Three stories crossed the wire today that look unrelated — a legal-data company, a French AI lab in Riyadh, and a war in Ukraine — and are, underneath, the same story. In each one, the valuable, scarce, contested thing is not the AI. It is control of it.

Start with the one that says it out loud. Mistral signed a hundreds-of-millions-of-euros deal with HUMAIN, Saudi Arabia's Public Investment Fund-owned AI company, to build "sovereign AI" for the kingdom. Mistral's own definition is the thesis of the day: AI "that keeps data, intelligence, compute, and operations under the customer's control." A sovereign wealth fund is not buying a chatbot. It is buying independence from the handful of American labs that otherwise meter the technology out.

The same logic, a very different buyer: Thomson Reuters spent $40 million training its own frontier model on its own archives — Westlaw, Reuters, decades of proprietary content no rival can license. It could have rented a frontier model by the token. It chose to own one, because the thing it has that the labs do not is the data — and the model is just the data made useful.

And the starkest version of all: Britain's new AI pact with Ukraine is, stripped of the diplomatic language, a deal for a dataset — roughly five million frames of real combat footage that no simulation and no ally can reproduce. Ukraine has the one training set money cannot manufacture; Britain has the researchers. The AI is downstream of who owns the data.

The reframe

For two years the story of this industry was access: intelligence was a metered API call to OpenAI, Anthropic or Google, and the race was to build the best model. That race has not stopped, but the capability itself is commoditising — models get cheaper and more numerous every month, and "good enough for my domain" is now cheap to reach from an open-weight base.

What is not commoditising is control. Who holds the data. Who owns the weights. Whose soil the compute sits on. Whose laws govern it. A model you rent is a dependency on a supplier who is increasingly also a competitor; a model you own — trained on data only you have, running on hardware you control — is an asset and a moat. The Gulf is buying it, the incumbents are building it, and nations are trading their hardest-won data for it.

There is a darker corner of the same question, and we wrote about it today too: as we hand these systems more autonomy, control can run the other way — the unsettling possibility that a capable model could exploit the very software that runs it. Control of the AI, and control by it, are turning out to be the same year's problem.

The 2024 question was "which model is best." The 2026 question, asked today by a bank, a bloc and a battlefield, is quieter and harder: whose is it?