On 27 January 2025, Nvidia lost 589 billion dollars of market capitalization in a single day. It is the largest loss of value in the history of the US market, with no comparable precedent. The reason: a Chinese company almost nobody outside specialist circles had heard of —DeepSeek— had published, seven days earlier, an open reasoning model, DeepSeek-R1, that performed on a par with the American models at a fraction of the cost. More precisely: the previous model, DeepSeek-V3, had been trained —according to the figures the company itself published for the final run— for some 5.6 million dollars, against the hundred-plus million GPT-4 cost. The thesis that had governed the sector since 2022 —"without Nvidia GPUs and Western capital you can't train a frontier model"— collapsed in 24 hours. The question is what has changed since then. And, above all, what has happened to Europe.
I'll confess that on 27 January I had bought it too. American supremacy looked like a fact of nature, not a defensible —and therefore attackable— position. I spent months repeating, like almost everyone, that without Nvidia's GPUs you trained nothing serious. I was wrong, and it's no comfort that nearly all of us were. It's worth looking slowly at what DeepSeek did, what it didn't do, and where it leaves us.
What DeepSeek Is
DeepSeek is a company founded in 2023 by Liang Wenfeng, a man in his forties trained at Zhejiang University, former manager of a quantitative fund called High-Flyer. The fund had accumulated, before the American restrictions, a significant quantity of Nvidia GPUs for its algorithmic trading operations. Public estimates of that stock vary wildly —from around 10,000 A100 units to higher speculative figures— so they're best taken as approximations, not as firm data. When Liang saw that the technical frontier of AI was starting to produce sector-wide returns, he decided to divert part of that infrastructure to an independent lab with a clear bet on open weights: models whose parameters are published and anyone can download.
DeepSeek-V2 had come out in mid-2024 without much noise outside technical circles. DeepSeek-V3 came out in late 2024, already carrying training-cost figures that drew attention. But the structural moment was DeepSeek-R1, published on 20 January 2025. R1 is a reasoning model with an explicit chain of thought —similar to OpenAI's o1, published a few months earlier— but with two decisive differences: open weights under an MIT license, and a technical paper describing the full methodology, including a reinforcement-learning approach with outcome verification (instead of human feedback) that opens a new path for scaling reasoning without expensive annotators.
In published benchmarks, R1 sits close to OpenAI's o1 on mathematics, programming and reasoning tasks. The casual user doesn't need to grasp the technical details. What they do need to grasp is the scale of the shift this produced in the sector's expectations.
The 27th of January, Hour by Hour
The sequence of events on Monday 27 January 2025 is worth recalling precisely, because it marks a before and an after.
Over the weekend of the 25th-26th, the technical crowd —Twitter/X, specialist forums, researcher mailing lists— had begun circulating the R1 paper with technical astonishment. The six-million-dollar training figure was discussed with skepticism. Does it hide prior infrastructure already paid for? Is it measuring only the final run and omitting the cost of earlier iterations? The doubts are legitimate and remain open. But even with a wide margin of doubt, the figures were an order of magnitude below those of the American models.
On Monday the 27th, at the open, Nvidia shares opened falling. Over the weekend Marc Andreessen had posted that R1 was "one of the most impressive and amazing breakthroughs" he'd seen. Institutional investors began to sell. The first move was in chips. Nvidia fell 10% in the pre-market. When the regular market opened, the fall accelerated to 17% by close. ASML fell. AMD fell. Broadcom fell. The data-center suppliers —Vertiv, Eaton— fell. The Nasdaq 100 lost 3%. The S&P 500 lost 1.5%. Jensen Huang, Nvidia's CEO, watched his personal wealth shrink by 20.1 billion in a single session.
Nvidia's loss of 589 billion dollars in market cap isn't just a stock-market figure. It's the material translation of a change in the sector's expectations. Investors were pricing in two things. First, that if competitive models can be trained with far fewer chips, future demand for Nvidia GPUs might be lower than the demand that had driven the price into the stratosphere during 2023-2024. Second, and more structural, that the "American monopoly" thesis —only a handful of companies with access to cutting-edge hardware can produce frontier models— had just taken a blow it wasn't going to recover from.
The Cost Paradox and the Technical Debate
The figures deserve qualification, because there's an honest technical discussion open and it can't be treated as settled.
The DeepSeek-V3 paper reported a final-run training cost of some 5.6 million dollars, calculated by multiplying GPU-hours consumed by the theoretical hourly price of an H800. That figure explicitly excludes other associated costs —prior research, failed iterations, prepared data, amortized infrastructure, staff. Legitimate critics have pointed out that the real total cost of the project is probably closer to several hundred million accumulated if you count everything. A project financed, let's remember, from the High-Flyer quant fund, which had prior access to GPUs acquired before the American export restrictions.
I was going to write that the 5.6-million figure is a misleading headline, and it is. But I'll stop myself, because fixating on whether it was six million or six hundred is missing the point. Even granting every qualification, two things remain true. One: the marginal cost of training a competitive model, once you have the methodology and the infrastructure, has fallen dramatically. The techniques DeepSeek demonstrates —a more efficient mixture-of-experts architecture, training on less data but better curated, reinforcement learning with verification— are public domain once the paper is published. Any lab in the world with access to a few thousand GPUs can now replicate the path. The barrier to entry sank.
