Reading time: about 8 minutes.
The claim
In late December 2024, the Chinese AI company DeepSeek — at the time known mainly within the AI research community — published a technical report on its new base model, DeepSeek-V3. The report stated that the model’s official training run had used about 2.8 million GPU-hours on Nvidia H800 chips, which it valued at about USD 5.6 million in compute. The H800 is a less powerful version of Nvidia’s chips that could still be exported to China under US export controls. The report added that this figure covered only the official training run, not prior research or experiments.
On 20 January 2025, DeepSeek released DeepSeek-R1, a reasoning-focused model built on V3. Its weights were openly downloadable, and DeepSeek claimed that its reasoning performance was comparable to leading US models. A technical paper followed on the preprint server arXiv on 22 January. In the following days, DeepSeek’s free chatbot app climbed to the top of the US App Store.
The market reaction
On Monday, 27 January 2025, financial markets reacted. Nvidia — the dominant maker of the advanced chips that power most large-scale AI training — saw its share price fall by approximately 17 percent in a single trading session. That drop erased roughly $589 billion of Nvidia’s market value in one day — according to Bloomberg and Forbes, the largest one-day loss of market value by any company in stock-market history, surpassing Nvidia’s own previous record of about $279 billion, set in September 2024. Other chip-related stocks, including Broadcom and Taiwan Semiconductor Manufacturing Company (TSMC), also fell sharply, and the technology-heavy Nasdaq 100 index dropped about 3 percent. Nvidia’s own public statement called R1 “an excellent AI advancement.”
Why investors reacted
Investors read DeepSeek’s claims as evidence that state-of-the-art AI capability might no longer require the enormous compute spending that firms such as OpenAI, Google, and Meta were pouring into their own models — spending that had justified Nvidia’s soaring chip demand and valuation. If a much smaller budget could produce comparable performance, the logic went, then the upstream chip-buying spree might be overbuilt.
What came later
In September 2025, the R1 research was published, after peer review, in the journal Nature. Supplementary material put the cost of R1’s final reinforcement-learning stage at about USD 294,000, run on 512 H800 chips for roughly 80 hours. Many headlines reported that DeepSeek had trained R1 “for $294,000.” Independent reporting pointed out that this figure covered only one stage on top of the V3 base model; adding the base model’s compute brings the total to about USD 5.9 million — still excluding hardware purchases, research, and staff.
Journalists who tested DeepSeek’s chatbot in the days after its release also found that it declined to answer some politically sensitive questions, for example about the events at Tiananmen Square in 1989.
Questions to consider
- Which layer of the AI value chain did DeepSeek’s claim come from, and which layer did it reprice — and why?
- Would the market have reacted the same way if the claim had come from a US start-up?
- R1 was released on 20 January, but markets reacted on 27 January. What does this delay tell you about how AI news moves markets?