The trigger and the panic
DeepSeek released R1 on January 20, 2025. Its free chat app climbed to the top of Apple's U.S. App Store within the week, ahead of ChatGPT, and the company had already told the press that training its V3 base model, the one R1 was built on, cost about $5.6 million. To investors watching a Chinese lab match Western frontier reasoning models at what looked like a fraction of the price, the combination read as evidence that the AI buildout underway across U.S. tech had been sized for hardware nobody needed to buy. [2]
The reaction arrived on January 27. Nvidia fell about 17% and lost roughly $589 billion in market capitalization, the largest single-day dollar loss for any company in U.S. stock market history. The damage reached well past chipmakers: Constellation Energy, the utility that had agreed to restart a reactor at Three Mile Island to power a Microsoft data center, lost close to 21% of its value, and Vertiv, which sells the power and cooling systems data centers run on, fell sharply alongside it. Each of these companies had been priced, in a different way, for a buildout that suddenly looked larger than it needed to be. [1][3]
The theory, and the objection raised the same day
Atlas interpretation: The sell-off priced in arithmetic, not sentiment. If a model with R1's benchmark results could be trained and served on a fraction of the GPUs the leading U.S. labs were assumed to need, then projected demand for Nvidia's chips, and for the power plants and cooling systems being financed around them, had been set too high. That was the entire chain of reasoning behind a day that erased $589 billion: not that DeepSeek would take customers from Nvidia, but that fewer total chips would need to be bought industry-wide to serve the same amount of AI usage. [1]
Not every technologist agreed that cheaper compute meant less compute purchased in aggregate. On the same day, Microsoft CEO Satya Nadella posted on X and LinkedIn: “Jevons paradox strikes again!”, linking to the nineteenth-century economic observation that making a resource more efficient to use tends to increase total consumption of it, not reduce it, because a lower cost opens up uses that were not worth the expense before. Nadella's argument was that more efficient AI models would mean AI got used for more things, which would mean more chips bought to run it, not fewer. [4]
The cost claim under scrutiny
The $5.6 million figure had a specific, narrow meaning that got flattened in the coverage. DeepSeek's own technical report put the number at $5.576 million and defined it as the rental-equivalent cost, at $2 per H800 GPU hour, of the 2.788 million GPU hours used for V3's official training run: 2.664 million hours for pre-training, 119,000 for context-length extension, and 5,000 for post-training. The company was explicit that this excluded prior research, failed experiments, and the ablation studies that did not make it into the final run. [5]
SemiAnalysis, an industry research firm, disputed the figure's use as a stand-in for what DeepSeek had spent to build its AI capability at all. The firm reported that DeepSeek's parent company had put more than $500 million into Nvidia GPUs over its history, and called the training-run number accurate but incomplete: it captured one run's compute rental price, not the cluster, staff, and repeated experimentation that made the run possible. DeepSeek did not publish a rebuttal figure of its own. The dispute was over what counted as the cost of the model, not over whether the reported run happened as described. [6]
How the theory aged
Nvidia did not get cheaper as a stock. It crossed $4 trillion in market capitalization on July 9, 2025, and less than four months later crossed $5 trillion on October 29, 2025, becoming the first company to reach either threshold. Coverage of the October milestone cited CEO Jensen Huang's own estimate of roughly $500 billion in AI chip orders extending into 2026 and his description of cloud GPU capacity as sold out, the opposite of a market absorbing a demand shock. [7][8]
Atlas interpretation: The demand-collapse theory did not survive contact with the rest of 2025. The big cloud providers did not treat DeepSeek's efficiency claims as a reason to slow their data center buildouts; the buildouts continued and capital spending guidance kept rising through the year instead of falling. Jevons paradox, cited half-seriously on the day of the panic, described what actually happened better than the theory the market was pricing in that afternoon: cheaper inference did not shrink the industry's chip bill, it financed a larger one. The cost dispute over DeepSeek's own number never fully resolved, but it stopped being the question that mattered to Nvidia's valuation. [7][8]
Sources
- Nvidia sheds almost $600 billion in market cap, biggest one-day loss in U.S. history
CNBC · Jan 27, 2025
- China's DeepSeek AI tops App Store: Here's what you should know
CNBC · Jan 27, 2025
- Power stocks plunge as AI energy needs questioned due to new China model
CNBC · Jan 27, 2025
- Microsoft CEO says AI use will 'skyrocket' with more efficiency amid craze over DeepSeek
GeekWire · Jan 27, 2025
- DeepSeek-V3 Technical Report
DeepSeek-AI · Dec 27, 2024
- DeepSeek's hardware spend could be as high as $500 million, new report estimates
CNBC · Jan 31, 2025
- Nvidia briefly touched $4 trillion market cap for first time
CNBC · Jul 9, 2025
- Nvidia becomes first company to reach $5 trillion valuation, fueled by AI boom
CNBC · Oct 29, 2025