Why did DeepSeek's app top the charts and send Nvidia's shares tumbling?

Yesterday, Monday, January 27, 2025, a small Chinese AI company became the main story in technology and finance at the same time. Over the weekend, DeepSeek's chat app climbed past ChatGPT to become the most-downloaded free app in the United States on Apple's App Store. When markets opened, investors sold shares of the companies that supply AI hardware, led by Nvidia, the chipmaker whose processors power most AI data centers.

The two events are connected by one idea: DeepSeek appears to have built top-tier AI models with far less money and computing power than people assumed was needed. This article explains what happened, why that idea moved markets, and what it means for anyone who uses AI.

What happened, hour by hour

DeepSeek is based in China and was founded in 2023 by Liang Wenfeng, a co-founder of the hedge fund High-Flyer, which CNBC reports still owns and funds it. In late December it released DeepSeek-V3, an open model it said took about two months and under $6 million to train. Earlier this month it released R1, a reasoning model that rivals OpenAI's o1. Both are free to download, and its chat app is free to use.

Here is Monday as the sources describe it:

Market value, or market capitalization, is the share price multiplied by the number of shares: a rough measure of what investors think a company is worth on a given day. A drop in market value is not money spent or lost by Nvidia itself. It reflects a change in what buyers are willing to pay for its shares.

DeepSeek itself felt the strain of sudden attention. Later on Monday it said it was temporarily limiting new registrations because of malicious attacks on its services, while existing users could still sign in.

Why cheaper training worried investors

To see why a free app could move a stock worth trillions, start with how AI models are built. Training a model means running enormous amounts of text through a neural network on thousands of specialized chips, often for weeks or months. Nvidia's graphics processors, or GPUs, dominate that market, and companies such as Alphabet, Meta and Amazon spend billions of dollars on them.

DeepSeek's technical report gives its arithmetic for V3, and you can check it yourself:

1. Total training time: 2.788 million GPU hours (one chip running for one hour counts as one GPU hour).
2. Assumed rental price: $2 per hour for an Nvidia H800, a chip with reduced capabilities that DeepSeek could buy despite US export limits.
3. Cost: 2.788 million × $2 = about $5.58 million.

That figure covers only the final training run, not earlier research or the hardware itself, and estimates of what R1 cost differ. Still, it is far below what investors associated with frontier models. If good models can be trained with fewer chips, some investors reasoned, companies might buy fewer of Nvidia's.

The other side of the argument

Not everyone read the news as bad for chipmakers. Analysts at Cantor wrote that more efficient AI will most likely make the industry want more computing, not less. Microsoft's CEO, Satya Nadella, pointed to the Jevons paradox, an economic idea: when a resource becomes cheaper to use, people often end up using much more of it. Cheaper AI could mean AI in many more products, each needing chips to run.

Yann LeCun, Meta's chief AI scientist, framed the story as a win for open research. DeepSeek built on published work, including Meta's Llama models, and by publishing its own methods it lets everyone learn from them.

There are also the questions CNBC lists from analysts: whether US export limits on chips are working, whether open models will undercut paid ones, and whether the huge spending plans announced this month, such as the Stargate project, still make sense. None has a settled answer today. This article describes the reported market moves only; it is not investment advice.

What does this week's DeepSeek moment mean for people who use AI?

For most users, the practical change is more choice. A capable assistant is now free from a company outside the usual group of American labs, and its models are open for anyone to download. That kind of competition tends to push prices down and features up across the industry.

There are trade-offs to weigh. DeepSeek is limiting sign-ups for now, so new users may have to wait. Like other Chinese chatbots, it reportedly steers away from some political topics. And as with any online assistant, think about what you type: avoid sharing personal, financial or work-confidential information unless you have read and accept the service's privacy terms.

Finally, treat the cost story with care. The $5.6 million figure is real arithmetic from DeepSeek's report, but it measures one training run under stated assumptions. The bigger lesson of the day is that efficiency now matters as much as scale in AI, and that the field can change direction quickly.