Are the US and China Really Running the Same AI Race?

One of the online papers I've come to value is in Italian, and thanks to a language model I read it more and more often. It offered me a view of the race between the United States and China that stands apart from all the hype, including the hype I help create.

Il Post starts from a line of Trump's. On 12 September Dario Amodei, chief executive of Anthropic, asked the industry to slow down, and Sam Altman and Elon Musk said they agreed. The next day the US president replied that "we can put guardrails, we can do this and that", but "whoever wins with AI wins". The paper asks the question nobody in Washington has asked in public: what does winning mean, and what happens to whoever loses?

Its answer is that there are two races. The United States is betting on a breakthrough: the first lab to reach artificial general intelligence, a system able to do everything a person can, takes a lead nobody will ever close. China, with less computing power because of US controls on chips, is betting on diffusion: open models that anyone can download and run, and AI built into the economy, from hospitals to infrastructure to surveillance.

The numbers support much of this reading. According to Epoch AI, a research institute that measures what models can do, every frontier model since 2023 has been built in the United States, and Chinese models arrive seven months later on average, with a gap ranging from four to fourteen. On usage the picture flips. In July Chinese models passed 60 per cent of the traffic on OpenRouter, the platform that routes developers' requests to different models, and held the top five places. It is a partial measure: that traffic leaves out enterprise contracts, where OpenAI and Anthropic still earn most of their revenue. In August 2025 China's State Council set itself a diffusion target: intelligent agents and devices in 70 per cent of key sectors by 2027.

The most useful point in the article is still the question about the finish line, and I offer my reading of it as a hypothesis. A race with no finish line never ends, and as long as it runs, any rule can be presented as unilateral disarmament in front of the rival. Trump used it exactly that way, against the companies asking to slow down. David Sacks, who co-chairs a White House advisory council on science and technology, answered those companies on social media that they are the ones setting the pace: "The easiest way not to build superintelligence is for you to agree not to build it." The administration holds two positions at once: the race is a national necessity, and the runners can stop whenever they like. On 23 September Yoshua Bengio told the UN Security Council in his own words that the race is "the product of choices, choices made by the companies themselves".

Of course the two-race scheme is tidier than reality, and Il Post admits as much in its last paragraph. Beijing has written general intelligence into its government documents. The Jamestown Foundation, a US think tank, traces the thread from the State Council plan of 2017 to Xi Jinping's call in 2018 to "bravely venture into the uncharted zones of AI's frontiers", up to April 2025, when he urged the Politburo to "seize the decisive opportunity and win the advantage" in AI. Open models fly a US flag too: on 24 July Nvidia promoted a letter presenting open models as a tool of US leadership, and on 3 September it agreed to buy Hugging Face, the platform where those models circulate. The White House action plan of July 2025 is titled "Winning the Race", and one of its sections is devoted to adoption.

One more detail makes Il Post's reading more interesting than the paper itself says. The idea that a great power is built by spreading a technology more than by inventing it has an author, the US political scientist Jeffrey Ding, who developed it in Technology and the Rise of Great Powers (Princeton, 2024). Studying Britain and steam, the United States and electricity, Japan and computers, Ding concludes that in the diffusion of AI the institutional advantage, meaning the schools, skills and firms able to absorb a technology, currently lies with the United States. The race Il Post assigns to China is the one in which, according to the man who theorised it, Washington starts ahead.

The frame of the race serves both sides anyway. In Washington it justifies the absence of rules; in Beijing, the mobilisation of an entire industrial system. On 20 September the US Treasury Secretary, Scott Bessent, called the two countries "the number one and the number two AI powers" while proposing a direct channel for incidents. Competitors in different races don't exchange phone numbers. Anyone watching from outside, which for a European is the natural position, can note that the German institute Merics read the Chinese target as a warning for Europe: China has already overtaken Germany and Japan in robots per worker. For anyone who uses these systems, the practical question shifts. The race asks which team you support. Diffusion asks what you do with the machine, and who chose the settings it comes with.

Today Trump and Xi meet in Washington. Nobody has drawn the finish line yet, and Il Post deserves credit for noticing. What remains to be seen is whether either of the two has any interest in drawing it.