Chinese AI models are narrowing the gap with American systems not only through access to their rivals’ outputs, US officials allege, but by making better use of limited computing resources.
The FBI, National Security Agency and Cybersecurity and Infrastructure Security Agency said six Chinese companies, including DeepSeek and Moonshot, had bought bulk subscriptions to US AI services and trained on their outputs since 2024. The agencies claimed the approach extracted “capabilities worth billions” and helped DeepSeek understate its training cost of 5.6 million dollars.
China’s foreign affairs ministry rejected the allegations, saying the country’s progress was the result of “high-level scientific and technological self-reliance” and describing the claims as “groundless”.
The accusations offer one possible explanation for the performance of Chinese models, which are now estimated to be close to their US counterparts. A Stanford report earlier this year put Anthropic’s leading model just 2.7 per cent ahead of DeepSeek’s.
How Chinese AI models are cutting computing costs
Analysts say Chinese laboratories have developed another important advantage: squeezing more value from every unit of computing power.
The approach centres on attention, the mechanism underlying large language models that allows them to assess how each part of a piece of text relates to the others. As the amount of text increases, the calculations required become more expensive.
Brendan Burke, a semiconductors and supply chain analyst at the Futurum Group, said Chinese laboratories had found ways to reduce the complexity of those calculations substantially while identifying the most relevant pieces of text.
“Chinese labs found algorithms that reduce the complexity of those calculations by an order of magnitude, and then achieve better results because they’re able to summarize the most relevant tokens,” he said.
US restrictions on Nvidia’s most advanced chips have limited China’s access to the highest-performing computing hardware, pushing its developers towards domestic alternatives such as Huawei. By contrast, the US has 74 per cent of the world’s computing capacity, according to a White House report, while major cloud companies are investing billions of dollars in new data centres.
“Because they had less compute to work with, they found that computationally efficient method instead of just throwing more compute at an inefficient technique, as U.S. labs initially did,” Mr Burke said.
He said leading US laboratories could afford to use large amounts of computing power to explore and test their base models, making them “token hogs”.
Ameya Kanitkar, co-founder of the AI measurement platform Larridin, said Chinese models including GLM 5.2 and Kimi 2.6 and 2.7 handled about 75 per cent of the engineering tasks tracked by the company “reasonably well”, at a fifth of the cost of US systems.
“Frontier U.S. models still have an advantage on the most complex tasks, but Chinese open-weight models are becoming more than capable enough for the majority of everyday enterprise engineering work,” he said.
The cost gap could become increasingly significant as AI takes up a larger share of corporate budgets. One-fifth of business leaders surveyed by McKinsey said expenses such as buying tokens were restricting their use of the technology.
US businesses experiment with Chinese models
Cost is not the only attraction for companies. DeepSeek’s R1 reasoning model was made available for download through platforms including Hugging Face, allowing businesses to run and adapt it themselves, including through US-based cloud providers such as Amazon Web Services.
That flexibility has helped reduce concerns among some US companies about sending data to a China-based provider. Chinese open-source models accounted for 41 per cent of total downloads on Hugging Face last year, a larger share than US models.
DoorDash chief executive Andy Fang said Moonshot AI’s Kimi was “cheaper” and offered “better quality” without reducing the quality of code. AI coding company Cursor has also used Kimi while developing its Composer 2 coding agent.
Airbnb and Siemens are experimenting with models from Alibaba and DeepSeek. Airbnb chief executive Brian Chesky described Alibaba’s Qwen as “fast and cheap”.
Chinese models are also being used for more specialised work. Thomson Reuters developed an internal system, Thomson-1, by adapting Alibaba’s open-source Qwen model for document review previously handled by Claude.
Ramp’s AI index showed that the proportion of businesses paying for platforms with access to open-source and Chinese-developed models rose from 4.5 per cent of AI-spending businesses in January to 6.1 per cent in July.
That shift does not mean US systems have been abandoned. Mike Finley, chief technology officer at enterprise AI analytics firm AnswerRocket, said American models remained months ahead on performance and provided the “existence proof” from which Chinese laboratories could develop their own innovations.
“The work they do would simply not be possible without the frontier labs blazing the trail,” he said.
