Six million dollars against five hundred billion

On 20 January 2025, a Chinese artificial intelligence company called DeepSeek released a reasoning model called R1 under the MIT licence, the most permissive of the standard open-source licences, which grants anyone the right to use, modify, and distribute the software for any purpose without restriction.¹

The following day, President Trump announced the Stargate Project, a $500 billion commitment to American AI infrastructure. The two events sat twenty-four hours apart—and the gap between their price tags did more to shape the year that followed than either announcement did on its own.

Within a week the DeepSeek app was the top free download in both the United States and China. On 27 January, NVIDIA lost close to $600 billion of market value in a single day, $589 billion by the closing figure, the largest one-day loss any company has recorded in United States history.

The cause was not the model's capabilities, formidable though those were. It was the cost. DeepSeek disclosed that training the underlying V3 model had taken 2,048 NVIDIA H800 processors and roughly 2.8 million processor-hours, which at a rental price of $2 an hour came to $5.576 million.² The company stated plainly that this covered the official training run alone and excluded the prior research and architectural experiments that preceded it. When the reinforcement learning phase that produced R1 was later costed at $294,000, the total still came in under six million dollars.

Six million dollars against five hundred billion. The disparity broke two narratives that had structured the Western AI industry's understanding of itself. The first was the scaling hypothesis: that competitive AI required ever-larger quantities of computing power, data, and capital, and that whoever held the deepest pockets would inevitably dominate. The second was the effectiveness of export controls: that restricting China's access to advanced chips would constrain its AI capability. DeepSeek had trained a frontier model on processors several generations behind the leading edge, using techniques that favoured efficiency over brute force.

A year on, the questions that followed remain unresolved. What does open-source AI mean when the most permissively licensed frontier model comes from a country whose data governance and censorship regime run against the values the open-source movement has historically claimed? What happens to the capital-intensive Western model when a laboratory most policymakers had never heard of reaches comparable results for a thousandth of the price? And what does "open source" mean at all in AI, when the term is used by companies whose practices bear no resemblance to the definition?

What DeepSeek built

R1 is a reasoning model, designed not merely to generate text but to work through problems step by step, producing chains of reasoning that can be examined and evaluated. The technical approach was distinctive. Rather than the supervised fine-tuning that characterised most frontier models, DeepSeek applied large-scale reinforcement learning directly to its base model, letting the system explore chain-of-thought reasoning without the extensive human-labelled data Western laboratories relied on.³

The benchmark results were striking. On AIME 2024, a mathematics competition benchmark, R1 scored 79.8 per cent. On GPQA Diamond, a graduate-level question-answering benchmark, it scored 71.5 per cent. On MATH-500 it scored 97.3 per cent.⁴ Those results placed it in the same tier as OpenAI's o1, the model built explicitly as a reasoning specialist, and they came from a training run costing a fraction of o1's estimated expense.

The efficiency had several sources. DeepSeek's team developed training techniques that cut the computation each step required. Reinforcement learning is inherently less data-hungry than supervised fine-tuning, needing fewer human annotations. And the team optimised its use of the H800, an export-restricted version of NVIDIA's H100 with reduced interconnect bandwidth, to extract performance the export control regime had been designed to prevent.

The openly documented methodology mattered as much as the model. DeepSeek published detailed technical papers explaining its approach, making the techniques available to anyone with the model and sufficient computing power. That transparency stood against the opacity of Western frontier laboratories, which had published no comparable detail for their reasoning models.

The point was settled more firmly than any technical report could settle it in September 2025, when the R1 work appeared as a cover article in Nature, the first major open-weight language model published after independent peer review. It went through three rounds of review with eight reviewers, producing 64 pages of reports across five months. The reviewed paper is also where the $294,000 figure for the reinforcement learning phase was established, on 512 H800 processors over roughly four days. Whatever else is contested about DeepSeek, its central technical claim has been through a process no Western frontier laboratory has yet submitted to.

What "open" means, and does not

The MIT licence under which R1 was released is a genuine open-source licence. It imposes no restriction on commercial use, no limitation on derivative works, and no threshold based on company size or user count. Anyone can download the model, modify it, deploy it commercially, and distribute their modifications without asking permission or paying royalties.

