Securing US Compute Advantage: Anthropic's Position on the 'Diffusion Rule'

Anthropic submitted an analysis and recommendations on the US Department of Commerce's 'AI Diffusion Framework.' The core point is that export controls should be used to maintain the US's Compute Advantage, thereby protecting national security and economic prosperity.

The 'Diffusion Rule,' announced in January 2025, establishes export controls on advanced AI chips and model weights, operating a 3-tier (Tier 1-3) system based on national security risk.

Anthropic argues for strengthening the policy based on the following rationale:

• Importance of Compute Advantage: As the technology gap widens, the AI training costs for countries using older chips will surge sharply.
• The DeepSeek Case: The case of the Chinese company DeepSeek demonstrates that chip restrictions are a substantive constraint, showing the effectiveness of export controls.
• Securing Infrastructure Leadership: The US should encourage domestic infrastructure investment to prevent AI infrastructure buildout from shifting overseas.
• Smuggling Threats: The rule must counter sophisticated smuggling methods, such as hiding chips inside pregnant-belly prosthetics or alongside live lobsters.

Additionally, Anthropic recommended the following to strengthen the rule:

1. Adjusting the Tier System: Allow Tier 2 countries with strong data center security to secure more chips through government-to-government agreements.
2. Lowering the Unlicensed Compute Threshold: Lower the unlicensed purchase limit (approximately $40 million) applied to Tier 2 countries to close smuggling loopholes.