“According to NVIDIA’s filings, NVIDIA technology development personnel helped DeepSeek achieve significant improvements in model training efficiency through ‘optimal co-design of algorithms, frameworks and hardware.’ This is while DeepSeek’s internal reports boast that ‘DeepSeek-V3 requires only 2.788 million H800 GPU hours for full training’ – less than US developers typically require for similarly configured models As frontier-scale models,” Molner wrote in the letter.
GPU hours are the number of hours an AI chip must run to train an AI model, while front-scale models refer to leading models produced by American companies such as OpenAI or Google’s Alphabet.
The documents in question cover Nvidia’s activities from 2024. Molner noted in his letter that while Nvidia provided help to DeepSeek, there was no public indication that DeepSeek’s technology was being used by the Chinese military. “Nvidia treated DeepSeek accordingly — as a legitimate business partner deserving of standard technical support,” Molnar wrote.
Nvidia said in response to Reuters: “China has more than enough domestic chips for all its military applications, with millions to spare. Just as it would be illogical for the US military to use Chinese technology, it does not make sense for the Chinese military to depend on American technology,” Nvidia said in a statement.