Chinese AI labs close gap with cheaper 'attention' algorithms, not just alleged training on US outputs

The U.S.-China AI race took a sharp turn this week when the FBI, NSA and CISA alleged that DeepSeek, Moonshot and four other Chinese companies extracted 'capabilities worth billions' by training on American model outputs since 2024. China's foreign affairs ministry called the accusations 'groundless.' But the Fortune reporting suggests the allegations are only one explanation for a narrowing performance gap — a Stanford report put Anthropic's top model just 2.7% ahead of DeepSeek's earlier this year. The other explanation is structural: U.S. restrictions on Nvidia's best chips pushed Chinese labs toward domestic alternatives like Huawei, forcing them to find cheaper ways to run the 'attention' mechanism that underlies every large language model. Futurum Group's Brendan Burke describes the result as algorithms that cut computational complexity by an order of magnitude, while U.S. frontier labs, with access to 74% of the world's compute per a White House report, could afford to be 'token hogs.'

The commercial consequence is already visible in enterprise workflows. Larridin's Ameya Kanitkar says Chinese models like GLM 5.2 and Kimi 2.6 and 2.7 handle around 75% of engineering tasks at a fifth of U.S. cost, and Hugging Face reported Chinese open-source models took 41% of downloads last year. DoorDash, Cursor, Airbnb and Siemens are all cited as experimenting with or adopting Chinese models, and Ramp's index shows the share of businesses paying for platforms with access to open-source or Chinese-developed models rose to 6.1% in July from 4.5% in January. That said, the article is careful to note this is not a wholesale replacement: U.S. models remain months ahead on the most complex tasks, and AnswerRocket's Mike Finley argues Chinese labs' work 'would simply not be possible without the frontier labs blazing the trail.' The test ahead is whether cost pressure — already constraining AI use for 20% of business leaders McKinsey surveyed — pushes more enterprises toward Chinese open-weight models, or whether performance gaps on frontier tasks keep the most demanding workloads on U.S. systems.