Nvidia’s stock price dropped to $118.50 a share at the end of trading yesterday, January 27—a nearly 17% drop from its opening price of $142.02. This wiped out over $589 billion in market capitalization for the company, which Forbes says is the biggest single-day loss for any company in history. Aside from this, other tech companies also faced losses, with Nasdaq tanking by 3.1%, while other tech giants like Arm, Broadcom, and Oracle all experienced share price drops greater than 10%.
This downward market movement is fueled by the release of DeepSeek R1, an open source LLM that could rival OpenAI’s o1 model in performance while being so optimized that it uses much less power to train. The Chinese AI company said that it took 11x less compute to train its latest model versus Meta’s Llama 3, with its DeepSeek-V3 Mixture-of-Experts (MoE) language model that has 671 billion parameters being trained of 2,048 Nvidia H800 GPUs in two months—approximately 2.8 million GPU hours. By comparison, Meta’s Llama 3, with its 405 billion parameters, took 54 days to train using 16,384 H100 GPUs, or around 30.8 million GPU hours.
Aside from using fewer GPU hours to train, the H800 chips that DeepSeek used to train its LLM have less performance than the H100 due to U.S. export restrictions. This means the Chinese company introduced optimizations that reduced its reliance on advanced hardware to create an AI language that could rival Western AIs.
“DeepSeek is an excellent AI advancement and a perfect example of Test Time Scaling. 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 a statement. “Inference requires significant numbers of NVIDIA GPUs and high-performance networking. We now have three scaling laws: pre-training and post-training, which continue, and new test-time scaling.”
However, because of this realization, many investors are balking at the future cost of AI hardware, especially as companies continue to spend billions on AI-focused infrastructure. Aside from the cost of acquiring so many advanced chips, they also use up a lot of electricity, and many experts are concerned about its impact on the national grid and power supply.
“If it’s true that DeepSeek is the proverbial ‘better mousetrap’ that could disrupt the entire AI narrative that has helped drive the markets over the last two years,” Annex Wealth Management Chief Economist Brian Jacobsen told Reuters. “It could mean less demand for chips, less need for a massive build-out of power production to fuel the models, and less need for large-scale data centers.”
