Anthropic’s Dario Amodei responds: doesn’t oppose open-weight models, but fears Chinese AI

Dario Amodei just drew the industry's most consequential distinction in months. Open-weight models are not the enemy, he says. The Chinese Communist Party is. Anthropic's CEO published his position Monday after Nvidia's Jensen Huang rallied Hugging Face, Meta, Microsoft, and Mistral behind an open letter opposing "premature restrictions" on open weights. That letter named no country. Everyone knows which country it meant.

Amodei's response is clarifying precisely because it refuses the easy binary. He calls open-weight models without dangerous capabilities a public good — free compute aside, they fuel businesses, developers, researchers. He has never advocated banning them. He emphasizes the point. The industry needed to hear it.

But he pivots instantly to the threat that actually keeps him awake. Authoritarian governments building models more powerful than America's to lock in permanent military superiority. To repress their own populations. The CCP is not the only such government, he allows. It is merely the most capable. That sentence does heavy lifting. It frames the conflict as structural, not incidental. The technology is neutral. The regime is not.

His threat model extends beyond cyber. Biological attacks. Open weights amplify that risk because guardrails cannot be enforced on released models. Once out, they cannot be recalled — a point he anchors to a UK AI Security Institute report. This directly contradicts the open-source catechism: that transparency lets defenders protect themselves. Amodei calls that faith dangerous. He may be right. The asymmetry favors offense. A single released model with pathogen-design capability outruns any patching cycle.

His policy prescriptions are aggressive. Maintain the chip embargo — already U.S. doctrine. Criminalize distillation, the prompt-bombing technique Chinese labs allegedly use to reverse-engineer American models. Sanction entities caught stealing IP. These are not safety measures. They are economic warfare tools. Amodei knows this. He deploys them anyway.

Then comes the fantasy. He supports a global model safety testing organization. He wants the entire world — including China — at the table. This is where the argument fractures. Beijing will not submit its frontier models to a U.S.-led audit regime. The CCP treats AI supremacy as sovereign survival. No treaty, no inspection regime, no multinational body will pry those weights loose. Amodei surely understands this. The proposal functions as rhetorical cover: it lets him sound cooperative while pursuing containment.

The industry should notice what he did not say. He did not call for compute governance on domestic open-weight releases. He did not demand licensing regimes for Llama or Mistral derivatives. He drew the perimeter exactly where Anthropic's commercial interests align with national security hawks. Closed models for the frontier. Open weights for the commodity layer. Chinese labs for neither.

Critics will call this convenient. They are not wrong. But convenience does not invalidate the threat assessment. The CCP has disclosed its intent: AI as the lever for military leapfrog and internal control. Distillation is happening. Chip smuggling networks exist. Amodei's fear is not performative. His policy toolkit, however, bears the fingerprints of a company that sells closed models to defense contractors.

The open letter Huang circulated represents the other flank: companies whose moats are compute and distribution, not model secrecy. They want the commons open. They want regulators away from their supply chains. Their argument is principled and self-interested simultaneously. So is Amodei's.

The middle ground is vanishing. Washington will pick a side. The CHIPS Act already chose. The next legislation will target distillation. Amodei just handed lawmakers the intellectual framework. Whether the resulting regime actually slows Beijing — or merely consolidates the American closed-model oligopoly — is the question the industry should be asking. Amodei has answered his own version. The rest of us get to live with the consequences.