Open Source

Six Million Dollars: DeepSeek, Open Source, and the Geopolitics of Efficient AI

When a Chinese AI lab released a frontier model under a genuine open-source licence for a fraction of the cost of its Western competitors, it broke two narratives at once — and raised questions that neither side of the Pacific was prepared to answer.

P J Laszkowicz

On January 20, 2026, 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, granting anyone the right to use, modify, and distribute the software for any purpose without restriction.1 Within days, the DeepSeek app was the number-one free download in dozens of countries. Within a week, NVIDIA had lost $600 billion in market value.2

The reason for the market convulsion was not the model's capabilities, though those were formidable — R1 matched or exceeded the performance of OpenAI's o1 model across mathematics, coding, and reasoning benchmarks.3 It was the cost. DeepSeek disclosed that the total training cost for the base model and R1's reinforcement learning phase was approximately $5.9 million — less than six million dollars in compute time, using 2,048 Nvidia H800 GPUs over fifty-five days.4 By comparison, OpenAI's o1 was estimated to have cost over $100 million to train. Google's Gemini and Anthropic's Claude had reportedly consumed comparable resources. The Stargate Project, announced by President Trump three days before DeepSeek's release, had pledged $500 billion in US AI infrastructure investment.5

Six million dollars against five hundred billion. The disparity shattered two narratives that had structured the Western AI industry's self-understanding. The first was the scaling hypothesis — the assumption that competitive AI required ever-larger quantities of compute, data, and capital, and that the companies and countries with the deepest pockets would inevitably dominate. The second was the effectiveness of export controls — the belief that restricting China's access to advanced chips would constrain its AI capabilities. DeepSeek had trained a frontier model on chips that were several generations behind the cutting edge, using techniques that prioritised efficiency over brute force.6

The questions that followed were more uncomfortable than the shattered narratives. What did it mean for open-source AI when the most permissively licensed frontier model came from a country whose data governance and censorship regime were antithetical to the values that the open-source movement had historically championed? What did it mean for the capital-intensive Western AI model when a lab most policymakers had never heard of achieved comparable results for a thousandth of the price? And what did "open source" even mean in the context of AI, when the term was being used by companies whose practices bore no resemblance to the definition?

What DeepSeek built

DeepSeek-R1 was a reasoning model — designed not merely to generate text but to think through problems step by step, producing chains of reasoning that could 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, allowing the system to explore chain-of-thought reasoning for complex problems without the extensive human-labelled training data that Western labs relied upon.7

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. On MATH-500, it scored 97.3 per cent.8 These results placed it in the same performance tier as OpenAI's o1, the model explicitly designed as a reasoning specialist — and they were achieved with a model whose training had cost roughly one-twentieth of o1's estimated expense.

The efficiency came from multiple sources. DeepSeek's team had developed novel training techniques that reduced the compute required for each step of the process. The reinforcement learning approach was inherently less data-hungry than supervised fine-tuning, requiring fewer human annotations and less curated training data. And the team had optimised its use of the H800 chips — export-restricted versions of NVIDIA's H100 with reduced interconnect bandwidth — to extract performance that the export control regime had been designed to prevent.9

The openly documented methodology was itself significant. DeepSeek published detailed technical papers explaining its training approach, making the techniques available for replication by anyone with access to the model and sufficient compute. This transparency stood in contrast to the opacity of Western frontier labs: OpenAI had not published detailed training methodology for its reasoning models, and Anthropic had shifted away from publishing detailed model specifications. The Chinese lab was, in a concrete and measurable sense, more open about its techniques than its American competitors.

What "open" means (and doesn't)

The MIT licence under which DeepSeek released R1 is a genuine open-source licence. It imposes no restrictions on commercial use, no limitations on derivative works, no thresholds based on company size or user count. Anyone can download the model, modify it, deploy it in a commercial product, and distribute their modifications — all without asking permission or paying royalties.10

This placed DeepSeek in sharp contrast with the companies that had most aggressively marketed themselves as champions of "open-source AI." Meta's Llama models, described by Meta as open source, were released under a community licence that imposed significant restrictions. Entities with more than 700 million monthly active users were required to obtain a separate licence from Meta.11 The licence prohibited using Llama's outputs to train competing AI models. Newer versions of the licence had excluded users in the European Union entirely from certain provisions.12 The Open Source Initiative had explicitly stated that Meta's Llama licence did not meet the Open Source Definition, calling on Meta to stop describing Llama as open source.13

The gap between DeepSeek's actual openness and Meta's claimed openness highlighted a broader problem in the AI industry: the term "open source" was being used with a flexibility that would have been unacceptable in any other domain of software development. In traditional software, the meaning of open source was settled: the Open Source Definition, maintained by the OSI since 1998, required that software be freely redistributable, that source code be available, that derived works be permitted, and that no restrictions be placed on fields of endeavour or persons.14 When Meta released a model with commercial restrictions and user-count thresholds, it was not open source by any recognised definition. It was source-available software with a marketing budget.

