US Chip Controls Slow China's AI Progress Despite High-Performing Models
Analysis shows that while Chinese AI models rival US benchmarks, severe compute shortages preserve a strategic frontier advantage for Washington.
The emergence of high-performing Chinese AI models does not signal the failure of US export controls. While Beijing continues to make technical strides, a significant gap remains in practical performance on complex tasks and the compute necessary for large-scale deployment.
Recent data from the Epoch Capabilities Index (ECI) highlights this divide. As of September 2026, the Chinese model Kimi K3 is ranked highly on the index with a score of 158, yet it still trails behind top US models such as Claude Fable 5 (163) and GPT-6 Astra (169). Moonshot AI, the developer of Kimi K3, has acknowledged that the model's overall performance still trails the best available models.
The Compute Gap
Since 2022, the US has maintained stringent export controls on advanced semiconductors and equipment to hinder China's AI capabilities. In response, Chinese firms have released 'open weights' models like DeepSeek V4 Pro and GLM-5.3 that perform well on public benchmarks. However, these firms face immense hurdles in scaling these models for millions of users due to a lack of hardware.
To bypass these restrictions, Chinese entities have turned to illicit channels and regulatory loopholes. According to Epoch AI, significant amounts of hardware have been diverted into China, with estimates suggesting approximately 300,000 H100-equivalent chips were smuggled into the country by the end of 2025. This smuggled compute represents roughly 25% of the total compute China acquired through legal channels or domestic production. Additionally, Chinese firms like Alibaba and ByteDance have utilized remote cloud access via data centers in Southeast Asia, such as Aivres, to circumvent direct hardware bans.
Strategic Implications
The current state of the AI race suggests that the goal of US policy is strategic slowing rather than total prevention. By limiting the scale of deployment and the ability to conduct frontier experiments, the US maintains a critical window to harden its own infrastructure against AI-driven threats. The ability to produce an impressive individual model is distinct from the ability to deploy that model at a national or industrial scale.
What's Next
Industry observers are now watching whether China can close the performance gap through algorithmic efficiency or if domestic chip production can offset the loss of US imports. While the ECI rankings show the gap is narrowing, the reliance on smuggling and rented cloud services indicates that China remains heavily dependent on the very technology the US is attempting to restrict. Whether these loopholes can sustain the growth of Chinese frontier AI remains unconfirmed.