AI has fueled Nvidia’s extraordinary rise to a $3 trillion market valuation. But on Monday, AI was the cause of a panic among Nvidia investors, sending its shares down almost 17% and wiping out nearly $600 billion in value.
The selloff was triggered by Chinese AI startup DeepSeek, whose latest V3 and R1 AI models appear to rival the best of any U.S. company, while having been trained at a fraction of the cost. Since Nvidia’s powerful graphics processing units are one of the biggest costs of developing the most advanced AI models, investors are suddenly, and radically, questioning their assumptions about the AI business.
While there still many unanswered questions about how DeepSeek developed its models, the upstart is clearly shaking up the AI market. The prognostications of Nvidia’s doom may be premature however. So too may be claims DeepSeek’s success mean the U.S. should abandon policies aimed at curtailing China’s access to the most advanced computer chips used in AI.
DeepSeek has said it has access to 10,000 of Nvidia’s older generation A100 GPUs—chips that were obtained before the U.S. imposed export controls that restricted the ability of Chinese firms to buy these top-of-the-line chips. It has also mentioned training V3 on Nvidia’s H800 chips, a chip Nvidia sells in China that is designed specifically to comply with U.S. export controls.
Either way, that is orders of magnitude less processing power than what U.S. companies typically use to train their most advanced AI models. For instance, Elon Musk’s Xai built a computing cluster, called Colossus, in Tennessee, that has 100,000 of Nvidia’s more advanced H100 GPUs.
What’s more, DeepSeek’s R1 model, an AI model it created to do well on math, logic problems, and coding, and which is designed to challenge OpenAI’s o1 “reasoning” model, is small enough to run on a laptop, where the primary processing power comes from a conventional central processing unit (CPU), rather than requiring access to many GPUs running in a datacenter.
It’s not just investors who have seized on this news. Critics of U.S. export controls on advanced computer chips have pointed to DeepSeek’s success as evidence that the trade restrictions aren’t working. Some even argued the export restrictions have backfired—meant to hobble China’s AI companies and keep them from catching up to the U.S., they have instead forced Chinese AI researchers to develop clever ways to make AI models that are much more efficient in their use of computer power.
“China’s achievements in efficiency are no accident. They are a direct response to the escalating export restrictions imposed by the US and its allies,” Angela Zhang, a professor of law at the University of Southern California and author of a book on Chinese tech regulation, wrote in an op-ed in the Financial Times last week. “By limiting China’s access to advanced AI chips, the US has inadvertently spurred its innovation.” AI skeptic Gary Marcus also echoed these arguments on his blog on Sunday.
Both these emerging narratives could prove shortsighted. That’s because DeepSeek’s impact could, counterintuitively, increase demand for advanced AI chips—both Nvidia’s and those being developed by competitors. The reason is partly due to a phenomenon known as the Jevons Paradox.
Named for 19th Century British economist William Stanley Jevons, who noticed that when technological progress made the use of a resource more efficient, overall consumption of that resource tended to increase. This makes sense if the demand for something is relatively elastic—the falling price due to the efficiency improvement creates even greater demand for the product.
Jevons Paradox could well come into play here. One of the things that has slowed AI adoption within big organizations so far has been how expensive these models are to run. This has had made it hard for business to find use cases that can earn a positive return on investment, or ROI. This has been particularly true so far for the new “reasoning” models like OpenAI’s o1. But DeepSeek’s models, especially its o1 competitor R1, are so inexpensive to run that companies can now afford to insert them into many more processes and deploy them for many more use cases. Taken across the economy, this may cause overall demand for computing power to skyrocket, even as each individual computation requires far less power.
Both Microsoft CEO Satya Nadella and former Intel CEO Pat Gelsinger made this point in posts on social media on Monday. Nadella explicitly referenced Jevons Paradox, while Gelsinger said “computing obeys” what he called “the gas law.” “Making it dramatically cheaper will expand the market for it…this will make AI much more broadly deployed,” he wrote. “The markets are getting it wrong.”
