AI-bubble burst: the cost of NVIDIA collapsed by $ 600 billion per day due to the Chinese startup Deepseek

Monday, January 27, 2025, will go down in history as one of the worst days for technological companies from around the world-the promotions of most of them fell against the background of the success of Chinese II-Startap Deepseek. NVIDIA had the worst of all – its capitalization collapsed by about $ 600 billion, which is the largest collapse in the history of the US stock market. And perhaps this is not the end.

Image source: nasdaq

At the time of preparation of this material, NVIDIA shares showed a daily fall of 17.8 % – for this company this is the most serious decline since March 2020. The market value of the largest producer of AI-chips in the world decreased by $ 600 billion to a mark of 2.89 trillion, which is a record fall in history. The previous record of 279 billion also belonged to NVIDIA and occurred in September 2024. Yesterday, NVIDIA was the most expensive company in the world, and today it rolled in third place after Apple and Microsoft, and the risk of falling even lower.

The example of NVIDIA securities was followed by shares of many other companies in the technological sector, one way or another related to AI. Broadcom shares lost 17.3 %, AMD – 8 %, Microsoft – 3 %, Palantir – 7 %. Perhaps Openai was lucky that it is not trading on the exchange, since its actions would most likely also be in the leaders of the fall. The NASDAQ Composite index lost 3.5 %, and the S&P 500 index fell 1.8 %. We add that companies that are not directly related to AI have suffered: for example, Constellation Energy and Vistra electricity suppliers lost 21 and 29 % of their cost, respectively. In total, the US stock market has lost more than 1 trillion of dollars of capitalization per day.

Image source: NVIDIA

Why did investors arrange a sale of shares and collapsed markets? It’s all about the Chinese startup Deepseek, which found a way to train advanced and models on a small number of calculation accelerators. For example, when learning one of its models on an impressive 685 billion parameters, Deepseek used only 2048 and accelerators of the NVIDIA H800 and spent $ 5.6 million. This is a scanty part of the expenses of Openai and Google to teach models of a comparable size.

In addition, last week DeepSeek released the “reasoning” model of AI R1, which surpassed the thinking Openai O1 in important tests. Moreover, the company published instructions on how to build a large language model with minimal costs that can independently learn and improve without control from a person. We add that DeepSeek spreads many of his developments completely free. Yes, and paid access to its most advanced systems is much cheaper than competitors – for example, R1 is available through a company API at a price, which is 90–95 % lower than that of Openai O1.

As a result, investors realized that neural networks can be taught not only according to the “buy and set as many calculation accelerators”, but also much more effective for the smaller number of the GPU. This threatens with a sharp drop in demand for NVIDIA products, the revenue of which more than 80 % depends on computing accelerators.

A number of analysts predicts NVIDIA a gloomy future and do not recommend buying shares of the company so far, although the prices for them have become tempting. Others, on the contrary, are sure that the company will be able to quickly recover, and the current decline should just be used to buy shares.

Image source: NVIDIA

Interestingly, Nvidia herself praised DeepSeek’s development. The company noted that the new DeepSeek R1 model is an “excellent AI achievement”, which does not violate US export restrictions. The statement also rejects the suspicions of some analysts and experts that the Chinese startup could not make the breakthrough that he announced.

At the same time, NVIDIA noted that its accelerators are needed not only for the training of AI models, but also for an infront-launching already trained systems. Moreover, for this you need a lot of GPU, especially with a large number of users. “The information requires a significant number of NVIDIA graphic processors and high -performance networks,” the company said.