Anthropic’s Dario Amodei Questions DeepSeek’s Cost Claims, Bats for Stronger AI Export Controls

In his recent blog, ‘On DeepSeek and Export Controls,’ Anthropic CEO Dario Amodei addressed claims surrounding DeepSeek’s AI model and its implications for US export restrictions on chips to China. He challenged the notion that DeepSeek has achieved a significant cost advantage over US AI companies.

Comparing the outputs, Amodei noted that the claims regarding DeepSeek’s achievement for $6 million, which cost American AI companies billions of dollars, are exaggerated.  

He explained that Anthropic’s Claude 3.5 Sonnet was trained for a few $10 million and was developed 9-12 months before DeepSeek’s model—disputing exaggerated claims about DeepSeek’s cost efficiency.

Amodei noted that DeepSeek’s model matches US models from 7-10 months earlier and was trained at a lower cost, though “not anywhere near the ratios people have suggested.” The cost reduction, he explained, aligns with the industry’s expected 4x annual decrease in AI training costs rather than being a disruptive breakthrough.

Amodei also emphasised the strategic importance of export controls in shaping the future of AI. He outlined two possible scenarios for 2026-2027: a bipolar world where both the US and China develop advanced AI, and a unipolar world where only the US and its allies take the lead AI.

Advocating for the latter, Amodei said, “If we can close [export control loopholes] fast enough, we may be able to prevent China from getting millions of chips, increasing the likelihood of a unipolar world with the US ahead.” This suggests that he supports the US’s AI monopoly rather than mere leadership.

Finally, Amodei addressed concerns over the effectiveness of current export controls. He argued that DeepSeek’s progress does not indicate a failure of these measures. Instead, DeepSeek’s chip inventory—comprising a mix of banned, previously allowed, and still-permitted chips—demonstrates that export restrictions are evolving and closing loopholes.