Nvidia says DeepSeek’s advances prove need for more of its chips

Nvidia says DeepSeek’s advances prove need for more of its chips

DeepSeek’s user surge highlights rising demand, addressing investor concerns following a selloff of Nvidia stocks.

Nvidia
A DeepSeek research paper revealed it had utilised around 2,000 of Nvidia’s H800 chips, designed to comply with US export controls. (Nvidia pic)
SAN FRANCISCO:
Nvidia on Monday said Chinese AI firm DeepSeek’s advances show the usefulness of its chips for the Chinese market and that more of its chips will be needed in the future to meet demand for DeepSeek’s services.

Nvidia issued a statement on Monday after its shares tumbled 17%to US$118.58 on investor concerns that DeepSeek had matched rivals such as OpenAI using far fewer Nvidia chips than US firms. Nvidia rival Advanced Micro Devices’ shares also slid more than 6% to US$115.01.

“DeepSeek’s work illustrates how new models can be created using that technique, leveraging widely-available models and compute that is fully export control compliant,” Nvidia said in its statement.

One of DeepSeek’s research papers showed that it had used about 2,000 of Nvidia’s H800 chips, which were designed to comply with US export controls released in 2022, rules that experts told Reuters would barely slow China’s AI progress.

The US microchip export controls were designed to freeze China’s development of supercomputers used to develop nuclear weapons and artificial intelligence systems.

Jimmy Goodrich, a senior adviser to the RAND Corp for technology analysis, said there are at least a dozen major supercomputers in China with significant numbers of Nvidia chips that were legal for purchase at the time that DeepSeek used them to learn how to become more efficient. Computing efficiency has also been a major focus of US AI firms.

“DeepSeek didn’t come out of nowhere – they’ve been at model-building for years. It’s been long known that DeepSeek has a really good team, and if they had access to even more compute, God knows how capable they would be,” Goodrich said.

DeepSeek was struggling on Monday to accommodate an influx of new users. Servicing new users is a process that AI firms call “inference”, which Nvidia said demonstrated that its chips will remain in demand.

“Inference requires significant numbers of Nvidia GPUs and high-performance networking,” Nvidia said in its statement.

Nvidia is currently selling a chip called the H20 that is designed to meet the most recent export control regulations. While the restrictions limit the chip’s usefulness for AI training, Goodrich said it is “probably the best chip in the world for inference.”

“How long will Washington allow the best inference chip in the world to be sold to China?” he added.

Stay current - Follow FMT on WhatsApp, Google news and Telegram

Subscribe to our newsletter and get news delivered to your mailbox.