Last reviewed
Jun 8, 2026
Sources
18 citations
Review status
Source-backed
Revision
v1 · 1,914 words
Add missing citations, update stale details, or suggest a clearer explanation.
The DeepSeek market crash of January 2025 was a sharp, single-day selloff in United States and global technology stocks on Monday, January 27, 2025, triggered by the rapid rise of the Chinese AI lab DeepSeek and its claim that it had trained frontier-quality models at a small fraction of the cost of leading US systems. Over the preceding weekend, DeepSeek's newly released DeepSeek-R1 reasoning model and its free chatbot app climbed to number one on the US Apple App Store, prompting investors to question whether the enormous capital expenditure underpinning the US AI and chip boom was justified. [1][2]
When US markets opened on January 27, chip and AI-infrastructure shares fell hard. NVIDIA dropped about 17 percent and lost roughly 589 billion to 593 billion dollars in market capitalization in one day, the largest single-day market-value loss for any company in US stock-market history. The technology-heavy Nasdaq Composite fell about 3.1 percent, and analysts estimated that close to 1 trillion dollars was wiped from US technology and AI-related stocks. [1][3][4] The episode became a focal point for debates over the "efficient AI" thesis, US-China AI competition, semiconductor export controls, and whether the spending behind the AI boom had inflated an AI bubble. Most affected stocks, including NVIDIA, recovered much of the lost ground over the following weeks. [5][6]
DeepSeek is a Chinese AI lab founded in July 2023 by Liang Wenfeng, who in 2015 had co-founded the quantitative hedge fund High-Flyer (Hangzhou High-Flyer Quantitative Investment). High-Flyer, which managed several billion dollars in assets and used machine learning for trading, accumulated a large stock of NVIDIA GPUs starting around 2021, providing the compute base from which DeepSeek operated. The lab pursued an open-weights strategy, releasing model weights publicly. [7][8]
On December 26, 2024, DeepSeek released DeepSeek V3, a large mixture-of-experts language model with 671 billion total parameters (about 37 billion active per token). The V3 technical report stated that the main training run used roughly 2.79 million GPU-hours on NVIDIA H800 accelerators, which the company valued at about 5.6 million dollars assuming a rental price of 2 dollars per GPU-hour. DeepSeek was careful to note that this figure covered only the final training run and excluded prior research, ablations, and data costs, but in popular coverage the number was widely repeated as the total cost and contrasted with the more than 100 million dollars reported for OpenAI's GPT-4. [9][10]
On January 20, 2025, DeepSeek released DeepSeek-R1, an open-weights reasoning model trained with large-scale reinforcement learning on top of the V3 base. R1 posted results competitive with OpenAI's o1 on math, coding, and reasoning benchmarks while being released under a permissive license, intensifying interest in low-cost approaches to building strong reasoning systems. Over the weekend of January 25 to 26, DeepSeek's consumer chatbot app surged up the charts, reaching the number one free-app spot on the US Apple App Store and displacing ChatGPT. [2][11]
As the news spread over the weekend and the app hit number one, investors increasingly questioned the thesis that ever-larger spending on chips and data centers was required to advance AI. When markets opened on Monday, January 27, 2025, a broad selloff hit semiconductor and AI-infrastructure names. NVIDIA, the most valuable beneficiary of the AI build-out, was hit hardest. [1][3]
NVIDIA closed down about 17 percent, erasing roughly 589 billion to 593 billion dollars of market value in a single session. CNBC and several outlets reported the loss as about 589 billion dollars, while figures citing Reuters put it near 592.7 billion dollars; the starting figure of about 593 billion is in the same range. By all accounts it was the largest one-day market-capitalization loss for any company in US history, exceeding prior records set by Meta and Amazon in 2022, each of which had lost more than 200 billion dollars in a day. The Nasdaq Composite fell about 3.1 percent and the S&P 500 about 1.5 percent. [1][3][4]
The damage extended well beyond NVIDIA. The table below summarizes representative one-day moves reported for January 27, 2025.