Two: the specific figures matter less than the political effect. Governments, investors, regulators and competitors acted as if the figures were real. Administrations budgeted accordingly. Investment and policy decisions were made on that basis. The sector's economic reality changed through perception, regardless of what the exact figure was. In the geopolitics of AI, what is believed is, for months or years, what operates. And what has been believed since 27 January 2025 is that China can.
The Sector's Immediate Response
The months that followed were months of accelerated reaction. Meta put out Llama 4 in 2025 with a more aggressive bet on open weights. And OpenAI, which until then had made an irony of its own name, ended up publishing GPT-OSS in August 2025, an open-weights model —something unthinkable in its strategy the year before. The European and Canadian labs, Mistral and Cohere among them, read the same message each in its own way.
And the US Bureau of Industry and Security (BIS), responsible for chip export controls to China, found itself in a difficult internal debate. If the controls had been designed to slow the development of Chinese AI while keeping the technical frontier in the United States, and China had published a competitive model despite the controls, what function were the controls actually serving? A hardline position argues that without the controls China would have arrived sooner. Another, more nuanced, argues that the controls accelerated Chinese innovation in efficiency: forced to do more with less, the Chinese teams developed techniques that then circulated globally and undermined the competitive position of the American frontier itself.
Whatever the reading, the sector changed. Until January 2025 the question was "when will China open the technical gap." After January 2025, the question is "to what extent has China already closed it." Stanford's AI Index 2026 confirms figures: Anthropic's leading model in March 2026 is only 2.7% ahead of the best Chinese model on comparative evaluations. The gap, whatever it was, is all but closed.
Europe Before DeepSeek
Here, as a European, comes the part I find hardest to write without sounding bitter. Europe, before 27 January 2025, had spent two years accepting a secondary role in the sector with a strange mix of fatalism and regulatory pride. The official story was: "we can't compete at the frontier, but we're the ones who regulate best." The AI Act was sold as leadership when structurally it was a defensive response to a sector being built off the continent.
The operational reality: no European company was in the global top tier of frontier models. Mistral, founded in Paris in 2023 by ex-researchers from Meta AI and DeepMind, was the only European lab with international presence, but it competed in a lower division. The big European corporations —SAP, Siemens, the telcos— consumed AI from American providers, they didn't produce it. European hyperscalers were nonexistent: AWS, Azure and Google Cloud cornered the continental cloud market with a combined share that sector estimates put at around 70%. The French government invested in Mistral, the German one talked about digital sovereignty, the European Commission published strategies, but the net flow of capital, talent and compute was outward.
The imagery of submission: Europe as a market where others' products get sold. A buyer of AI, not a producer. The AI Act as a tariff wall dressed up as citizen protection. European industry as an integrator of American technology into traditional sectors —banking, industry, health. That was the position. And it was taken as inevitable.
Europe After DeepSeek
DeepSeek did something the American dynamic had not managed: it woke Europe from its resignation. If China could build a competitive model with fewer resources, the question "why not Europe?" stopped being rhetorical. The structural excuse —"we don't have enough GPUs, we don't have enough capital, we don't have an ecosystem"— became harder to defend. If what mattered wasn't the raw size of the cluster but the efficiency of the method, the continent that knew most about industrial efficiency could, in theory, recover ground.
The weeks that followed were the first in five years in which the European conversation about AI sounded different. Bruegel published an analysis of Europe's "compute gap" with concrete proposals. The Commission began to talk seriously about sovereign compute capacity. The Eurogroup discussed specific funds. And, above all, Mistral found the air to present in Brussels, in April 2026, its document titled European AI: A Playbook to Own It, a roadmap for building AI in Europe on European terms.
In September 2025 Mistral closed a Series C round of 1.7 billion euros (a touch under 2 billion dollars) led by ASML —the Dutch firm that makes the lithography machines used to make frontier chips. The deal has considerable geopolitical elegance: the most relevant European actor at the sector's physical bottleneck comes in as investor of the most relevant European actor at its technical bottleneck. In the ecosystem of mutual dependence, ASML and Mistral hold each other up. It's not total independence, but it's a step.
In November 2025, France and Germany announced, alongside Mistral and SAP, an agreement to build a sovereign AI stack for public administrations, with the binding agreement expected to be signed in mid-2026 and the rollouts staggered over the 2026-2030 period. Public procurement, for years dismissed in discussions of innovation, is starting to become the lever with which European governments sustain capacity of their own. The story changes. It's not frontier leadership, but it's no longer submission.
How much has Europe's projection really changed? Honestly, still little in absolute figures. European investments remain a fraction of American or Chinese ones. Talent still leaves. Compute still concentrates off the continent. But, for the first time in five years, there's vocabulary in circulation that isn't merely defensive. Digital sovereignty, compute gap, open weights as a European strategy, integration with ASML as a continental asset. The conversation has rearmed. Action will lag. But the conversation always changes before action does.
