The trillion-dollar mystery surrounding DeepSeek’s Nvidia GPUs

At the center of the story of DeepSeek’s breakthrough achievement with its R1 models lies the Nvidia hardware that powered the servers that trained those models.

In December 2024, DeepSeek researchers released a paper that outlined the development and capabilities of the new DeepSeek-V3 large language model. In the paper, the researchers said they were able to train their powerful, efficient model over 2.78 million GPU hours of computing time on a cluster of only 2,048 Nvidia H800 GPUs. That is a very small number of GPUs for a model that matched or beat OpenAI’s state-of-the-art o1 model in some benchmarks.

For comparison, Meta trained its Llama 3.1 models on two clusters, using a total of 39.3 million GPU hours with 49,152 Nvidia H100 GPUs. Last week, Mark Zuckerberg said that Meta is planning on ending 2025 with over 1.3 million GPUs.

Released in 2023, the H800 is a GPU that’s similar to the H100 but is tailored for the Chinese market to comply with US export controls concerning national security parameters that the Biden administration rolled out in 2022. Reuters reported that the main thing Nvidia changed in the H800 was that it “reduced the chip-to-chip data transfer rate to about half the rate.”

But The Wall Street Journal reports that government officials found the H800 exploited technical loopholes that met the strict requirements of the ban, but still gave Chinese buyers very powerful AI chips. To close the loophole, in October 2023, the US government banned the export of H800s as well.

It appears that DeepSeek was able to acquire its H800s during that short window of availability.

DeepSeek’s claims are drawing suspicion from some observers in the AI industry, but most appear to be just speculation. Scale AI CEO Alexandr Wang told CNBC that he suspected DeepSeek has “about 50,000 H100s, which they can’t talk about obviously because it is against the export controls that the United States has put in place,” and in a tweet, Elon Musk replied, “Obviously.” Musk, meanwhile, has bragged about xAI’s “Colossus supercluster,” which is powered by 100,000 H100 GPUs, and that he plans to scale up to 1 million of the expensive Nvidia chips.

There have been reports of H100s being smuggled into China through a series of intermediaries on the black market, but no evidence that DeepSeek did so.

Adding to the confusion, DeepSeek cofounder Liang Wenfeng said that the company does own a cluster of 10,000 Nvidia A100 GPUs, a cheaper and less powerful AI chip.

The H100 has earned a status of being one of the most coveted pieces of computer hardware in the AI age. Even when other chips are used, the power is sometimes expressed as a number of “H100-equivalent” GPUs.

Nvidia is in the process of rolling out its next-gen H200 Blackwell GPUs, and last year CEO Jensen Huang hand-delivered the first DGX H200 server to OpenAI headquarters.