Chinese AI labs have found ways to squeeze more value from limited compute resources, narrowing the performance gap with their US counterparts. US agencies say six Chinese firms have been training on outputs from American models, while analysts point to algorithmic shortcuts that reduce the cost of the attention mechanism at the heart of large language models.
How Chinese firms are closing the gap
Earlier this week US officials from the FBI, the National Security Agency and the Cybersecurity and Infrastructure Security Agency warned that companies such as DeepSeek and Moonshot have bought bulk subscriptions to US models and used the generated data to train their own systems. The agencies claim the practice has extracted "capabilities worth billions" since 2024 and allowed DeepSeek to report a training cost of $5.6 million, far below the true expense.
China's foreign affairs ministry dismissed the accusations as "groundless" and attributed the country's AI progress to a policy of high-level scientific and technological self-reliance.
Efficiency through smarter attention
The key advantage, according to analysts, lies in a refined use of the attention mechanism introduced by Google researchers in 2017. Attention lets a model compare each token with every other token, but the calculation becomes more expensive as the context window grows.
"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 summarise the most relevant tokens," said Brendan Burke, semiconductors and supply-chain analyst at Futurum Group.
Restricted access to Nvidia's top chips forced Chinese firms to rely on domestic alternatives such as Huawei, limiting their compute power. A White House report notes that the United States controls about 74 % of global compute capacity, while its hyperscalers invest billions in data-centre expansion.
Faced with less compute, Chinese labs opted for computationally efficient methods rather than simply scaling up hardware, whereas US frontier labs often consume large numbers of tokens while exploring new model architectures.
Cost advantage in enterprise use
Data from the AI measurement platform Larridin shows that Chinese models such as GLM 5.2 and Kimi 2.6/2.7 handle roughly 75 % of engineering tasks at one-fifth the cost of US models. While US models retain an edge on the most complex problems, Chinese open-weight models are becoming sufficient for most everyday enterprise engineering work.
Cost considerations matter as AI spending takes up a larger share of corporate budgets. A McKinsey survey found that 20 % of business leaders say token-related expenses are limiting their AI adoption.
Western firms warming to Chinese models
Open-source availability has helped Chinese models gain traction. DeepSeek released its R1 reasoning model on platforms such as Hugging Face, allowing companies to download, fine-tune and run the model on US cloud providers like Amazon Web Services.
Hugging Face reported that Chinese open-source models accounted for 41 % of total downloads last year, surpassing US models. Andy Fang, CEO of DoorDash, said using Moonshot AI's Kimi was "cheaper" and "better quality" without compromising code quality. AI coding startup Cursor also uses Kimi for its Composer 2 agent, while Airbnb and Siemens are testing Alibaba and DeepSeek models. Airbnb CEO Brian Chesky described the Qwen model as "fast and cheap".
Even specialised tasks are seeing a shift. Thomson Reuters built an in-house model, Thomson-1, by adapting Alibaba's open-source Qwen model for document-review work previously handled by Claude.
According to Ramp's AI index, the share of businesses purchasing platforms that include open-source and Chinese-developed models rose to 6.1 % of total AI-spending firms in July, up from 4.5 % in January.
Future outlook
US firms are unlikely to abandon domestic models entirely. Mike Finley, chief technology officer at AnswerRocket, said the output of US AI companies still serves as the "existence proof" that enables Chinese labs to innovate.
"The work they do would simply not be possible without the frontier labs blazing the trail," Finley explained.
As Chinese models become more cost-effective and increasingly open, enterprises may continue to blend US and Chinese solutions, driving further optimisation and competition in the global AI market.