Nvidia Worked to ‘Co-Design’ DeepSeek Model, US Lawmaker Says

(Bloomberg) -- Nvidia Corp. provided technical support that helped DeepSeek improve its breakthrough artificial intelligence model despite US export controls designed to restrict the Chinese startup’s access to high-end American chips, according to the Republican head of the House China committee.

DeepSeek achieved cutting-edge performance with its R1 model thanks to what Nvidia called “an optimized co-design of algorithms, frameworks and hardware” for using its H800 processors, Representative John Moolenaar wrote in a letter Wednesday to US Commerce Secretary Howard Lutnick. Nvidia also proposed offering DeepSeek as an enterprise-ready product to be deployed on its hardware, Moolenaar wrote, citing records obtained from the chipmaker.

Most Read from Bloomberg

“In effect, Nvidia’s technical support allowed DeepSeek to extract near-frontier performance from ‘deprecated’ H800 chips, undermining the export-control bottlenecks that US policy was designed to impose,” he wrote. Nvidia’s internal reporting shows that DeepSeek-V3 requires only 2.8 million H800 GPU hours for its full training, according to the letter.

Nvidia created the H800 as a hobbled version of its H100 chip in 2023 to comply with existing export control rules, and the processor was allowed to be sold to Chinese customers until October of that year. The letter offers more detail on the extent to which Nvidia was actively working to help DeepSeek design the best model possible in the face of semiconductor constraints.

The records obtained by the committee included communications between Nvidia and DeepSeek from June 2024 to May 2025.

Moolenaar said Nvidia’s collaboration with DeepSeek should spur tough enforcement of US conditions for allowing shipments of the company’s H200s to China following President Donald Trump’s decision in December to ease restrictions on some AI chip sales to the world’s second-largest economy. The Commerce Department has since spelled out terms for winning approval for H200 sales licenses, including a requirement for rigorous procedures to prevent unauthorized use of the technology.

In a statement, an Nvidia spokesperson said the “administration’s critics are unintentionally promoting the interests of foreign competitors — America should always want its industry to compete for vetted and approved commercial businesses, and thereby protecting national security, creating American jobs, and keeping America’s lead in AI.”

Commerce Department spokespeople had no immediate comment.

DeepSeek’s launch of its R1 model last year rattled markets, erasing 3% from the tech-focused Nasdaq 100 Index in one day as investors weighed the impact of an AI model developed at lower cost and with more technical constraints. Shares rebounded, but fears in Washington about how to maintain a US lead in artificial intelligence have lingered.

In his letter, Moolenaar renewed accusations — raised by his panel in April — that DeepSeek has connections to the Chinese Communist Party and Beijing’s military. He said the Chinese military has adopted DeepSeek models inside “military hospitals and defense mobilization planning units,” calling it an added reason for caution in clearing any H200 exports to China.

The Nvidia spokesperson rejected Moolenaar’s concerns that the company’s chips could aid China’s military aims. “China has more than enough domestic chips for all of its military applications, with millions to spare,” the spokesperson said. “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.”

Moolenaar has endorsed House legislation calling for more congressional oversight of AI chip export controls and a potential ban on sales of any processors more advanced than Nvidia’s Hopper generation, which includes the H200. That product line is second-best to the Blackwell chips currently on the market and two generations behind the upcoming Rubin series. An 18-month lag behind the latest Nvidia products was part of the Trump administration’s justification for allowing H200s to be exported to China.

Moolenaar requested a briefing from Lutnick by Feb. 13 on the H200 rule enforcement and another recommendation that the Commerce Department use its authority to consider restricting the use of Chinese AI models in the US.