Dario Amodei publishes 'On DeepSeek and Export Controls'

Anthropic’s chief executive Dario Amodei published an essay responding to the market and policy reaction to DeepSeek’s R1 release, arguing that it strengthened rather than undermined the case for US export controls on chips to China.

Amodei argued the achievement being celebrated was narrower than reported. R1 itself, he wrote, “essentially replicates” the reinforcement-learning approach OpenAI had already shown with o1; the real advance was DeepSeek-V3, the base model beneath it, which he said was trained for markedly less than comparable Western models. But he rejected treating the widely circulated cost figures as evidence of a discontinuous leap, noting the field’s training costs fall roughly fourfold a year on trend regardless, which made V3’s economics “on-trend at best” — and pointed out that DeepSeek’s total spending, including on chips, still ran to roughly a billion dollars, comparable to major US labs.

On policy, Amodei made the case that cheaper, more efficient training makes chip controls more important, not less: what matters is preventing China from assembling the millions-of-chips-scale clusters he argued would be needed to build near-future frontier systems, since a black market can plausibly move thousands of chips but not the volumes required at that scale without detection. He was explicit that this was an argument about state capability rather than about the individuals doing the work, declining to characterise Chinese AI researchers as adversaries.

The essay was widely read as the clearest rebuttal from a rival lab to the narrative that DeepSeek had proven Western compute advantages didn’t matter, and became a standard reference point in the export-control debate that intensified through 2025.