
This is our second article about Nvidia and DeepSeek today, where we delved deeper into DeepSeek's claims. Many people may be surprised to know that DeepSeek actually uses Nvidia GPUs. DeepSeek, a Chinese AI firm, has emerged as a major player in artificial intelligence, developing models that rival leading competitors like OpenAI’s GPT-4. However, due to U.S. export restrictions on advanced AI chips, DeepSeek has been forced to rely on Nvidia H800 GPUs instead of the more powerful H100 GPUs. This decision has had significant cost implications, highlighting the financial burden imposed by geopolitical constraints.
The numbers: H800 vs. H100
DeepSeek reportedly used ~2,000 Nvidia H800 GPUs to train its models. If they had access to H100 GPUs, they would have only needed ~715 GPUs due to the H100’s significantly higher performance (2.8x more compute power).
Cost comparison
Using H800 GPUs (actual):
~2,000 units at an estimated $25,000 each
Total cost: ~$50 million
Using H100 GPUs (hypothetical):
~715 units at an estimated $30,000 each
Total cost: ~$21.45 million
The impact of restrictions
Higher cost: DeepSeek was forced to spend $28.5M more than they would have with unrestricted access to H100s.
Inefficiency: More GPUs mean higher power consumption, cooling costs, and operational inefficiencies.
Training speed: H100s would have enabled faster model training, reducing overall project timelines.
Necessity, not strategy: The use of H800s was not a cost-saving measure but rather a forced adaptation to U.S. sanctions.
The contradiction: Nvidia’s most advanced chips allow lower costs
Despite Nvidia's stock decline, the reality is that Nvidia’s high-end chips, like H100s, actually help reduce AI training costs. If DeepSeek had access to them, it could have completed its training at a much lower cost and with fewer GPUs.
H100s are ~2.8x faster than H800s, meaning DeepSeek could have used only ~715 H100s instead of 2,000 H800s to achieve the same AI training results.
Total cost with H100s (~$21.45M) would have been ~57% cheaper than what DeepSeek actually spent (~$50M) using H800s.
Electricity and cooling costs would also have been significantly lower with fewer GPUs.
So why did Nvidia’s stock drop?
Despite the fact that Nvidia’s best chips are still the most cost-effective, investors reacted negatively because DeepSeek’s success suggested AI firms might prioritize efficiency rather than just raw GPU power. This led to concerns that:
Future AI firms might not automatically buy Nvidia’s most expensive GPUs.
Companies may optimize architectures to use fewer chips, reducing demand for high-end Nvidia GPUs.
There could be a long-term shift towards cost-efficiency over maximum performance, leading to lower revenue growth for Nvidia.
The core contradiction
Nvidia’s most advanced chips actually help reduce AI training costs.
DeepSeek spent more because they did not have access to them.
Yet Nvidia’s stock fell because investors saw this as proof that AI firms could survive without the most expensive GPUs.
Nvidia’s perspective
Despite restrictions, Nvidia benefits as Chinese AI firms like DeepSeek continue purchasing large quantities of H800 GPUs, proving that demand remains strong even for downgraded chips. However, the long-term risk is that AI firms may now prioritize efficiency over raw power, potentially disrupting Nvidia’s pricing power.
DeepSeek’s case highlights the hidden costs of export restrictions, not just in hardware procurement but in the overall efficiency of AI development. Ironically, had DeepSeek had access to Nvidia’s best chips, it would have trained its model at a significantly lower cost. However, this situation has sparked broader concerns among investors that AI firms may start looking for alternative, cost-effective solutions instead of Nvidia’s most expensive GPUs. If these limitations persist, Chinese firms may eventually seek homegrown alternatives to reduce dependence on Nvidia and other U.S. suppliers. Meanwhile, Nvidia continues to profit, albeit through lower-performance products. The AI race remains heavily influenced by geopolitics, with technology firms navigating complex trade barriers to stay competitive.
Comments