Anthropic submitted a detailed response to the U.S. Department of Commerce’s interim final rule called the "Framework for Artificial Intelligence Diffusion" (the Diffusion Rule). In its filing, Anthropic argues that stronger export controls on advanced semiconductors are necessary to preserve America’s compute advantage, which it says is essential for national security and economic prosperity as more powerful AI systems are developed.
What the Diffusion Rule does
Published in January 2025, the Diffusion Rule imposes export controls on advanced AI chips and model weights worldwide. It establishes a three-tier system based on national security risk: Tier 1 covers close allies with relatively few restrictions, Tier 2 covers most other countries with some limits, and Tier 3 covers adversarial states subject to strict controls.
Anthropic notes that earlier U.S. policy makers correctly identified AI as central to strategic competition with China and that export controls are a legitimate tool to maintain U.S. leadership in AI.
Why compute advantage matters
Anthropic bases its recommendations on several core points:
- Advanced AI systems require massive compute resources to train and run. The United States currently leads in advanced semiconductor technology, and export controls take advantage of the industry trend that computing power roughly doubles every two years. Maintaining that lead slows rivals’ progress.
- According to Anthropic’s analysis, growing efficiency gaps mean that by 2027 countries relying on older chips could face training costs up to ten times higher than those with access to cutting-edge American technology.
DeepSeek as an illustrative case
Anthropic cites the Chinese AI lab DeepSeek as evidence that chip restrictions can meaningfully constrain capabilities. It says DeepSeek has made progress using chips obtained before controls took effect, but that restrictions have forced Chinese firms to rely on less efficient domestic chips. Anthropic also reports that DeepSeek needs roughly 2–4 times more power to reach comparable results to U.S. firms when constrained to non-frontier chips.
Risks from offshoring infrastructure and declining domestic production
The company warns that without strong controls, the buildout of frontier-scale AI infrastructure could move overseas — mirroring past shifts in industries such as solar panels and batteries — which would undermine U.S. strategic advantage. Anthropic highlights long-term trends in semiconductor production: U.S. share of global semiconductor manufacturing fell from 40% in 1990 to about 12% today, and roughly 90% of the world’s leading-edge semiconductors are produced outside the United States. That offshoring is presented as a strategic vulnerability to be avoided for AI infrastructure.
The Diffusion Framework attempts to address this risk by including domestic compute requirements and limits on foreign compute, creating incentives for large-scale advanced compute to be hosted on U.S. infrastructure by U.S. companies.
Chip smuggling is a significant threat
Anthropic emphasizes documented cases of sophisticated chip-smuggling operations linked to China, involving hundreds of millions of dollars of chips. The filing notes creative smuggling methods reported in the record, including processors hidden in prosthetic baby bumps and GPUs shipped alongside live lobsters. It also points to rapid creation of shell companies in third countries by Chinese firms as a mechanism to evade export controls.
Anthropic’s three recommendations to strengthen the rule
-
Adjust the tiering system
- Allow Tier 2 countries with robust data-center security to receive more chips via government-to-government agreements that reduce smuggling risk and align technology controls.
-
Lower the no-license compute threshold for Tier 2 countries
- Currently, Tier 2 countries may acquire the equivalent of 1,700 NVIDIA H100 chips without a license — roughly $40 million worth of technology. Anthropic argues this creates a loophole exploitable through multiple sub-threshold purchases and recommends lowering the threshold so more transactions require review.
-
Increase funding for export enforcement
- Export controls are only effective if enforced. Anthropic recommends greater resources for the Bureau of Industry and Security to improve enforcement capability.
Risk of delaying implementation
Anthropic warns that delaying the rule would invite further stockpiling; it notes that some Chinese firms engaged in aggressive stockpiling ahead of the May 15, 2025 implementation date. The filing argues that any pause would undermine the rule’s effectiveness at a critical moment.
Conclusion
Anthropic concludes that the strategic window to strengthen American export controls is now. By tightening the Diffusion Framework, the United States can encourage development of transformative AI technologies domestically and ensure alignment with American values and interests. Anthropic’s filing references prior public interventions by its leaders, including a Wall Street Journal op-ed by Dario Amodei and Matthew Pottinger ("Trump Can Keep America’s AI Advantage," January 6, 2025) and other writings and interviews by Dario Amodei discussing competition with China and export controls.
Sources and references cited in the filing
- U.S. Department of Commerce: "Framework for Artificial Intelligence Diffusion," January 2025
- Anthropic submission to the Diffusion Rule, published April 30, 2025
- Planned implementation date of the rule: May 15, 2025
- Current no-license threshold for Tier 2: equivalent to 1,700 NVIDIA H100 chips (approximately $40 million)
- Semiconductor production statistics cited in the filing: U.S. share from ~40% in 1990 to ~12% today; ~90% of leading-edge semiconductors produced outside the U.S.
- Mention of Chinese AI lab DeepSeek as an example cited in the filing
