THE WASHINGTON POST: As China advances in AI, Trump faces a new test in the technology race

For months, America’s top artificial intelligence companies have appeared to be comfortably far ahead of their Chinese rivals, launching new flagship models with capabilities that have dazzled their most voracious users.

The new systems could make entire video games, solve knotty academic math problems and identify long-overlooked security holes in computer code.

But in recent days the size of that lead has been called into in question, with the arrival of a wave of new Chinese models that experts say are rapidly closing the gap. And, with the most powerful American models from Anthropic and OpenAI becoming mired in government red tape, the latest Chinese models, Moonshot AI’s Kimi K3 and Alibaba’s Qwen3.8, are set to soon become freely available to anyone with powerful enough computers to run them.

For those worried about America’s lead eroding, an attempt last week to hack American AI company Hugging Face has quickly become a cautionary tale.

The company said in a blog post that when it tried to use commercial AI tools in its defence, it ran into guardrails designed to stop them being used to mount an attack. The solution? Turning to a Chinese model.

At a time when competition on AI has become a key part of the tussle between the United States and China for world leadership, the new releases provide a fresh look at the state of the contest to shape the future a vital technology.

“We’re talking about whether it is the United States or China that provides the digital lifeblood of the global economy,” said Ryan Fedasiuk, a fellow at the American Enterprise Institute, a Washington think tank.

Aware of the stakes, the Trump administration said its approach to the new technology was designed to help the US win the race, aiming to sweep away regulations that might slow American innovation. But in recent months it recalibrated its approach as it has also been forced to contend with the growing computer security risks posed by AI. Officials temporarily hit the brakes on the latest American models, worried about them falling into the hands of hackers but also delaying their public release for several weeks.

“It’s a really hairy position for the US government,” said Sam Bresnick, a fellow at Georgetown’s Centre for Security and Emerging Technology who studies competition with China. “The administration is trying to find its feet.”

The government has a range of options to try to keep the competition at bay and slow the Chinese down, but it is unclear whether the administration will take new action. The White House did not respond to a request for comment on how it is planning to respond.

Jamieson Greer, the US trade representative, said Tuesday that multiple arms of the government were scrutiniinsg how China is developing new AI systems.

“We’re taking a very close look at how China is propagating its AI development to make sure our companies compete ... on a level playing field,” Mr Greer said in an interview on CNBC.

The top systems made by San Francisco’s OpenAI and Anthropic do still appear to hold a lead when it comes to raw power, experts said. British government security researchers concluded that the Chinese model Hugging Face used, which was released in June, was about as good at computer hacking as American offerings from four to seven months ago. But that’s down from a six- to 10-month lag they observed last year.

In cybersecurity, the lead is important because it gives the government and business time to harden their defences before the widespread arrival kinds of sophisticated attacks that models like Anthropic’s Mythos can carry out.

Rob Lee, the chief AI officer at SANS Institute, a cybersecurity training centre, said that while the American industry has been debating how to respond to Mythos’s powerful abilities, Chinese developers have continued to be at work.

“We know by end of year there is going to be a Chinese open weight model that is on par with Mythos,” he said.

Further complicating the picture are the starkly different philosophies of the top AI companies on either side of the Pacific. American companies’ models are closed, made available to customers via subscription plans to cover the vast cost of making them. In contrast, the Chinese companies’ models are open, meaning almost anyone can make their own copy and make changes to it provided they have access to enough computing power.

That’s a cheaper proposition for businesses that don’t need the best of the best, and has won the Chinese models ardent defenders in Silicon Valley. They include influential figures close to Mr Trump, who have warned in recent days against taking measures to block the Chinese tech and tying American companies in regulations.

“America is tying itself in knots: politicians and bureaucrats are banning new data centres, piling on state regulations, and pushing for new federal agencies to pre-approve frontier models,” the investor and former Trump AI adviser David Sacks wrote on X last week. “This is how you lose the AI race.”

While the Chinese AI systems themselves are catching up, the US appears to still have a decisive advantage when it comes to accessing the powerful computer chips and data centres needed to run the technology.

The world’s top AI chips are still made by American companies and are subject to export restrictions. The Trump administration has opened the door to some less powerful chips being sent to China, although an official told lawmakers at a hearing last week that very few had been exported.

Moonshot said on Sunday that it would not allow new subscribers to sign up to preserve resources for its existing customers. “Kimi K3 has received far more love than we expected,” the company wrote on X.

Professor Fedasiuk said if the trend of Chinese models improving continues, it would point to a shift in the nature of the competition between the two countries: “The race will really be about deploying frontier capability not building it.”