That placed DeepSeek in sharp contrast with the companies that had marketed themselves most aggressively as champions of open-source AI. Meta's Llama models, which Meta described as open source, were released under a community licence carrying significant restrictions. Entities with more than 700 million monthly active users had to obtain a separate licence from Meta. The licence prohibited using Llama's outputs to train competing models. The Open Source Initiative stated plainly that the licence did not meet the Open Source Definition and asked Meta to stop calling Llama open source.⁵

The gap between DeepSeek's actual openness and Meta's claimed openness pointed at a broader problem. In traditional software the meaning of open source was settled: the Open Source Definition, maintained by the Initiative since 1998, requires that software be freely redistributable, that source code be available, that derived works be permitted, and that no restriction be placed on fields of endeavour or persons. A model released with commercial restrictions and user-count thresholds is not open source by any recognised definition. It is source-available software with a marketing budget.

The Initiative's attempt to adapt the definition for AI produced its own controversy. The Open Source AI Definition, published in October 2024, introduced the concept of "data information" as an alternative to requiring release of the full training dataset, on the grounds that complete publication was impractical given privacy, copyright, and scale.

Bradley Kuhn of the Software Freedom Conservancy objected on the record. His argument was not that the compromise was inelegant but that it was backwards. The Initiative had reasoned, in his account of it, that "there was no point in publishing a definition that no existing AI system could currently meet", engineering the standard to fit the systems that existed rather than setting one those systems should aspire to reach. That inversion, he wrote, undermined the community process and amplified the role of stakeholders who stood to gain financially from a retroactive declaration that their systems were open source. His substantive objection concerns reproducibility: without sufficient requirements on the licensing and public disclosure of training sets, the public cannot reproduce the scientific process by which these systems are built. He also faulted the speed of the process, which he considered open to regulatory capture, against an Open Source Definition that had developed over decades.

The result is a definition that R1 meets, with its MIT licence, documented methodology, and public weights, and that many models marketed as open source by Western companies do not.

The export control question

DeepSeek operates under the jurisdiction of the People's Republic of China. Its models are subject to Chinese data governance and censorship obligations. The company discloses little about its structure, its funding, or its relationship with the Chinese state. Asked about Tiananmen Square, Xinjiang, or Taiwan, its chatbot declines or answers consistently with government positions.

This creates a problem existing governance frameworks are not equipped to resolve. The model itself is genuinely open, and anyone can download, inspect, modify, and deploy it. The organisation that produced it operates inside a governance system that is not open at all. The MIT licence grants freedoms to users of the model. It says nothing about the freedoms of the people who built it, the data governance under which it was trained, or the political constraints shaping what it will discuss.

The conclusion drawn most widely from January 2025 was that chip export controls had failed. The most detailed analysis of that question reaches something more careful. CSIS argues that DeepSeek's success "in large part reflects the lagging impact of the flawed first package of U.S. AI chip export controls in October 2022", which is a judgement about a badly designed early regime rather than about export controls as an instrument. It also quotes DeepSeek's own chief executive, Liang Wenfeng, from a July 2024 interview: "Money has never been the problem for us; bans on shipments of advanced chips are the problem." On the CSIS reading, Chinese firms prioritised efficiency innovation precisely because constraint made it the most attractive research path available to them. That describes export controls working in an unintended direction rather than failing.

Bruegel is blunter about the consequence. The prevailing view had been that building large language models required billions of dollars of high-end chips, confining the field to the largest American firms. DeepSeek overturned it: with the right talent, similar results can be had for far less money. The clear implication, Bruegel writes, is that "US hegemony in AI is no longer guaranteed", and that controls intended to contain China technologically appear ineffective. It expects the most likely American response to be a doubling down on export controls, and notes this is poor news for Europe, which is unlikely to operate comfortably inside either ecosystem.

The capital question

If competitive AI can be built for six million dollars rather than hundreds of millions, the economic structure of the Western AI industry comes into question.

Stargate represented the apotheosis of the capital-intensive approach: $500 billion committed, backed by OpenAI, Oracle, SoftBank, and the Emirati sovereign wealth fund MGX, to build infrastructure at a scale that would, in the project's framing, secure American AI dominance for a generation. The investment thesis rested on the scaling hypothesis. More computing power, more data, and more infrastructure would produce more capable models, and whoever invested most would win.

R1 did not disprove scaling. Larger models trained on more data can still outperform smaller ones in some domains. It demonstrated that the relationship between investment and capability is not linear, and that a team with the right techniques can reach comparable results with a fraction of the resources.

