DeepSeek’s sudden rise put a hard question in front of investors: what happens to the AI boom if capable models can be built for far less money than Wall Street expected?
The Chinese company’s claim of a $5.6 million artificial intelligence breakthrough helped wipe almost $600 billion from Nvidia’s market value on Monday. It also shook confidence in a central assumption behind the recent AI trade: that large technology companies would keep spending heavily on advanced chips to stay ahead.
Why DeepSeek hit Nvidia so hard
Nvidia had become the world’s most valuable company last year as investors bet that Big Tech’s demand for powerful AI processors would keep rising. Jensen Huang, Nvidia’s chief executive, has predicted $1 trillion worth of AI data centers will be built in the next few years.
That confidence was tied to the idea of an AI scaling law. Senior leaders at AI start-ups such as OpenAI and Anthropic popularized the view that models become more capable when they are trained with more data and more computing resources.
DeepSeek challenged that narrative with its V3 model and then its R1 model. In December, the company released V3, which it said was comparable to models from OpenAI and Google while being trained on a fraction of the budget, at $5.6 million. DeepSeek said it used just 2,048 Nvidia chips, which could have been obtained without breaching US export controls that have limited China’s access to the newest products from US chipmakers.
Then it unveiled R1, a reasoning model described as comparable to OpenAI’s o1. The company also released a research paper explaining how the model was made, and its chatbot reached the top of the iPhone’s US App Store chart over the weekend.
The market reaction was severe
The sell-off did not stop with Nvidia. The Philadelphia Semiconductor index shed 9.2 percent, its worst daily drop since March 2020. Nvidia’s share price fell 17 percent, creating $6.75 billion in profits for short sellers, according to calculations by data group S3 Partners.
For short sellers who had bet against Nvidia’s high share price, Monday was a major validation. One short seller with interests in several large AI companies argued that open-sourced code from a Chinese entity, arriving before major US tech earnings, sent a message that AI models may be becoming commoditized.
That view lands directly on the largest US technology companies. Stephen Yiu, chief investment officer of Blue Whale Growth, said DeepSeek had leveled the playing field. His fund, backed by billionaire Peter Hargreaves, had reduced exposure to the Magnificent Seven group of big US technology companies last month because of concerns over their large AI spending.
Yiu’s broader point was that AI competition had appeared to require billions of dollars just to enter. If DeepSeek’s approach lowers that barrier, it could be positive for AI adoption, development, and penetration, even if it unsettles the companies that benefited from high entry costs.
Silicon Valley sees another possibility
Not everyone read the sell-off as a lasting blow to Nvidia. Several investors and executives argued that cheaper AI may expand the market rather than shrink it.
Pat Gelsinger, recently forced out as chief executive of Intel, said he was buying Nvidia stock on Monday. In a LinkedIn post, he argued that lower AI costs would expand the market and described DeepSeek as an engineering achievement that would drive wider AI adoption.
Jordan Jacobs, co-founder of AI investor Radical Ventures, also bought more Nvidia shares as the stock fell. He framed the issue around the two main uses for Nvidia chips: training and inference. Training is the process of building a model. Inference is the process of using a finished model to answer user requests.
That distinction matters because Nvidia has said it now earns just as much revenue from inference chips as from training chips. Huang has also argued on a recent podcast that demand for inference is about to go up by a billion times because newer AI models reason, plan, and take more steps before answering complex queries.
Nvidia made a similar argument on Monday. The company said DeepSeek was an excellent AI advancement and a perfect example of Test Time Scaling, referring to systems that use more computing resources after a user asks a question or assigns a task. Nvidia also said inference requires significant numbers of Nvidia GPUs and high-performance networking.
The cost question is still unsettled
DeepSeek’s reported efficiency is the core of the debate, but not everyone accepts the simplest version of the cost story. Dylan Patel of chip consultancy SemiAnalysis has estimated that DeepSeek and its sister company, the hedge fund High-Flyer, have access to tens of thousands of Nvidia GPUs that were used to train R1’s predecessors.
Patel said DeepSeek had spent well over $500 million on GPUs over the history of the company. His view is that the final training run may have been efficient, but that it came after significant experimentation and testing.
G Dan Hutcheson at TechInsights also argued that the market may have focused on the wrong target. He said he did not see DeepSeek as a major hit to Nvidia, but as a bigger challenge for companies such as OpenAI that are trying to sell AI services.
There is also a China-specific chip story. Huawei has emerged as Nvidia’s main competitor in China for inference chips. The Financial Times has previously reported that Huawei has been working with AI companies, including DeepSeek, to adapt models trained on Nvidia GPUs so they can run inference on Ascend chips.
What DeepSeek changes
DeepSeek arrived as the US was moving to assert AI leadership over China and as major US technology companies prepared to report earnings. US President Donald Trump said DeepSeek should be a wake-up call for our industries that we need to be laser focused on competing to win.
For investors, the immediate shock was clear: Nvidia lost almost $600 billion in market value, semiconductor stocks fell sharply, and short sellers gained. But the longer-term meaning is less settled.
If DeepSeek shows that strong models can be built and shared more cheaply, it could pressure companies that sell proprietary AI services. If cheaper models lead to much broader use, it could also increase demand for inference chips and networking. That is why the same event looked like a threat to some investors and a buying opportunity to others.
The central question is no longer only whether AI models can become more capable with more resources. It is also whether more efficient models can make AI common enough that the infrastructure demand keeps growing anyway.