Now the question becomes exactly what kind of computer power will be needed. Nvidia’s top-of-the-line GPUs are optimized for training the largest large language models (LLMs), such as OpenAI’s GPT-4 or Anthropic’s Claude 3-Opus. The company has less of an edge when it comes to what AI researchers and developers call inference—that is using a fully trained AI model to perform a task. Here some of Nvidia’s rivals, including Advanced Micro Devices (AMD) and upstarts such as Groq, have claimed they can run AI applications faster and much more efficiently in terms of power consumption than Nvidia’s GPUs. Alphabet’s Google and Amazon’s AWS also build their own AI chips, some of which are optimized for inference.
Television graphics are seen in the window of Nasdaq headquarters in Times Square, as Nasdaq fell nearly 4 percent this morning on January 27, 2025 in New York City. European and Asian stock markets mostly slid Monday and Wall Street was forecast to open sharply lower on talk that a cheaper Chinese generative AI program, DeepSeek, can outperform big-name rivals, notably in the United States. (Photo by Bryan R. SMITH / AFP) (Photo by BRYAN R. SMITH/AFP via Getty Images)
Some of these rivals could indeed begin to erode Nvidia’s dominant market position. (The company currently controls more than 80% of the market for data center-based AI computing; when the cloud providers’ bespoke silicon is excluded, Nvidia’s share may be as high as 98%.) But Nvidia is unlikely to lose this dominance quickly or entirely. Its GPUs can also be used for inference—and its GPU programming software, CUDA, has a large and loyal developer community that is unlikely to defect overnight. If overall demand for AI computer chips increases due to Jevons Paradox, Nvidia’s overall revenues could still continue to climb, even if its marketshare drops, as it will own a smaller percentage of a larger, and growing, pie.
Another reason that demand for advanced computer chips is likely to continue to grow has to do with the way reasoning models like R1 work. Whereas previous kinds of LLMs became more capable if they used more computer power during training, these reasoning models use what is called “test time compute”—they provide better answers the more computing power they use during inference. So while one might be able to run R1 on a laptop and get it to output a good answer to a tough math question after, say, an hour, giving the same model access to GPUs or AI chips in the cloud might allow it to produce the same answer in seconds. For many business applications of AI, latency, or the time it takes a model to produce an output, matters. The less time, generally the better. And to get that time down with reasoning models still requires advanced computing chips.
For these reasons, it probably still makes sense—if the U.S. sees it as a national security priority to make it more difficult for China to compete on AI—to continue to restrict the country’s access to the most cutting-edge computer chips. Miles Brundage, an AI policy expert who recently left OpenAI, made this point on a podcast over the weekend, saying that even if DeepSeek proved that powerful AI models could be built on fewer, less advanced chips, it would still always be an advantage to have access to more advanced chips than not.
“I think everyone would much prefer to have more compute for training, running more experiments, sampling from a model more times, and doing kind of fancy ways of building agents that, you know, correct each other and debate things and vote on the right answer,” Brundage said. “So there are all sorts of ways of turning compute into better performance, and American companies are currently in a better position to do that because of their greater volume and quantity of chips.”
So export controls might still slow China down when it comes to using AI everywhere it would like to—which would give the U.S. an advantage economically and perhaps militarily, when it comes to deploying AI and reaping its benefits.
On top of this, there’s another argument for why this may not be quite as bad news for Nvidia and U.S. national security policy as investors and critics think: it’s entirely possible DeepSeek has been less than truthful about how many top-flight Nvidia chips it has access to and used to train its models.
Many AI researchers doubt DeepSeek’s claims about having trained its V3 model on about 2,000 of Nvidia’s less capable H800 computer chips or that its R1 model was trained on so few chips. Alexandr Wang, the CEO of AI company Scale AI, said in a CNBC interview from Davos last week that he has information that DeepSeek secretly acquired access to a pool of 50,000 Nvidia H100 GPUs (its latest model). It is known that HighFlyer, the hedge fund that owns DeepSeek, had amassed a substantial number of less capable Nvidia GPUs prior to export controls being imposed. If this is true, it is quite possible that Nvidia is in a better position than investor panic would suggest—and that the problem with U.S. export controls is not the policy, but its implementation.
Leave a Reply