DeepSeek Reportedly Used Banned Blackwell Chips for Training, Accused by Anthropic of Distillation Attacks

DeepSeek launched R1 in January 2025, achieving performance close to that of leading U.S. models with about $5.6 million in chip-computing costs and sending shockwaves through the AI industry over its low-cost approach. The controversy now centers on two alleged shortcuts: obtaining advanced chips subject to a U.S. export ban and using Claude outputs for training through “model distillation,” raising questions about U.S.-China AI competition and the effectiveness of export controls.

On February 23, 2026, U.S. officials said DeepSeek trained its latest model in China using Nvidia Blackwell chips, possession of which could itself violate export controls. The model could be released as early as the following week. The same day, Anthropic accused DeepSeek, Moonshot AI and MiniMax of using about 24,000 fake accounts to interact with Claude more than 16 million times. DeepSeek accounted for more than 150,000 of those interactions. The three companies did not immediately respond.