| Company / index | Approx. move (Jan 27, 2025) | Notes |
|---|---|---|
| NVIDIA | down ~17% | ~589B to ~593B dollars lost, record one-day loss [1][3] |
| Broadcom | down ~17% | AI networking and custom-silicon supplier [3] |
| Micron | down ~12% | memory for AI systems [3] |
| AMD | down ~6% | data-center GPUs and CPUs [3] |
| Constellation Energy | down ~20% | nuclear power tied to AI data-center demand [4] |
| Vertiv | down ~30% | data-center cooling and power infrastructure [4] |
| Nasdaq Composite | down ~3.1% | tech-heavy index [4] |
The selloff notably struck AI-power and infrastructure stocks such as Constellation Energy and Vertiv, reflecting fears that more efficient models would soften projected demand for electricity, cooling, and chips. ASML and other European and Asian chip-equipment makers also fell in overseas trading. Estimates of the total value erased from US technology and AI-linked stocks clustered around 1 trillion dollars. [1][4]
Public and industry reactions were swift and mixed. US President Donald Trump, speaking the same week, called DeepSeek's progress a "wake-up call" for American technology companies, while arguing it could ultimately be positive if it pushed US firms to compete and innovate at lower cost. The development landed days after the administration had publicized the Stargate Initiative, a planned multi-hundred-billion-dollar US AI-infrastructure program, sharpening the contrast between large US capital plans and DeepSeek's lean claims. [12]
Venture capitalist Marc Andreessen, co-founder of Andreessen Horowitz, called DeepSeek-R1 "AI's Sputnik moment" in a January 26 post, framing it as a competitive shock comparable to the 1957 Soviet satellite launch. The "Sputnik moment" label was widely repeated, though some commentators argued it overstated the situation and risked fueling an arms-race framing. [13]
NVIDIA itself downplayed the threat. In a statement on January 27, the company called DeepSeek's R1 "an excellent AI advancement and a perfect example of Test Time Scaling," adding that the work was "fully export control compliant" and that "inference requires significant numbers of NVIDIA GPUs and high-performance networking." Several Big Tech leaders struck a similar note. Microsoft chief executive Satya Nadella invoked the Jevons paradox, writing that "as AI gets more efficient and accessible, we will see its use skyrocket, turning it into a commodity we just can't get enough of," implying that cheaper AI would expand, not shrink, overall demand. Microsoft's Nadella, Google's Sundar Pichai, and Apple's Tim Cook subsequently praised DeepSeek's efficiency work on earnings calls. [14][15]
A central question after January 27 was whether the selloff was justified or an overreaction, and whether DeepSeek's headline cost figures told the whole story. Skeptics argued that the roughly 5.6 million dollar number covered only V3's final training run and excluded research, staffing, and the cost of the underlying GPU fleet. Scale AI chief executive Alexandr Wang publicly claimed it was his understanding that DeepSeek had access to a large cluster of more advanced NVIDIA H100 GPUs that it could not openly discuss because of US export controls. [16]
The semiconductor research firm SemiAnalysis estimated that DeepSeek's parent operation controlled on the order of 50,000 Hopper-class GPUs (a mix that it described as including H100, H800, H20, and A100 units) and that total GPU-server capital expenditure was on the order of 1.6 billion dollars, far above the publicized training-run cost. DeepSeek and its defenders maintained that the V3 figure was an accurate, narrowly defined measure of one training run and was never meant to represent total investment. The dispute underscored how export controls, which had pushed China toward export-compliant H800 and H20 chips, shaped both DeepSeek's methods and the credibility debate around its claims. [17][18]
A second strand of the debate, advanced by NVIDIA, Nadella, and several analysts, held that cheaper, more efficient models would increase aggregate AI usage and therefore chip demand over time, consistent with the Jevons paradox. Wall Street research firms such as Bernstein characterized fears of an imminent collapse of AI-infrastructure spending as overblown, helping set the stage for a rebound. [3][14]
The crash became a watershed reference point in the "efficient AI" debate, demonstrating in dramatic financial terms that algorithmic and engineering efficiency, not only raw spending, could move frontier capability and that markets were highly sensitive to that possibility. It sharpened scrutiny of the AI capital-expenditure thesis and fed the broader argument over whether the sector had inflated an AI bubble. [5][6]
Despite the historic one-day loss, the selloff proved largely temporary. NVIDIA and other chip stocks recovered much of the decline within weeks, and NVIDIA had recouped most of the lost value by late February 2025. In a February 2025 interview, NVIDIA chief executive Jensen Huang argued that investors had misread R1, treating it as a sign that AI would need far less compute, when in his view reasoning models that "think" at inference time would drive demand for more computation, not less. [5][6]
The episode also intensified US policy attention on Chinese AI capabilities and export controls. DeepSeek continued to release widely used open models, and the broader competitive dynamic it crystallized, low-cost open models from Chinese labs challenging higher-cost US systems, remained a defining theme of AI in 2025. Later in 2025, DeepSeek published additional cost disclosures (including a peer-reviewed figure of about 294,000 dollars for the R1 reinforcement-learning stage on top of the multimillion-dollar base model), which renewed but did not settle the debate over how to account for the true cost of training frontier models. [10][18]