THE ECONOMIST: AI labs in the US are under threat from cheap Chinese rivals as demand soars

Spare a thought for America’s leading developers of artificial intelligence. They pour blood, sweat and gazillions of dollars into each new release of a frontier model.

Yet before they have time to catch their breath, their competitors have already caught up.

It would be bad enough if rivals that charge for their technology were the only ones nipping at their heels. But the likes of Anthropic and OpenAI must also sprint to stay just a few steps ahead of labs that give their models away for free.

On July 16 Moonshot AI, a Chinese start-up, released Kimi K3 — a model that is only slightly less capable than Fable or Sol, the latest offerings from Anthropic and OpenAI, respectively.

Its parameters (or “weights”) will be made publicly available at no cost later this month, meaning it can be downloaded and run on any server capable of doing so.

On July 19 Alibaba, a Chinese internet giant, also unveiled a preview version of a new addition to its Qwen family of models that the company says is second only to Fable.

According to analysis by Epoch AI, a research firm, the most powerful open-weight models, which typically come from China, are a matter of months behind the technology’s frontier.

Adoption of these systems is spreading rapidly. In May customers of OpenRouter, a popular marketplace for models, used the systems offered by the top American labs about as much as those of the top Chinese ones, going by their consumption of tokens, the industry’s preferred unit of output.

The following month use of American models was up by two-thirds — but had tripled for Chinese ones.

That is causing consternation not just among America’s leading model-makers, but also in its government, which fears the country losing its lead in AI and creating a new opening for hacking. Some within the Trump administration are said to be pushing for various measures to slow the spread of Chinese models, such as adding their makers to trade blacklists or warning American companies against their use.

There are also grumbles about Chinese labs supposedly using American models to train their own systems, a process known as “distillation”. Yet the American government’s unpredictable approach to granting access to advanced models such as Fable and Sol is also spurring adoption of China’s open-weight alternatives.

Even before K3, China’s open-weight models were drawing interest around the world. GLM-5.2, a model launched in June by Z.ai, another Chinese lab, is better at so-called agentic tasks—such as creating software—than anything currently available from Google, another of America’s top model-makers. Models from DeepSeek and MiniMax, two other Chinese labs, have also proved popular.

According to a report by Mozilla, which makes the Firefox browser, open-weight models remain behind the frontier in reasoning and complex agentic work, but are now widely perceived as “good enough” for many tasks.

They also cost a lot less — an important advantage as companies grapple with the rising cost of AI. Artificial Analysis, another research firm, tracks the cost it takes selected models to perform each task in its battery of tests. The most capable system offered by DeepSeek costs an average of $US0.04 per job. K3, which is not yet open-weight, costs an average of $0.95. Users of Fable can expect to pay $US2.75.

Part of the cost difference comes from the fact that more or less any cloud provider can sell access to an open-weight model, driving down profit margins.

DeepSeek, for example, offers access to its v4 flash model at a cost of $US0.28 per million tokens, whereas the cheapest third-party provider offers it for $US0.18 per million.

What is more, many open-weight models are trained with a specific focus on running efficiently, partly in response to the challenges Chinese labs face in gaining access to computing power owing to American restrictions on advanced chips.

Innovations like “mixture of experts” models, which activate only the part of an AI system that is relevant to a query, have been used extensively by open-weight developers, with labs such as DeepSeek and Alibaba making big strides in the technique.

The appeal of open-weight models, however, goes deeper. The three weeks in which Fable was unavailable globally, owing to a hurried attempt by America’s government to withhold the model from all but American citizens, prompted many companies to examine their dependence on individual labs.

It encouraged them to experiment with open-weight alternatives that can be accessed through many providers and perhaps even run on internal servers. The episode served as a “shot across the bow”, notes Raffi Krikorian, Mozilla’s chief technologist.

Still, China’s open-weight models may not be entirely dependable for long.

Its government is said to also be looking at ways to control foreigners’ access to its technology, which could involve restricting the ability to download its open-weight models outside China. America’s labs would surely be delighted.