Capacity is committed. Supervision is not.

On 11 August 2026, Mistral announced that it is assembling an anchor group of European enterprises whose multi-year commitments will underwrite the compute it intends to build. The commitments convert into what the company calls European Compute Units, a claim on Mistral-built capacity that participants can spend across inference, training or model adaptation as their needs change. The stated target is 200MW across Europe by the end of 2027 and up to a gigawatt by the end of 2030. Amadeus, ASML, Capgemini, Caisse des Dépôts and CMA CGM are all named in the announcement.

Most of the commentary has been about sovereignty, and that is a reasonable place to start. It is not the part that should worry anyone running a delivery organisation.

The shape of the commitment

Read past the gigawatt figure and look at the contract structure. Demand is locked in first. Capital is deployed after. Lenders want the offtake secured before the substations are ordered. That is not how technology gets bought. It is how power stations get built.

This is a genuinely serious piece of capital allocation, and the enterprises signing it are not naive buyers. A commitment of that shape has a term. It has a value. It has a business case, a board paper and a person whose name is on it. It survives a reorganisation because it is a contract rather than an intention.

Now look at the other side of the ledger.

What nobody is signing

I have yet to see an organisation make a multi-year, costed, board-visible commitment to reviewing what its AI systems produce. Not a policy saying that all AI-assisted code is reviewed. A commitment, in the same units as the compute one: a term, a number, an owner.

The asymmetry is not an oversight by anyone in particular. It follows from three things.

Capacity has a unit and supervision does not. Megawatts, GPU hours, compute units. These are countable, so they can be contracted, and once contracted they can be planned around. There is no comparable unit for the human attention required to read what comes out. Nobody writes “four hundred substantive reviews per month, guaranteed until 2030” into an agreement, because the industry has not agreed what a substantive review is.

Capacity scales by signature and review scales by hiring. One of those can be doubled in an afternoon. The other moves at the speed of recruitment, onboarding and the slow accumulation of enough context to know when something looks wrong. Capacity is elastic in a way that judgement is not, and this is the part that is easy to say and hard to feel until the volume arrives.

Capacity sits in the capital plan and review sits in a policy. A budget line has a defender at planning time. A paragraph in a governance document has nobody, and the first thing that happens to it under pressure is nothing, which is the problem. It stays exactly as written while everything around it changes.

Written down is not the same as holding

There is a small illustration of this in the same announcement, in a place most readers will skip. Mistral’s regional endpoints let a customer choose whether inference runs in Europe or the United States. The commitment is stated plainly, and then qualified in the same sentence: processing takes place in the selected region, subject to limited transfers to sub-processors that may occur outside it.

That is not a criticism of Mistral. Naming the exception is more honest than most residency claims manage, and every provider has sub-processors. It is an illustration of something more general. A control that is written down still has edges, and the edges are where it stops working. The written version and the working version are different artefacts, and only one of them survives being asked a hard question.

The same distinction applies with far more force to the supervision side, where most organisations have only ever produced the written version.

What a real commitment would look like

If the compute commitment is contracted, costed and owned, then the commitment to check its output should be specified the same way. Three moves make that concrete.

• Put a term on it. Match the length of the review commitment to the length of the capacity commitment. If one runs to 2030, the other should too. A twelve-month commitment to check five years of output is not a plan, it is a hope with a start date.
• Give it a number. Not a policy paragraph. A budget line, with the headcount, tooling and time it implies. Budget lines get defended. Paragraphs get inherited.
• Measure substance rather than throughput. Count what review caught and what it halted, not how many changes it processed. Volume tells you a queue is moving. It tells you nothing about whether anybody read the change.

The third of those is the one that changes behaviour, and it is also the one most likely to be resisted, because throughput metrics flatter everybody and substance metrics do not.

The question to put in the board paper

Both commitments should appear in the same document, side by side, in the same units. Five questions make the comparison visible:

• What is our compute commitment, in term and in value?
• What is our supervision commitment, in those same two units?
• Who owns it by name, and what happens if they leave?
• What did our review gate stop in the last quarter?
• Could we show that to a regulator, or only describe it?

The fourth question is the one that does the work. A control that has never stopped anything is a description of a control. It may be perfectly well written and still be inert, and the difference between the two is invisible from outside until somebody asks.

The regulatory position reinforces the point without doing the arguing for it. Obligations under the EU AI Act attach to roles and to what systems put into the world. Commission enforcement powers arrived on 2 August 2026. None of that moves because the GPUs moved. Where processing happens is a data residency question. It is not the question anyone is going to ask about the code your systems now write.

The uncomfortable version

European compute capacity under European control is worth having, and the buildout is a better outcome than the alternative. It does not create the gap described here. It prices it.

A gap that costs very little at today’s volumes costs a great deal at ten times the volume. What the anchor commitments do is guarantee that the volume is coming, on a known timetable, with the money already spent.

Which leaves a simple statement of position, and it is not a comfortable one. You have financed the production of work you have not financed the checking of. That is a choice an organisation is allowed to make. It should at least be made deliberately, in a document where both numbers appear.

---

Jason Cruz Falzon is the creator of the Cognitive Delivery Framework and the Cognitive Delivery Index, a governance and measurement operating model for AI-assisted software delivery. He writes on the difference between governance that is documented and governance that demonstrably changes behaviour at the point of action. cognitivedelivery.co.uk · jason@cognitivedelivery.co.uk

Source: Mistral AI, “In-region inference, open models, and new European infrastructure for sovereign AI”, 11 August 2026. A first-party announcement, read as such. Figures are stated targets rather than built capacity. Nothing here is legal advice.