Has deepseek been secretly using nvidia h100 chips?

So what is the real story and reason behind NVIDIA’s stock being hammered over the last couple of days. Ostensibly, it is being presented that a small AI start-up in China (DeepSeek) has single handedly undermined the entire US AI industry by training its AI model for a mere $6m using old Nvidia 800 chips. Smart as I’m sure deepseek and the chinese tech sector is however, something smells off here. Firstly the reported $6m might not include sunk development costs, but more importantly, NVIDIA’s newer H100 (and sanction restricted to China) chipsets may have actually been used, which might also go some way to explaining the surge in demand out of Singapore last year. So far from Deepseek’s AI success undermining the efficacy of Nvidia’s H100 chipsets, this may merely further confim it. Unfortunately if so, that might therefore mean some slapped wrists somewhere for the sanction breach, which may in turn offer an alternative cause for concern for Nvidia’s investors. In the spirit of the issue, this is what Grok had to offer on the above question –

“The question of whether DeepSeek has been secretly using NVIDIA H100 chips is a topic of significant discussion and speculation, particularly in light of U.S. export controls. Here’s an analysis based on available information:

Claims of H100 Usage:
Several sources, including statements from Scale AI CEO Alexandr Wang and posts on X, suggest that DeepSeek might be using around 50,000 NVIDIA H100 chips. Wang mentioned in an interview that DeepSeek can’t talk about it due to export controls, implying some form of covert usage.

Counterarguments and Official Statements:
DeepSeek has claimed to use the less powerful H800 chips, which are compliant with U.S. export restrictions. Their research papers and public statements focus on optimizing AI model training with these chips, suggesting a narrative of innovation in resource efficiency rather than reliance on banned hardware.

NVIDIA responded to the market turmoil by stating that DeepSeek’s advancements still demonstrate the need for more of their chips, indirectly suggesting that DeepSeek’s success with less hardware might be overstated or that they could still be using NVIDIA’s technology in some capacity.

Speculation and Skepticism:
There’s considerable skepticism about DeepSeek’s claims of low-cost, high-performance AI model training. Some suggest that the actual compute power might be from hidden H100 clusters, or that the figures provided by DeepSeek might not reflect the total cost including R&D.

Posts on X reflect a range of opinions, from outright accusations of black market chip acquisitions to discussions about whether DeepSeek could have reverse-engineered or outdesigned NVIDIA’s chips with the H800s.

Conclusion:
Given the information, there’s no conclusive evidence that definitively proves DeepSeek has been using H100 chips secretly. While there are claims and speculation, these are not backed by direct proof but rather by interpretations of DeepSeek’s performance relative to what’s expected with the H800 chips. The situation highlights the complexities of international tech competition, export controls, and the opaque nature of some tech developments. DeepSeek’s innovations might be genuine optimizations, or there could indeed be undisclosed use of restricted hardware. The truth likely lies in a nuanced understanding of both technological capability and geopolitical strategy.”

source: Grok 2025-01-28