Congressional review of company records indicates that Nvidia personnel actively supported a Chinese artificial intelligence startup, DeepSeek, providing technical guidance that the House Select Committee on China says contributed to the development of models later used by the Chinese military. The documents cite hands-on collaboration between Nvidia staff and DeepSeek engineers focused on aligning algorithms, software frameworks, and hardware to improve efficiency.
Representative John Moolenaar, in a letter referenced by the committee, highlighted the outcome of that collaboration. He noted that DeepSeek-V3 required "only 2.788M H800 GPU hours for its full training," a training cost the committee described as substantially lower than what is typical for comparable U.S. models. The committee framed Nvidia’s role as an "optimized co-design of algorithms, frameworks, and hardware," language that underscores the technical depth of the engagement.
DeepSeek captured global attention last year after producing models that matched high performance using markedly less compute than American equivalents. That efficiency has raised alarms among some U.S. lawmakers, who warn it could represent a route by which China narrows the technology gap without relying solely on the level of compute that U.S. firms typically deploy.
The records seen by the committee span activities in 2024, a timeframe when DeepSeek was treated as a conventional commercial partner by Nvidia. "Nvidia treated DeepSeek accordingly - as a legitimate commercial partner deserving of standard technical support," Moolenaar wrote in his correspondence with the committee.
Nvidia responded to the committee's findings by pushing back on the implication that its assistance translated into a military dependence on American hardware. In a company statement, Nvidia said it would be illogical for the Chinese military to rely on U.S. technology any more than it would be for the American military to depend on Chinese technology.
The matter has also intersected with recent executive-branch decisions on chip exports. The Trump administration approved certain sales of higher-performance H200 chips to China under restrictions, a move that has drawn intense criticism from some lawmakers concerned those chips could enhance Beijing’s military capabilities and reduce the U.S. edge in artificial intelligence.
Market reaction to the committee’s disclosures has been muted so far. Nvidia’s shares were reported to be up 0.3% in after-hours trading on Wednesday following the release of the committee materials.
Key points
- Nvidia engineers provided technical assistance to DeepSeek, focusing on joint optimization of algorithms, software frameworks, and hardware - impacting the semiconductor and AI sectors.
- DeepSeek-V3 achieved full training with 2.788M H800 GPU hours, a level cited by lawmakers as notably lower than comparable U.S. models - relevant to compute cost and AI deployment strategies.
- Recent U.S. approvals for sales of more powerful H200 chips to China have intensified scrutiny and concern among policymakers about military applications and competition in AI technology.
Risks and uncertainties
- Potential circumvention of export controls - scrutiny centers on whether technical assistance helped China close the AI compute gap without restricted hardware, affecting national security and export control policy debates in the defense and technology sectors.
- Regulatory and political consequences for semiconductor companies - increased congressional attention could lead to policy responses that affect market access and supply chains in semiconductors and cloud compute services.
- Policy backlash from chip sales - decisions permitting controlled sales of advanced chips to China carry the risk of fueling criticism that such transfers may strengthen Beijing’s military-related AI capabilities, with knock-on effects for defense contractors and related technology firms.