U.S. chipmaker Nvidia provided technical guidance that helped China-based DeepSeek make its artificial intelligence models far more training-efficient, according to a letter from the chair of a House of Representatives committee that reviewed internal company records. The committee chairman said those improved models were subsequently used by elements of China's military.
Representative John Moolenaar, a Republican from Michigan who chairs the House Select Committee on China, set out the claims in a letter to U.S. Commerce Secretary Howard Lutnick. Moolenaar wrote that documents obtained from Nvidia showed company technology development staff worked with DeepSeek on an "optimized co-design of algorithms, frameworks, and hardware" that resulted in major gains in training efficiency for DeepSeek's models.
In the letter, Moolenaar cited an internal Nvidia figure showing that "DeepSeek-V3 requires only 2.788M H800 GPU hours for its full training" - a level he said is below the GPU-hour consumption typically associated with frontier-scale models developed by leading U.S. firms. GPU hours measure how long AI chips run while training a model, and frontier-scale models refer to the most advanced systems built by large U.S. developers.
The committee's documents cover Nvidia's activities in 2024. Moolenaar noted that when Nvidia provided the assistance, it was not publicly evident that DeepSeek's technology was being employed by the Chinese military. "Nvidia treated DeepSeek accordingly - as a legitimate commercial partner deserving of standard technical support," he wrote.
Nvidia's H800 chip, which the company designed specifically for the China market, was sold in China before H800 chips were placed under U.S. export controls in 2023. The committee letter links the technical help to the H800 hardware that DeepSeek used in training its models.
The Housing committee chair's account comes amid U.S. official concerns that DeepSeek has supported Chinese military AI capabilities. U.S. officials have said they believe DeepSeek is aiding China’s military.
Nvidia responded in a statement emphasizing that China has vast domestic semiconductor capacity. "China has more than enough domestic chips for all of its military applications, with millions to spare. Just like it would be nonsensical for the American military to use Chinese technology, it makes no sense for the Chinese military to depend on American technology," the company said.
The Commerce Department and the Chinese embassy in Washington did not immediately respond to requests for comment. DeepSeek was not available to comment outside of business hours in China.
Separately, earlier this month the U.S. administration approved sales of Nvidia's more powerful H200 chips to China, subject to restrictions that bar sales to entities that assist the Chinese military. The H200 is a step up in capability from the H800 chips DeepSeek used. That approval has been criticized by policymakers who view it as a potential risk to U.S. advantages in AI and national security.
Regulatory and oversight concerns raised in letter
Moolenaar's letter argued that assurances from sellers about the end use of chips are insufficient without robust licensing controls and enforcement. He wrote: "If even the world’s most valuable company cannot rule out the military use of its products when sold to (Chinese) entities, rigorous licensing restrictions and enforcement are essential to prevent such assurances from becoming superficial formalities."
He added that chip sales to ostensibly non-military end users in China "will inevitably result in a violation of the military end use restrictions," expressing a view that sales to civilian companies can still lead to military adoption.
Implications for markets and technology sectors
- Semiconductor manufacturers face increased scrutiny over how their products and technical support are used abroad.
- AI developers and cloud providers could see heightened regulatory measures and export licensing requirements.
- Defense and national security policymakers may press for stricter enforcement of end-use restrictions and export controls.
The committee's findings, based on internal company records, have amplified debates in Washington about the balance between commercial engagement and national security when advanced computing technology is sold or supported overseas.
Questions now center on whether current export controls, corporate due diligence, and licensing practices are adequate to prevent unintended military use when advanced chips and technical expertise cross borders. The records cited by the committee show that targeted technical collaboration can materially reduce the computational resources needed to train large AI models - a consideration that intersects commercial competitiveness and strategic risk.