The UK organisation for the business of Open Technology

The most powerful models Europe lost access to this year happen to be called the Fable series. It is the kind of coincidence you cannot improve on. The suspension did not create a vulnerability so much as expose a fable that had been told for years: that you did not need to build the substrate of artificial intelligence, only to use it well. Own the application layer, the story went, and let others burn the capital underneath. When the layer underneath was switched off from Washington, the story switched off with it.

The lesson most people drew was the obvious one – you cannot rent a substrate and call it sovereignty. That is true, and it is also the least interesting thing to say about it. Of course frontier models are expensive; of course whoever pays for the compute holds the leverage. If the case for decentralised AI stopped at “we need our own GPUs,” it would be a procurement problem, and procurement problems get solved or they don’t. The reason the suspension matters is that it exposed a deeper dependency, one that does not disappear even if you buy the hardware.

Here I have to bring in my philosophical background and talk about the French philosopher Gilbert Simondon, because he described the shape of the problem before there was a machine to attach it to. Individuation – the process by which a subject actually forms – is, in his account, two things at once. It is collective: a mind does not form in isolation but through feedback loops with a milieu, other minds, other practices, the slow friction of people who do not agree. And it is contextual: it happens inside a language, a craft, a tradition dense enough with shared references to sustain real disagreement. There is no individuation in general. There is only individuation within a milieu specific enough for thought to occur.

Hold those two commitments against the frontier proposition – one model, trained on everything, served to everyone through a single interface – and the mismatch is structural, not economic. A model like that can be extraordinarily responsive and still be uncontestable. You can prompt it; you cannot argue with it the way you argue with a colleague or a tradition. It flattens the milieu into a generic surface. Close the loop through which a subject individuates inside one company’s API, and the individuation that results is not free. It is administered.

This is the argument for decentralisation that does not reduce cost or anti-monopoly politics. It is a claim about the conditions under which people think. You need models embedded in particular languages, particular domains, particular communities of practice – not because the local is romantically superior to the universal, but because the loop through which a subject forms requires a milieu specific enough to sustain genuine collective thinking. One model trained on everything by a single entity cannot provide that. It can simulate responsiveness. It cannot simulate the situated friction of a real epistemic community. The frontier logic is not only concentrated. It is epistemologically impoverished – and plurality is not a concession to scarcity but a condition of the kind of thinking that matters.

The strange thing, the thing that gives this argument political weight it would otherwise lack, is that the CFO arrives where the philosopher does. As AI stops answering questions and starts running operations – procurement, compliance, logistics, decisions at every level – the model a company operates through becomes a layer of its governance. Every inference call constrains a process. No firm should run its core operations through a provider that reads its data and learns its patterns while serving its competitors on the same infrastructure. The commons argument says closed cognitive infrastructure is incompatible with collective thought. The competitive argument says closed operational infrastructure is incompatible with autonomy. Simondon and the CFO, by entirely different routes, demand the same substrate: plural, contextual, governed by those who depend on it.

The uncomfortable part is that wanting this does not make it buildable, and the tradition we reach for first does not fit. We call the alternative open source, but there is no open source AI in the sense there is open source software. Every capable open-weight model – Llama, Granite, the Mistral weights, Qwen, and OLMo, the one that comes from a research institute rather than a company – is a corporate release with a permissive licence attached, and the licence is the whole fragility. Open source software worked because the people who built it owned the means of producing it – a laptop, a connection, time – and because the GPL made openness irreversible: once code was free it could not be enclosed again. There is no GPL for weights. You can open them, but if the data, the methodology, and the training code stay closed, what you have is an artifact a corporation can withdraw tomorrow. A community built on a revocable gift is not a commons. It is a tenancy.

So the commons tradition has to borrow from one it has always ignored – the tradition that knew how to provide things too expensive for any individual and too important to leave to monopoly. Railways, electricity, telecommunications each posed exactly this, and each was answered with an institutional invention: common carriage, public utilities, cooperatives, rural electrification. Capital-intensive infrastructure that is a precondition for participation cannot be governed by the same logic as the activities it enables. It is worth noting who else has reached this conclusion. When Leo XIV placed AI infrastructure, data, and algorithms under the universal destination of goods, a tradition working from entirely different premises landed exactly here. Compute has a universal destination, or it has an owner.

And the challenge is double, because it is not only compute. It is also data – the permissibly licensed material models learn from, before it is scraped into proprietary pipelines and sold back as a service with nothing returning to the commons that produced it. Common Corpus, the two-trillion-token open dataset we built at Pleias, is one very modest attempt at that second layer: a demonstration that what models learn from can itself be open, and governed by something other than extraction. But without public compute, even open data feeds an enclosed pipeline. The two commons rise or fall together.

What is missing is the institutional form, and it is within reach precisely because the software world already invented most of it. A copyleft licence for AI: use the weights, modify them, build on them, but derivatives stay open, and commercial serving returns a share to the commons that produced them. Couple that to cooperative governance and an inference-revenue loop – training as the capital cost, inference as the demand that funds the next run – and the commons sustains itself rather than begging for subsidy, the way energy cooperatives and credit unions always have: those who depend on the infrastructure govern it, and the surplus stays.

None of this exists at full scale. All of it is being attempted in fragments. Whether those fragments converge into a form with the weight to hold is a political question, not a technical one – and for the first time the coalition is real, because the philosopher who wants plural infrastructure and the company that wants operational autonomy have finally begun to ask for the same thing.

First published by OpenUK in 2026 as part of AI Openness: 3 Years on from Llama 2

©OpenUK2026

Download Thought Leadership Download Report View all thought leadership