Anthropic backs US 'AI Diffusion' chip export framework

Anthropic published a formal comment backing the Commerce Department’s Framework for Artificial Intelligence Diffusion, the tiered chip-export rule the outgoing Biden administration had issued on 13 January 2025 and which was due to take effect that May. Submitted to the Federal Register as the incoming Trump administration weighed whether to keep, rewrite or scrap the rule, the comment argued for retaining and tightening it rather than dropping it.

Anthropic’s specific recommendations included lowering the no-licence compute threshold for “Tier 2” countries — then set at roughly 1,700 Nvidia H100-equivalent chips, about $40 million worth of hardware — to close a loophole it said let buyers split large orders into multiple under-the-limit purchases, and increasing funding for the Bureau of Industry and Security’s enforcement capacity. The company argued chip controls were already working, citing what it described as evidence that Chinese developers, including DeepSeek, needed two to four times more power to achieve comparable results to US models because of restricted access to advanced chips.

The position put Anthropic at odds with Nvidia, which had lobbied against the framework as written; Nvidia’s chief scientist publicly accused Anthropic of “telling tall tales” and exaggerating chip-smuggling methods, including a claimed case of chips hidden in prosthetic baby bumps, to justify tighter controls that would depress Nvidia’s addressable market in China. The underlying rule did not survive in its original form regardless: days before its 15 May 2025 compliance deadline, BIS announced it was rescinding the Diffusion Rule rather than enforcing it, leaving the tiering fight unresolved and folding into the Trump administration’s own, separately negotiated approach to AI chip exports.