DeepSeek, a Chinese artificial intelligence startup, trained its latest AI model on NVIDIA Corp.'s most advanced Blackwell chips, a senior Trump administration official said. The model is expected to be made public in the coming weeks.
Such use of Blackwell chips could conflict with existing U.S. export controls because the United States has prohibited sales of its most advanced processors to China. The official's comments raised questions about whether the training activity complied with those restrictions.
According to the same source, DeepSeek may take steps to remove technical indicators that would reveal the model had been trained on American hardware. Those indicators could otherwise provide clues about the underlying infrastructure used in development.
Separately, AI startup Anthropic has accused DeepSeek and two other Chinese firms - Moonshot and MiniMax Group Inc - of harvesting Anthropic's data to build their own models. The allegation was made earlier on Monday and does not include technical detail in the statements released.
Observers note these developments at a time when U.S. lawmakers are debating the appropriate scope of technology export limits to China. Beijing has largely resisted U.S. efforts to restrict tech shipments, and the emergence of activity involving advanced U.S.-designed chips could add pressure to those policy discussions.
Details in the official account indicate the Blackwell processors used by DeepSeek are likely part of a cluster housed at a data centre in Inner Mongolia. It was not immediately clear how U.S. authorities obtained information about the physical location or the chip cluster.
The confluence of alleged use of U.S.-designed high-end chips and accusations of data harvesting by multiple firms highlights two intersecting tensions: hardware access under export rules and intellectual property or data sourcing practices in model training. Companies and markets tied to AI development and advanced semiconductors may face heightened scrutiny as regulators and lawmakers weigh responses.
At this stage, public details remain limited. DeepSeek's model release timeline has been described as imminent, and the firm's potential steps to obscure technical indicators suggest an awareness of the sensitivity surrounding hardware provenance in training workflows.
Summary
DeepSeek trained its forthcoming AI model on Nvidia Blackwell chips despite U.S. restrictions on sales of such processors to China, an official said. The company may remove indicators of American hardware use; Anthropic has separately accused DeepSeek and two other Chinese AI firms of harvesting its data. The situation could feed into ongoing policy debates over technology exports and data practices.