decides to stay ambiguous and not put forward a vision to at least begin the conversation, I think that would be a missed opportunity.”

The White House and the Department of Commerce did not immediately respond to requests for comment.