Chinese AI Labs Narrow U.S. Gap Through Greater Efficiency

Chinese artificial intelligence companies are narrowing the gap with U.S. rivals partly by finding ways to make better use of limited computing resources.

Chinese AI developers are using more computationally efficient techniques to compete with U.S. companies despite restrictions on access to advanced semiconductors. [Photo: Fang Zhe, Xinhua]

Fortune reported that six Chinese AI companies have been accused by U.S. officials of buying bulk subscriptions to American AI services and using their outputs to train their own systems.

The companies identified in the report include DeepSeek and Moonshot. Chinese officials have rejected the allegations as “groundless.”

Those allegations are one possible explanation for the rapid progress of Chinese AI models.

But analysts cited by Fortune say another important factor is efficiency: Chinese laboratories have developed techniques that allow them to obtain more performance from limited computing resources.

One area of innovation involves the “attention” mechanism that underpins modern large language models. Introduced in the landmark 2017 paper “Attention Is All You Need,” the mechanism allows AI models to determine relationships among different pieces of information in a sequence.

According to Brendan Burke, a semiconductors and supply-chain analyst at Futurum Group, Chinese researchers have developed techniques that reduce the computational complexity associated with attention by an order of magnitude while summarizing the most relevant tokens.

The pressure to improve efficiency has been intensified by U.S. restrictions on China's access to the most advanced Nvidia chips, encouraging Chinese companies to develop domestic alternatives and more efficient methods of using available computing power.

Fortune noted that the United States continues to hold a substantial advantage in overall computing capacity.

The cost difference can be significant. Ameya Kanitkar, co-founder of AI measurement platform Larridin, told Fortune that in enterprise workflows tracked by the company, Chinese models handle about 75% of engineering tasks “reasonably well” at roughly one-fifth the cost of U.S. models.

That does not mean Chinese models have overtaken U.S. frontier systems.

Fortune reported that U.S. models still retain an advantage on the most complex tasks, while Chinese open-weight models are becoming increasingly capable for routine enterprise engineering work.

The shift is already beginning to appear among businesses.

Fortune reported that companies including Airbnb and Siemens are experimenting with Chinese-developed models, while Thomson Reuters has adapted Alibaba's Qwen model for document-review work.