Anthropic National Security Policy Head Discusses US-China AI Competition, US AI Models Lead China by 6-9 Months
Tarun Chhabra, head of national security policy at Anthropic, stated at the Aspen Security Forum that US AI models are 6-9 months ahead of China; the productivity of AI hardware (data centers, chips) is 30-40 times that of China, with smaller advantages in energy and talent; GLM 5.2 is the most advanced model in China, distilled from Claude and GPT data; without GPU restrictions, China's AI should be on par with or even ahead of the US.
This viewpoint highlights the critical role of hardware restrictions in the US-China AI gap.
In market mechanisms, hardware advantages support US AI leadership, event-driven investments in chips and data centers, with beneficiaries of export controls under pressure, while China accelerates independent hardware development.
Source: Public Information
ABAB AI Insight
Anthropic has previously commented little on Chinese models; Chhabra's remarks continue the US AI security circle's assessment of the effects of hardware restrictions, similar to other think tanks' analyses of the US-China AI gap.
In terms of capital pathways, the US maintains hardware export restrictions, directing resources towards domestic chips and data centers, motivated by the desire to extend the lead window through supply chain control while promoting coordination among allies.
Models like DeepSeek have been accused of distillation; this is the first time GLM 5.2 has been specifically named by Anthropic. Currently, the US-China AI competition is dominated by hardware barriers, and the 6-9 month gap reflects the current situation.
Essentially, this is a regulatory change; export controls are reshaping the global AI supply chain, with the mechanism being that hardware gaps amplify model performance differentiation, concentrating capital in a US-led AI hardware ecosystem, while China accelerates independent alternatives to narrow the gap.
ABAB News · Cognitive Law
Hardware leads by 30 times, models only lag by 6-9 months, regulation determines the pace.
Distilling data is easy, independent hardware is difficult; the root of the gap lies in the supply chain.
Energy and talent are catching up; AI leadership relies on hardware barriers.