The $5.6 Million AI Model That Wiped $589 Billion Off Nvidia in a Single Day

On Monday, January 27, 2025, Nvidia’s stock fell about 17% and, according to Bloomberg, erased $589 billion in market value, the largest one-day loss in stock market history. Reuters, using LSEG data, put the figure at about $593 billion; either way, it was a record. The trigger was not an earnings miss or a scandal. It was a free AI app from a Chinese startup most investors had never heard of.

That startup was DeepSeek, and the story of that day is still one of the best case studies in how AI news, hype and markets collide.

What DeepSeek released

DeepSeek, an AI lab backed by the Chinese quantitative hedge fund High-Flyer, released two models in quick succession:

• DeepSeek-V3, in late December 2024, a large “mixture-of-experts” language model whose technical report claimed performance competitive with leading closed models.
• DeepSeek-R1, on January 20, 2025, a reasoning model that “thinks” step by step before answering. Its paper reported results comparable to OpenAI’s o1 on math, coding and reasoning benchmarks.

Crucially, DeepSeek published the model weights under the permissive MIT license, so anyone could download, run and modify R1. Its consumer app shot to the top of Apple’s U.S. App Store that weekend.

The $5.6 million number, explained

The figure that spooked investors came from the V3 technical report. DeepSeek said the full training of V3 took about 2.788 million GPU hours on Nvidia H800 chips, and that at an assumed rental price of $2 per GPU hour, that came to about $5.576 million.

Compared with the hundreds of millions or more that U.S. labs were believed to spend on frontier models, it looked like DeepSeek had done the same thing for a rounding error. If world-class AI was that cheap, why would anyone need so many of Nvidia’s most expensive chips?

But read the report’s own fine print. DeepSeek stated explicitly that the figure covers only the official training run of V3, and excludes the costs of prior research and ablation experiments on architectures, algorithms and data. It also does not include the cost of buying and running the cluster, staff salaries, or the work that went into R1’s reasoning training. The number was real, but it was the cost of the final lap, not the whole race.

Why it was still a genuine breakthrough

Dismissing DeepSeek as a misleading headline would be a mistake too. The work contained real engineering advances that the wider industry studied closely:

• Efficient architecture. Its mixture-of-experts design activates only a fraction of the model’s parameters for each token, cutting compute per answer.
• Low-precision training. The V3 report described training with FP8, a compact number format that saves memory and compute.
• Reasoning via reinforcement learning. The R1 paper showed that large-scale reinforcement learning could teach a model to reason, including a version, R1-Zero, trained without the usual step of supervised fine-tuning first.
• Working within export limits. The H800 is a version of Nvidia’s hardware made for the Chinese market under U.S. export restrictions. DeepSeek’s results showed constraints can drive efficiency.

What the market got wrong

The sell-off rested on a simple story: cheaper AI means fewer chips. In the following months, Nvidia’s stock recovered and the largest tech companies kept raising their AI infrastructure spending.

Many economists and tech leaders pointed to an old idea called the Jevons paradox: when a resource becomes more efficient to use, total consumption often rises rather than falls, because new uses become affordable. Cheaper reasoning models make it practical to put AI into far more products, run agents that think longer, and serve many more users. All of that still runs on chips.

Reasoning models also shifted where compute goes. Instead of spending most of it once during training, they spend a lot every time they answer, by generating long chains of reasoning. That kind of demand grows with usage.

The lasting lessons

1. Training-cost headlines need context. Always ask what a number includes. “Final training run” and “total cost to build” are very different things.
2. Open weights travel fast. Because R1 was downloadable, developers worldwide could test and build on it within days, which amplified its impact far beyond the app.
3. The frontier is global. DeepSeek proved a lab outside the U.S. giants could reach near-frontier performance, and that efficiency can be a competitive weapon.
4. Markets overreact to narratives. A single day’s price move told a dramatic story. The months after told a more complicated one.

The bottom line

DeepSeek did not make AI chips obsolete, and it did not build a frontier model for the price of a house. What it did was show that clever engineering can close the gap with far fewer resources, and publish the results for everyone. That shook up assumptions about cost, about open models and about who can compete. The $589 billion day was the market’s first, panicked attempt to price that in.

Sources

• Bloomberg, “Nvidia’s $589 Billion DeepSeek Rout Is Largest in Market History”, Jan. 27, 2025
• Reuters, “DeepSeek sparks AI stock selloff; Nvidia posts record market-cap loss”, Jan. 27, 2025
• DeepSeek-AI, “DeepSeek-V3 Technical Report” (arXiv:2412.19437)
• DeepSeek-AI, “DeepSeek-R1: Incentivizing Reasoning Capability in LLMs via Reinforcement Learning” (arXiv:2501.12948)
• DeepSeek-R1 model repository on GitHub