On 20 January 2025, a Chinese AI laboratory released a model called DeepSeek R1 under the MIT licence, one of the most permissive terms in software. R1 matched the performance of the best frontier models available at the time, had been trained in roughly two months, and carried a documented final-training-run cost of 5.576 million dollars. By the close of trading that day, Nvidia had lost 589 billion dollars in market capitalisation, the largest single-day loss in the history of any company. Every headline that week called R1 "open source". Almost none of them said what that word was supposed to mean.
[BOOK IMAGE: row 128 - DeepSeek R1 breaking upward]
That gap is the subject of this article, not R1's benchmark scores. The word "open" in "open-source AI" has no agreed definition, and nobody, not a board, not a procurement team, not an enterprise architect, can govern a word that does not pin anything down. A vendor's datasheet calling a model "open" is a claim about a label, not a fact about the model, until somebody has checked it against a definition specific enough to fail something.
Testable once, in 1998
The last time this word carried this kind of ambiguity, the industry settled it, at least for software. On 3 February 1998, a small group met in Palo Alto and coined the term "open source" as a deliberate rebrand, aimed at corporations who could not say "free" without sounding like they were taking a side. Bruce Perens then adapted the Debian Free Software Guidelines into the Open Source Definition: ten criteria a licence had to satisfy, with no partial credit. A licence met all ten, or it did not.
[BOOK IMAGE: row 37 - the Open Source Definition, ten criteria]
That binary line, not any sentiment about collaboration or community, is what made the term useful. It gave a lawyer, or a procurement officer, something to check a licence against and answer with a yes or a no, not an opinion to form about a company's intentions. AI broke that test before anyone built an equivalent one for the new substrate, and the years since have been an argument about what the replacement should look like.
The same test, for a different substrate
Start with the company whose own name made the argument first. OpenAI incorporated as a non-profit in December 2015, when its name was still a description rather than a brand. GPT-2 was released in stages in 2019, citing misuse concerns; GPT-3 arrived in 2020 as an API only, with no weights released at all. OpenAI restructured into a "capped-profit" entity in 2019, took one billion dollars from Microsoft and then a further ten billion in 2023, and on 28 October 2025 completed its conversion to a fully for-profit public benefit corporation, removing the non-profit's governance role entirely.
[BOOK IMAGE: row 130 - the OpenAI logo, on an open laptop]
Meta took a messier route to roughly the same place. It released Llama in February 2023 by invitation to researchers; a copy leaked within days regardless. Llama 2 followed in July 2023 as a downloadable model permitting commercial use, but carrying a restriction: any product reaching 700 million monthly active users needed a separate licence from Meta, a threshold set high enough to exclude precisely the handful of companies large enough to compete with Meta directly. The restriction reappeared, essentially unchanged, through Llama 4 in April 2025. The weights were downloadable and the architecture was documented; the training data was not disclosed, and the user-threshold clause is exactly the kind of restriction the 1998 Open Source Definition prohibits outright. The Open Source Initiative concluded Llama was not open source. Meta disagreed and kept using the word anyway. Google's Gemma, released in February 2024, used a "Responsible AI" licence carrying its own restrictions on top. None of these companies were lying about what they released. Each was applying its own definition of "open", tuned to its own commercial position, and none of those definitions was the one the industry had agreed to in 1998.
In 2022 the Open Source Initiative began building an Open Source AI Definition, OSAID for short, to settle for AI the question the 1998 definition had settled for software. OSAID v1.0 landed on 28 October 2024, requiring four elements: the model weights, the architecture, the training code, and sufficient information about the training data for a practitioner to retrain the model. That fourth element was the one nobody could agree on. A model's weights can be public while the data used to produce them stays a trade secret, so OSAID took a middle position, sufficient information rather than the data itself, which Bradley Kuhn of the Software Freedom Conservancy called a capitulation in public, on the record, while the process was still running.
Run real models against the four elements and the split is stark. Pythia, from EleutherAI, and OLMo, from the Allen Institute for AI, passed. So did DeepSeek R1. Llama 4, GPT-4, Gemini and Claude did not. The three that passed are also, not coincidentally, the three a practitioner can actually download in full from a public repository such as Hugging Face and run today; the four that did not pass are not available on those terms, whatever their own marketing calls them.
[BOOK IMAGE: row 131 - the openness taxonomy's four quadrants][BOOK IMAGE: row 132 - Hugging Face, hosting the models that passed]
The Linux Foundation, which had no stake in the OSI's binary outcome, launched a parallel effort at roughly the same time: a Model Openness Framework grading a model on a scale rather than sorting it into a single pass or fail. Its existence is itself evidence that OSAID's binary answer, carefully built as it was, did not end the argument even inside the community that builds open models.
None of this is only an argument about words on a licence page. The manuscript's own framing is that the AI enclosure is neither purely legal, the way Unix's was, nor architectural, the way the cloud's was, but infrastructural: the barrier is physical. Building and operating a hundred-node GPU cluster capable of training a serious model costs between 40 million and 400 million dollars a year. That is not a community resource, and no definition of "open" changes who can afford to build one. A definition covering the weights, the code and the training data says nothing at all about who can afford to run either.