The OSI's attempt to adapt the Open Source Definition for AI — the Open Source AI Definition, version 1.0, published in October 2024 — had produced its own controversy. The definition introduced the concept of "data information" as an alternative to requiring the release of full training datasets, arguing that complete dataset publication was impractical due to privacy, copyright, and scale constraints.15 Critics, including Bradley Kuhn of the Software Freedom Conservancy, argued that the definition had been "calibrated to what industry could meet, not what openness requires," and that substantial acrimony had marked the drafting process, with working groups that favoured requiring open training data being effectively overruled.16

The result was a definition that DeepSeek's R1 met — MIT licence, openly documented methodology, publicly available model weights — but that many of the models marketed as "open source" by Western companies did not. The irony was difficult to miss: the most genuinely open frontier AI model in the world had come from a country that most Western policymakers associated with opacity and control.

The geopolitical paradox

DeepSeek operated under the jurisdiction of the People's Republic of China. Its models were subject to Chinese data governance requirements and censorship obligations. The company itself disclosed little about its organisational structure, funding sources, or relationship with the Chinese state. When DeepSeek's chatbot was asked about sensitive political topics — Tiananmen Square, Xinjiang, Taiwan — it declined to answer or provided responses consistent with Chinese government positions.17

This created a paradox that existing AI governance frameworks were not equipped to resolve. The model itself was genuinely open: anyone could download, inspect, modify, and deploy it. The organisation that produced it operated within a governance system that was, by Western standards, not open at all. The MIT licence granted freedoms to users of the model. It said nothing about the freedoms of the people who built it, the data governance under which it was trained, or the political constraints that shaped what it would and would not discuss.

Western policymakers who had framed AI competition as a US-China zero-sum race found themselves in an uncomfortable position. The export controls on advanced chips — the centrepiece of the US strategy to maintain AI dominance — had not prevented DeepSeek from building a competitive model. The CSIS analysis observed that DeepSeek's success "reignited doubts about the effectiveness of stringent US chip export controls," noting that China had demonstrated it could "overcome barriers like limited access to top-tier chips by boosting efficiency or compensating for lower-quality hardware with quantity."18 The Bruegel analysis went further, arguing that the emergence of DeepSeek had "disrupted assumptions about who gets to develop powerful AI" and that the most likely US response — doubling down on export controls — risked accelerating rather than constraining Chinese innovation.19

The Chatham House analysis framed the moment as "a new, more unpredictable era for AI," in which the relationship between capital investment and AI capability was less straightforward than either the Stargate investment thesis or the export control strategy assumed.20 The Royal United Services Institute questioned whether DeepSeek constituted a genuine "Sputnik moment" for the US-China AI competition, arguing that the analogy was imprecise but that the underlying challenge — a competitor achieving frontier capability through an unexpected path — was real.21

The capital question

The implications of DeepSeek's efficiency extended beyond geopolitics. If competitive AI could be built for six million dollars rather than hundreds of millions, the entire economic structure of the Western AI industry was called into question.

The Stargate Project, announced on January 17, 2026, represented the apotheosis of the capital-intensive approach: $500 billion in committed investment, backed by OpenAI, Oracle, SoftBank, and the Emirati sovereign wealth fund MGX, to build AI infrastructure at a scale that would — in the project's framing — ensure American AI dominance for a generation.22 The investment thesis rested on the scaling hypothesis: more compute, more data, more infrastructure would produce more capable models, and the countries and companies that invested most would win.

DeepSeek's release three days later did not disprove the scaling hypothesis — larger models trained on more data could still outperform smaller ones in some domains. But it demonstrated that the relationship between investment and capability was not linear. A team with the right techniques could achieve comparable results with a fraction of the resources. This had implications that extended well beyond the US-China competition.

For smaller countries, DeepSeek's efficiency meant that frontier AI was no longer the exclusive preserve of entities that could mobilise billions in capital. Academic institutions, open-source communities, and governments with modest budgets could download, fine-tune, and deploy a frontier model for purposes that reflected their own values and priorities — whether those were healthcare in Sub-Saharan Africa, climate modelling in the Nordic countries, or education in Southeast Asia. The MIT licence imposed no restrictions on who could benefit.

For the AI industry itself, the capital question was existential. If the moat around frontier AI was not compute but technique — not hardware but algorithmic innovation — then the billions invested in data centres, GPU clusters, and power infrastructure might represent not a competitive advantage but a sunk cost. The venture capital model that had funded the American AI industry assumed that capital-intensive barriers to entry would protect incumbents from competition. DeepSeek suggested that those barriers were lower than anyone had assumed.

The open washing problem

DeepSeek's genuine openness threw into sharper relief the practices of Western companies that claimed the label without earning it.

Meta was the most prominent example, but not the only one. The company had invested heavily in positioning its Llama models as the standard-bearers of "open-source AI," describing them in press releases, blog posts, and congressional testimony as open source. The Llama Community Licence told a different story. It prohibited using model outputs to train competing systems. It required companies above 700 million monthly active users to obtain a separate licence. It gave Meta the unilateral right to revoke the licence for violation of its terms.23 By any established definition of open source, Llama was not open source. It was a proprietary model distributed under terms that gave Meta strategic advantages while capturing the reputational benefits of openness.