The New Map
What's left after DeepSeek looks more like a map with four blocs than the race between two American companies it was until 2024.
The United States still holds the most capable technical frontier, but its lead is measured in small and shrinking margins. OpenAI, Anthropic, Google DeepMind, xAI, Meta AI. Unlimited capital, concentrated talent, massive infrastructure, but the profitability problem is still open and energy is still an operational constraint.
China has emerged as the second pole with DeepSeek, Alibaba (Qwen), Tencent and other labs. An aggressive open-weights strategy, cross-cutting state funding, a huge internal market that validates products before international deployment. A demonstrated capacity to compete at the frontier.
Europe is starting to sketch a third pole with Mistral, Aleph Alpha, Silo AI and other minor labs. A distinctive bet on sovereignty and on integration into public administrations. Limited capital and talent. Time will tell whether it's a pole or a wish.
And the global south is starting to appear on maps that used to ignore it. The adoption of open Chinese models in India, in Africa, in Latin America, in Southeast Asia produces a polycentric system where the citizen of Jakarta or Lagos doesn't necessarily depend on the American stack. "Inclusivity" as a Chinese banner has its propaganda side, but it also materially produces technical capacity deployed in places where the American offering arrives expensive or doesn't arrive at all.
I write "monopoly" and correct myself on rereading it, because it isn't one anymore. The decisions on investment, regulation, talent, infrastructure, industrial policy are now made from another premise: the AI frontier is contested. What leaves me uneasy isn't that China arrived. It's that it arrived publishing the method. The paper was right there, free, for anyone with a few thousand GPUs and the will to copy the path. And nobody in the West, neither the investors pricing in infinite chip demand nor the European regulators who mistook an AI Act for an industrial strategy, had reckoned with that possibility. The barrier to entry that was taken as physical turned out to be, in good part, a belief. Beliefs collapse in a day.
Definitions
Open weights is the model-publishing strategy in which the trained parameters are distributed openly so anyone can download, run and modify them. It differs from pure open source, because it may not include the training data or the full code.
Reasoning model is a language model optimized to produce explicit chains of thought before giving the final answer. It's more expensive to run, but more capable on mathematical, scientific or logical tasks; DeepSeek-R1 and OpenAI's o1 are the references.
BIS (Bureau of Industry and Security) is the agency of the US Department of Commerce responsible for the export controls on sensitive technology, including frontier AI chips to China since 2022.
Compute gap is the gap in AI-dedicated compute capacity between regions. Europe discusses it as a structural deficit relative to the United States and China, and the discussion has taken on new force after DeepSeek.
References
Bloomberg (January 2025), in its account of the stock-market crash of 27 January, put Nvidia's loss of capitalization at 589 billion dollars. Available at https://www.bloomberg.com/news/articles/2025-01-27/asml-sinks-as-china-ai-startup-triggers-panic-in-tech-stocks
CNBC (27 January 2025) documented Nvidia's 17% fall and the loss of nearly 600 billion in capitalization in one session, used here for the crash figures. https://www.cnbc.com/2025/01/27/nvidia-sheds-almost-600-billion-in-market-cap-biggest-drop-ever.html
Fortune (28 January 2025) reported, based on the Bloomberg Billionaires Index, the roughly 20.1-billion-dollar fall in Jensen Huang's wealth that day, the figure cited in the article. https://fortune.com/2025/01/28/nvidia-founder-jensen-huang-20-billion-blow-net-worth-deepseek-bloodbath/
DeepSeek (January 2025), DeepSeek-R1: Incentivizing Reasoning Capability in LLMs via Reinforcement Learning, the technical paper describing R1's methodology. The training-cost figure of some 5.6 million dollars appears in the technical report of the earlier model, DeepSeek-V3.
Stanford HAI (2026), AI Index Report 2026, source of the figure for the 2.7% lead of the leading US model over the best Chinese model. https://hai.stanford.edu/ai-index/2026-ai-index-report
Mistral AI (2026), European AI: A Playbook to Own It, the roadmap presented in April 2026. https://europe.mistral.ai/
García-Herrero, A. and Martens, B. (2026), Europe needs a strategy to close the artificial intelligence compute gap, analysis published by Bruegel. https://www.bruegel.org/analysis/europe-needs-strategy-close-artificial-intelligence-compute-gap
SAP (November 2025), note on the alliance with Mistral AI and the French and German governments for sovereign AI in public administration, the basis for the agreement's figures. https://news.sap.com/2025/11/sap-mistral-ai-new-alliance-european-sovereign-ai/
Lehdonvirta, V. (2022), Cloud Empires, MIT Press.
Lee, K.-F. (2018), AI Superpowers, Houghton Mifflin Harcourt.
Allison, G. (2017), Destined for War, Houghton Mifflin Harcourt.
También te interesa
- Who OpenAI Really Is
- Anthropic and OpenAI, Two Ways of Understanding AI
- Gemini, Claude and ChatGPT. Do You Really Have to Choose?
No comments yet
No comments yet. Be the first.
Leave a comment