[BOOK IMAGE: row 133 - the physical substrate, an Nvidia GPU chip]
The board consequence
None of this is a story about which model performs best. It is a story about whether a word on a datasheet can be trusted to mean anything, and what a board does before it finds out the hard way. I have sat in enough architecture reviews to know what the gap looks like from the buyer's side: an enterprise architect choosing infrastructure in 2026 cannot audit the training of the model they are about to deploy, cannot verify that a vendor's fine-tuning preserved the properties the business actually needs, and cannot reproduce the system at all if the vendor withdraws it or changes the terms. That sentence, not a benchmark score, is the one a due-diligence checklist should ask a vendor to answer before a contract is signed.
"Open" is doing a lot of work on AI datasheets, and under a rigorous definition, most of the models a board will actually meet in a vendor pitch do not clear it. Released weights are not open source by themselves; a permissive-sounding label is a claim, not a property, until it has been checked. The governance move available to any board is unglamorous and cheap: ask the vendor which definition of "open" they mean, hold the model against OSAID's four elements or a comparable named standard, and do that before any dependence gets built on the word rather than on what the word turns out to actually describe.
The definitional fight itself is worth naming as what it is: a community process, not a vendor decision. The Open Source Initiative is the same nonprofit body that wrote the 1998 Open Source Definition, and it ran the OSAID process the same way, in the open, drafting in public between 2022 and October 2024 and taking criticism as it went. Bradley Kuhn's public objection that the training-data compromise was a capitulation is itself part of that record, not a leak. No vendor got a veto over the final text, and no vendor volunteered to write it. That is the open-source dimension of this argument, and it is also why a vendor's own "open" claim is not the same kind of statement as the OSI's: one is marketing copy, the other is a standard anyone can hold a model against and get a checkable answer.
The sovereignty question underneath this is definitional before it is technical. An organisation, or a country, cannot claim independence over a model it cannot inspect, reproduce, or contest, and a vendor's "open" label does not establish any of those three things by itself. The Open Source AI Definition, published by the Open Source Initiative on 28 October 2024, is the closest thing to a checkable standard the industry has: four elements, weights, architecture, training code, and sufficient information about training data to retrain the model, applied consistently enough that Pythia, OLMo and DeepSeek R1 passed while Llama 4, GPT-4, Gemini and Claude did not. The harder barrier sits underneath even that standard: a hundred-node GPU cluster capable of training a comparable model costs between 40 million and 400 million dollars to build and run for a year, a physical enclosure no licence changes.
Pull up the datasheet for the AI model your organisation is about to depend on, and find the sentence that defines what "open" means there. Is there one, and would it survive being checked against OSAID's four elements?
This is one of seven weekly series in The Hamberger Report. Subscribe on LinkedIn and the next one arrives in your feed.
The views expressed in this article are entirely my own, informed by morethan 30 years of professional experience in architecture, security, andtechnology leadership in New Zealand. I write as director of Te PonoLimited; the views are personal and do not represent the position of anyclient, any government agency, or the New Zealand government. My commentaryon legislation and policy is analytical, drawing on publicly availablesources and my professional expertise in architecture, security, and AIgovernance, and it is politically neutral.
Andreas Hamberger is a New Zealand-based enterprise architect and technology strategist. Over 30 years, he has moved from compiling kernels on a 486 to leading cloud, cyber, and AI transformation programmes across government, banking, transport, and aviation. He founded Yoper Linux, served as a technology specialist for Novell during the Linux Wars, and is the author of "Generative AI: Skynet or Heaven" and "Space Mafia." He can be reached at andreas@thehambergerreport.com. His latest book, Free as in Theft: The Hidden History of Open Source Software, traces the cycle of openness, adoption, enclosure and resistance that runs from IBM's 1956 consent decree to the AI licensing disputes of today.
This article was produced with AI assistance under my direction. Research,drafting and images pass through a pipeline I built and govern: automatedgates for source verification, forbidden language and political neutrality,and my own review before anything is published. The tools include Claude,Gemini and Openart. The frameworks, arguments and editorial judgements aremine and are the same discipline I apply to the AI systems I audit forclients. AI accelerated the work; the thinking, and the responsibility forit, are mine.
References
[1] Hamberger, Andreas. "Free as in Theft: The Hidden History of Open Source Software." Te Pono Limited, 2026. ISBN 978-0-473-78455-3. (Primary source for all Chapter 6 and Chapter 22 facts: the 3 February 1998 Palo Alto coining meeting and the Open Source Definition's ten criteria; the 20 January 2025 DeepSeek R1 release, its MIT licence, training time and cost, and the Nvidia market-capitalisation loss; OpenAI's founding, staged releases, capped-profit restructuring, Microsoft investments and for-profit conversion; the Llama releases, the 700-million-monthly-active-user restriction, and the OSI's conclusion; Gemma's licence; the OSAID process, OSAID v1.0's four elements, Bradley Kuhn's criticism, the pass and non-pass model lists, and the Model Openness Framework; the 40-to-400-million-dollar GPU cluster figure and the practitioner limitations it creates.)