China and the United States, two recipes for the AI cake

Last Saturday, the CEO of Anthropic, Dario Amodei, published on his website the article "We must pace the frontier", in which he asked the AI industry to slow down the pace of improvement of their models. Five days later, the rotating chairman of Huawei, Eric Xu, responded to him from Shanghai: Chinese AI providers "still do not have enough computing or model capacity" to perceive the same risks that are being discussed in the US, and perhaps "they need to accelerate the pace until they can notice them too". One proposes to brake because he already sees them (and also because he does not want any rival to get ahead of him); the other wants to run to get there. Underneath this debate lie two technically very different ways of building artificial intelligence.

Open weights versus closed weights

The most visible difference between the US and China is who can look inside AI models. The American ones Claude (Anthropic), GPT (OpenAI), and Gemini (Google) are closed-weight: no one outside each company knows the exact configuration of the millions of parameters that determine their behavior. In contrast, the Chinese ones DeepSeek, Qwen (Alibaba), Kimi (Moonshot), and GLM (Z.AI) publish theirs with open licenses: anyone can download, inspect, modify, and run them on their own computer without having to pay for each token consumed.

This openness of Chinese models favors their spread, but it also implies a loss of control for their creators. On the Hugging Face repository, there are already more than 150,000 models derived from Qwen, 2.6 times more than Meta AI. In return, once an open model is published, its weights cannot be withdrawn or modified remotely, as Anthropic and OpenAI can do with theirs.

On the other hand, Washington sees it as a leak route for US technology. A joint alert from the National Security Agency, CISA, and the FBI accuses six Chinese companies (DeepSeek, Moonshot, Alibaba, MiniMax, StepFun, and Z.AI) of distilling variants of Claude, GPT, Gemini, and Grok on an industrial scale, that is, of training their models based on the responses of the American ones, a much cheaper method than starting from scratch.

Brute force versus efficiency

The other difference, more technical, is how models are trained. OpenAI's Stargate project plans to invest 500 billion dollars and 10 gigawatts of capacity, and the big five (Microsoft, Alphabet, Amazon, Meta and Oracle) will add up to 690 billion just this 2026. The principle is simple: whoever puts more resources into it (money, data) gets more computing power.

In contrast, DeepSeek demonstrated in 2024 that it can be achieved with a fraction of that expenditure: the final training of DeepSeek-V3 took 2.79 million processor hours, about 5.6 million dollars at the usual rental price, thanks to two techniques: FP8 precision, which performs calculations with fewer decimals and, therefore, faster and cheaper, and the mixture of experts (MoE) architecture, which activates only a part of the neural network in each response instead of the entire model. The figure does not include the total cost of the project, but it shows that algorithmic efficiency compensates for the lack of economic muscle.

Stanford University's annual report on the state of AI, published in April, summarizes the race: the performance gap between the best American models and the Chinese ones is only 2.7 percentage points, even though private investment in AI in the United States is 23 times that of China. Three years ago this gap exceeded 17 percentage points.

The chip bottleneck

The Chinese obsession with efficiency is mandatory. US technology export restrictions have left Nvidia practically out of the Chinese data center market: the company itself acknowledges that it has not delivered a single H20 chip to China in the last quarter. Huawei is responding by manufacturing its own chips: it plans to produce nearly 600,000 Ascend 910C in 2026, double that of last year, made by SMIC at 7 nanometers, two generations behind Nvidia's high-end range.

To compensate for the power lacking in each chip, Huawei interconnects many more; its CloudMatrix system matches the performance of an Nvidia unit with four times more hardware and consuming four times more electricity. Eric Xu acknowledged this in Shanghai: Huawei is "not yet as advanced" as its North American rivals in semiconductors, but the restrictions have pushed its customers toward national technology, and they are now pursuing "complete self-sufficiency" in chips and supply chain.

Two concepts of risk

This technical difference translates into the way risks are understood, which are monopolizing public debate in a way that is probably exaggerated. In the US, the debate revolves around whether a model can become uncontrollable, and Anthropic and OpenAI have spoken openly about it. There are also those who go beyond the self-regulation proposed by Amodei: Palantir CEO, Alex Karp, proposed last Thursday to nationalize AI laboratories and impose civil and criminal liability on those who "are not responsible". In China, experts largely agree on the substance, but not on the form. Beijing considers AI a powerful but governable technology, and is already drafting the world's first national standard for the safety of AI agents.

The official Chinese response to Amodei's article took only a few hours to arrive. Foreign Ministry spokesperson, Guo Jiakun, said that "fear, confrontation and misunderstood competition" harm global AI governance, and the state newspaper Global Times branded the text a "Cold War manual". The day before, at the BRICS summit, Xi Jinping had announced an open-source AI community among the countries that make up the bloc. Xi and Donald Trump will meet face-to-face on September 24 in Washington, and Trump has already said that AI will have a prominent weight there.

Mind you, the porosity between the two blocs became evident when Hugging Face had to investigate the attack by OpenAI agents it suffered in July: it ended up doing so with GLM-5.2, an open-weight Chinese model from Z.AI, because the American ones, being more restricted, proved less useful for the forensic work.

Europe, between two waters

Between these two blocs, in Europe we do not have any frontier laboratory on par with the Americans or the Chinese. Our supposed champion, the French company Mistral AI, raised 3 billion euros in new funding last week, but its main data center runs on 13,800 Nvidia processors, so sovereignty is very relative. The EU AI Act has had to delay the obligations for high-risk systems until 2027, and a Commission spokesperson has limited themselves to reminding companies that they must demonstrate that they are safe to operate in the Union. Meanwhile, Christine Lagarde, the president of the European Central Bank, adds fuel to the fire by warning that our continent may become "disconnected" from AI.