For smaller countries this mattered immediately. Frontier AI stopped being the exclusive preserve of entities able to mobilise billions. Academic institutions, open-source communities, and governments with modest budgets could download, fine-tune, and deploy a frontier model for purposes reflecting their own priorities, whether healthcare in Sub-Saharan Africa, climate modelling in the Nordic countries, or education in Southeast Asia. The MIT licence restricts nobody.

For the industry the question was sharper. If the moat around frontier AI is technique rather than computing power, then the billions spent on data centres, processor clusters, and power infrastructure may represent a sunk cost rather than a competitive advantage. The venture capital model funding the American industry assumed capital-intensive barriers to entry would protect incumbents. R1 suggested those barriers were lower than anyone had assumed.

The open washing problem

DeepSeek's genuine openness threw the practices of Western companies claiming the label into relief.

Meta was the most prominent example rather than the only one. The company invested heavily in positioning Llama as the standard-bearer of open-source AI, describing it that way in press releases, blog posts, and congressional testimony, on terms its own licence did not support. What it distributed was a proprietary model on terms that preserved its strategic advantages while capturing the reputational benefit of openness.

The compromise the Open Source Initiative reached, requiring information about data rather than the data itself, was pragmatic. It also created a category of open-source AI offering freedom to use without freedom to audit. A researcher can download R1 and run it. They can examine the weights, study the architecture, and test for bias or harmful output. What they cannot do, under even the most permissive licence, is fully reproduce the model without the training data that shaped its behaviour.

That limitation applies to every open-weight model regardless of licence. It is particularly consequential for systems trained under data governance regimes that are opaque by design, because a model trained on data curated under censorship requirements will reflect those requirements in ways that may not be detectable from the weights alone.

A year on

The year that followed settled some things and left the important questions open.

The efficiency thesis held, and the industry reorganised around it. Reasoning capability stopped being something locked behind a proprietary interface and became a component teams could download and deploy, which moved the conversation from comparing models to deploying them, cutting their cost, and integrating them into real systems. Hugging Face's assessment of the year records the shift in the adoption data. Among models created in the past year, downloads of Chinese models overtook those of every other country including the United States. Baidu went from no releases on the platform in 2024 to more than a hundred in 2025, and ByteDance and Tencent each increased their releases eight or nine times over. Open release stopped being a short-term tactic and became part of long-term competitive strategy, and Western startups increasingly default to models developed in China.

Which returns the original questions with more force than they carried a year ago.

Can open-source AI governance accommodate models from jurisdictions with fundamentally different values? The MIT licence is jurisdiction-neutral by design, granting identical freedoms regardless of where the model was built. The model is not jurisdiction-neutral. It carries the imprint of the data governance, political constraint, and organisational culture that shaped its training. A governance framework that evaluates the licence and not the provenance will miss the most consequential dimension of what "open" means.

Can the Western capital-intensive model survive competition from efficient alternatives? That depends on whether scaling holds at the frontier, or whether techniques like DeepSeek's close the gap permanently. If efficiency wins, the concentration of AI capability in a handful of well-capitalised companies is contingent rather than inevitable, and the contingency has already been demonstrated.

And who benefits most from genuinely open AI? The optimistic answer is the public: researchers, smaller companies, developing countries, and civil society organisations able to address their own problems without depending on proprietary platforms. The less optimistic answer is the next generation of companies that will build commercial products on open foundations without contributing back, capturing the value of openness without bearing its cost.

DeepSeek did not answer these questions. It made them impossible to avoid. Six million dollars bought a model the world's wealthiest companies had spent billions trying to build, and a year later the implications are still unfolding.

Footnotes

1. DeepSeek, "DeepSeek-R1 Release," 20 January 2025. ↩

2. DeepSeek-V3 technical report, December 2024. The figure covers 2.788 million H800 processor-hours across pre-training, context length extension, and post-training, priced at $2 per processor-hour. The report states explicitly that it excludes the costs of prior research and of ablation experiments on architectures, algorithms, and data. ↩

3. DeepSeek-R1 technical report, January 2025. ↩

4. Ibid. AIME 2024 at 79.8 per cent, GPQA Diamond at 71.5 per cent, MATH-500 at 97.3 per cent. ↩

5. Open Source Initiative, "Meta's Llama Licence Is Still Not Open Source," 2024; Meta, "Llama 3 License," 2024. ↩