Chinese AI labs are closing the distance with U.S. rivals by doing more with less.
U.S. officials alleged earlier this week that six Chinese AI companies bought bulk subscriptions to American rivals and trained on the outputs, a method the agencies said allowed DeepSeek to understate its $5.6 million training cost. China’s foreign affairs ministry rejected the accusations and called them 'groundless.'
Beyond that dispute, analysts say Chinese labs have built a real advantage around efficiency. Brendan Burke, a semiconductors and supply chain analyst at Futurum Group, said Chinese labs found algorithms that reduce the complexity of attention calculations 'by an order of magnitude' and get better results by summarizing the most relevant tokens.
The backdrop matters. The U.S. restricted China’s access to Nvidia’s best chips, pushing Chinese firms toward domestic alternatives like Huawei. Burke said that because they had less compute to work with, they developed a computationally efficient method instead of 'just throwing more compute at an inefficient technique.' By contrast, the U.S. has 74% of the world’s compute, according to a White House report, and hyperscalers are pouring billions into more data centers.
Ameya Kanitkar, cofounder of the AI measurement platform Larridin, said that in the enterprise workflows his company tracks, Chinese models like GLM 5.2 and Kimi 2.6 and 2.7 handle around 75% of engineering tasks 'reasonably well' at a fifth of the cost of the U.S. ones. He said frontier U.S. models still have an advantage on the most complex tasks, but Chinese open-weight models are becoming capable enough for most everyday enterprise engineering work.
That cost gap is starting to matter. McKinsey found that 20% of business leaders surveyed said AI-related costs like buying tokens are constraining their use of the technology.
Chinese models are also winning attention through openness. DeepSeek’s R1 reasoning model was made available for download through platforms like Hugging Face, letting companies run and adapt versions themselves rather than depend on a closed model. Hugging Face said Chinese open-source models accounted for 41% of total downloads last year, a larger share than U.S. ones.
The market has already voted on one simple fact: capital rewards clear rules and lower costs. When Chinese labs are forced to squeeze value from limited compute, they are building models that can compete on economics, not just hype.
For U.S. firms, the warning is plain. Bigger budgets do not guarantee better margins if the product burns tokens like fuel. In AI, as in every other market, efficiency is not a slogan. It is leverage.