The OSI acknowledged, in its analysis of the Open Source AI Definition process, that it had calibrated the definition to accommodate industry realities — recognising that requiring full training data disclosure would effectively disqualify every frontier model, including those whose developers had genuine open-source intentions.24 The compromise — requiring "data information" rather than data itself — was pragmatic. But it created a category of "open-source AI" that provided freedom to use without freedom to audit. A researcher could download DeepSeek R1 and run it. They could examine the model weights, study the architecture, and test for bias or harmful outputs. What they could not do, even with the most permissive licence, was fully reproduce the model without access to the training data that shaped its behaviour.25

This limitation applied to all open-weight models, regardless of licence. But it was particularly consequential for AI systems trained under data governance regimes that were opaque by design. A model trained on data curated under Chinese censorship requirements would reflect those requirements in ways that might not be detectable through model weights alone. The MIT licence granted freedom to deploy the model. It did not grant transparency about the values embedded in its training.

The questions that linger

DeepSeek demonstrated that frontier AI does not require frontier capital. It demonstrated that export controls designed to constrain a competitor's capabilities may instead drive innovation in efficiency. It demonstrated that the most permissively licensed frontier AI model in the world could come from a jurisdiction whose approach to information governance was the antithesis of the open-source movement's founding values. And it demonstrated that the Western AI industry's use of the term "open source" was, in many cases, a marketing strategy rather than a commitment.

The questions that follow from these demonstrations do not have tidy answers.

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

Can the Western AI industry's capital-intensive model survive competition from efficient alternatives? The answer depends on whether the scaling hypothesis holds at the frontier — whether the largest models, trained on the most data, will continue to outperform efficient competitors, or whether techniques like DeepSeek's will close the gap permanently. If efficiency wins, the concentration of AI capability in a handful of well-capitalised companies is not inevitable but contingent — 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, civil society organisations that can use open models to address their own challenges without dependence 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 costs.

DeepSeek did not answer these questions. It made them impossible to avoid. Six million dollars bought a model that the world's wealthiest companies had spent billions trying to build. The implications of that fact — for industry, for governance, for the distribution of AI power across the globe — will unfold over years. The six million dollars was the beginning of the conversation, not the end.

Footnotes

  1. DeepSeek, "DeepSeek-R1 Release," API documentation, 20 January 2026. ↩

  2. Chatham House, "Trump, Stargate, DeepSeek: A New, More Unpredictable Era for AI?", February 2025. ↩

  3. DeepSeek-R1 benchmark results: AIME 2024 (79.8%), GPQA Diamond (71.5), MATH-500 (97.3%). ↩

  4. IT Pro, "DeepSeek's R1 Model Training Costs Pour Cold Water on Big Tech's Massive AI Spending," January 2026. ↩

  5. Sovereign Magazine, "China's DeepSeek Takes On US Tech Giants: What This Means for Project Stargate," January 2026. ↩

  6. CSIS, "DeepSeek, Huawei, Export Controls, and the Future of the US-China AI Race," 2025. ↩

  7. GitHub, deepseek-ai/DeepSeek-R1, January 2026. ↩

  8. DeepSeek-R1 technical report, January 2026. ↩

  9. CSIS, "DeepSeek, Huawei, Export Controls, and the Future of the US-China AI Race," 2025. ↩

  10. MIT Licence, full text. ↩

  11. Meta, "Llama 3 License," 2024. ↩

  12. Open Source Guy, "Significant Risks in Using AI Models Governed by the Llama License," January 2025. ↩

  13. Open Source Initiative, "Meta's LLaMa License Is Still Not Open Source," 2024. ↩

  14. Open Source Initiative, "The Open Source Definition," version 1.9. ↩

  15. OSI, "Open Source AI Definition 1.0," October 2024. ↩

  16. The New Stack, "The Case against OSI's Open Source AI Definition," 2025. ↩

  17. Observer Research Foundation, "Global AI's 'DeepSeek Moment': Impact and Implications," 2025. ↩

  18. CSIS, "DeepSeek, Huawei, Export Controls, and the Future of the US-China AI Race," 2025. ↩

  19. Bruegel, "The Geopolitics of Artificial Intelligence after DeepSeek," 2025. ↩

  20. Chatham House, "Trump, Stargate, DeepSeek: A New, More Unpredictable Era for AI?", February 2025. ↩

  21. RUSI, "What Would a 'Sputnik Moment' for US-China AI Competition Look Like?", 2025. ↩

  22. White House, "The Stargate Project," January 2026. ↩

  23. Meta, "Llama 3 License," 2024; OSI, "Meta's LLaMa License Is Still Not Open Source," 2024. ↩

  24. Kluwer Copyright Blog, "Open Source Artificial Intelligence Definition 1.0 — A 'Take It or Leave It' Approach for Open Source AI Systems?", 2025. ↩

  25. The New Stack, "The Open Source AI Definition: What the Critics Say," 2025. ↩