Anthropic Response to AI Export Controls Framework

Anthropic has submitted a detailed analysis and set of recommendations to the U.S. Department of Commerce regarding the "Framework for Artificial Intelligence Diffusion" (the Diffusion Rule). The company asserts that maintaining America's compute advantage through strict export controls on advanced semiconductors is essential for national security and economic prosperity.

The Diffusion Rule and Tiered Risk System

The Diffusion Rule, published in January 2025, establishes global export controls on model weights and advanced AI chips. To manage these controls, the framework implements a three-tier system based on national security risk:

Tier 1: Close allies with minimal restrictions.
Tier 2: Most other countries with some limitations.
Tier 3: Adversarial nations subject to strict controls.

Strategic Rationale for Compute Advantage

Anthropic argues that the U.S. must capitalize on its lead in semiconductor technology to slow the progress of adversarial AI development.

The Efficiency Gap

Because computing power typically doubles every two years, the gap between cutting-edge American technology and older chips used by adversaries is widening. Anthropic estimates that by 2027, countries relying on older chips could face AI training costs ten times higher than those using the latest U.S. technology.

Evidence from DeepSeek

Anthropic cites Chinese AI labs, specifically DeepSeek, as evidence that export controls are effective. The company notes that DeepSeek has openly acknowledged chip restrictions as their primary constraint, requiring them to use 2-4x more power to achieve results similar to U.S. companies. Anthropic suggests that while DeepSeek likely used frontier chips for initial training, future development will be forced onto less efficient Chinese-made chips.

Preventing Infrastructure Offshoring

To avoid a repeat of the strategic vulnerabilities seen in the solar panel and battery industries, Anthropic emphasizes the need to keep AI infrastructure development within the U.S. The company highlights a significant decline in U.S. semiconductor production, which fell from 40% of the global share in 1990 to 12% today, with 90% of leading-edge semiconductors now produced outside the U.S. The Diffusion Framework's domestic compute requirements and foreign compute caps are designed to ensure that frontier-scale training infrastructure is built in America by American firms.

Addressing Enforcement and Smuggling

Anthropic identifies chip smuggling as a major threat to the efficacy of export controls. The company reports that China has established sophisticated operations involving shell companies in third countries and creative smuggling methods, such as hiding processors in prosthetic baby bumps or packing GPUs with live lobsters.

Anthropic's Policy Recommendations

Anthropic recommends three specific enhancements to the Diffusion Rule to close loopholes and improve enforcement:

Adjust the Tiering System: Allow Tier 2 countries with robust data center security to obtain more chips via government-to-government agreements that align technology controls and prevent smuggling.
Lower the No-License Compute Threshold for Tier 2: Currently, Tier 2 countries can purchase the equivalent of 1,700 NVIDIA H100 chips (approximately $40 million) without government permission. Anthropic recommends lowering this threshold to prevent smugglers from making multiple small purchases to avoid scrutiny.
Increase Enforcement Funding: Provide enhanced resources to the Bureau of Industry and Security to improve the overall effectiveness of export controls.

Implementation Timeline

Anthropic warns against delaying the implementation date of May 15, 2025. The company notes that Chinese firms have already engaged in aggressive stockpiling in anticipation of this deadline, and any pause in implementation would further weaken the rule's effectiveness.