The Sovereignty Gap

That was Anton Carniaux, director of public and legal affairs at Microsoft France, testifying on June 18, 2025, before a French Senate inquiry into public procurement and European digital sovereignty. The senators asked the only question that has ever mattered here: can you guarantee that French citizens' data will never be handed to US authorities without French authorization? Carniaux, to his credit, did not dance. "No, I cannot guarantee that, but, again, it has never happened before."

The second clause is doing an enormous amount of work. It is not the load-bearing one.

I am writing about this thirteen months late, which turns out to be the useful vantage. A sworn admission is a fixed point; what moved around it since is the actual story, and there was no way to see that in July 2025.

i · the statute was never hidden

The mechanism is the 2018 CLOUD Act, which says American companies produce data on lawful American demand regardless of which country the servers are sitting in. That is the entire statute in one sentence. It is not a loophole, an oversight, or something the lawyers missed — it is the explicit, stated design, and it has been publicly available reading since 2018. Every procurement officer in Europe who signed a data-residency addendum in the years since was signing a document the CLOUD Act had already pre-empted, in writing, before the ink.

One honest concession before going further, because it changes where this lands. The CLOUD Act is not a uniquely American overreach. The UK's Investigatory Powers Act, as amended in 2024, reaches any service used by people in the UK regardless of where the provider is incorporated; other states have their own versions of the same instinct. If extraterritorial legal reach were an American peculiarity, the answer would simply be leave the US clouds, and this would be a procurement memo rather than an argument. It isn't. Every jurisdiction you could flee to has a hand of its own.

Which means sovereignty is never absolute and never purchasable. It is always relative to a jurisdiction you have chosen, with your eyes open, and defended as a posture you maintain. The failure in Paris was not choosing the wrong vendor. It was believing the property could be bought as a feature at all.

ii · three years building the wrong thing

Here is the part that should actually unsettle you, and it is the engineering part.

Pierre Lagarde, Microsoft's technical director for the French public sector, testified alongside Carniaux and described three years of work building exactly what was promised: customer data held within the EU at rest, in transit, and in processing, application logs included. Boundary enforcement, the whole build. I believe him. I have no reason to think it does not work as specified. It probably works beautifully.

And it is completely irrelevant to the question that was asked.

I want to state that carefully, because the sweeping version — engineering can't solve this — is false, and falsifiable by exactly the readers most inclined to argue. Engineering absolutely can solve it. If the provider holds no usable keys, a CLOUD Act order served on the American parent produces ciphertext and a shrug. Customer-managed hardware security modules, client-side encryption, air-gapped deployments, or a locally-owned operating entity all change the answer. We know this because those designs exist, in market, in Europe — Bleu in France, Delos in Germany, disconnected Azure Local installations — and they exist precisely because the CLOUD Act works the way Carniaux described.

So the precise claim is far more damning than the sweeping one. It is not that engineering cannot close this gap. It is that Microsoft spent three years building the one architecture that cannot close it, and sold it into a market that was buying the other thing. Residency without key custody is an expensive, immaculate answer to a question nobody asked. Data residency is a network topology problem. Sovereignty is a jurisdiction problem. Microsoft genuinely solved the first and shipped it to customers paying for the second, and the difference between those two is a piece of paper served on a Delaware corporation that still holds the keys.

That gap is not a lie, exactly. It is worse than a lie, because everyone involved could point at something real — and because the choice of which architecture to build was made by people who understood the difference perfectly.

iii · what a year did to "it has never happened before"

So: thirteen months. What moved?

The law did not. As of 2026 there is still no instrument repealing or blunting the CLOUD Act's extraterritorial effect, and no absolute guarantee is available from any US-headquartered provider. That much is exactly where Carniaux left it.

The engineering moved, somewhat, and in a revealing direction. In November 2025 Microsoft announced a substantial sovereignty push — the EU Data Boundary extended to AI workloads and telemetry, in-country processing for Microsoft 365 Copilot across fifteen nations, Microsoft 365 Local running on Azure Local at far larger scale. Through 2026 it has added confidential computing, encryption during processing. Every one of those is real work, and not one of them moves the key custody question. Independent analysis of the 2026 sovereign cloud reaches the same place the senators did: Microsoft remains a US-headquartered provider subject to lawful access requests, and sovereignty risk is only partially answered. The partner-operated entities are where the actual answer lives, and they remain the exception rather than the default.

The vocabulary moved most of all, and this is the part I did not anticipate. Critics stopped arguing about datacenter maps and started arguing about the word. Civo's Mark Boost called the announcements "data residency, not true sovereignty — if the company is still governed by US law, the data ultimately sits under US jurisdiction." Nextcloud's Frank Karlitschek called it "sovereignty washing." And in October 2025 the EU published a Cloud Sovereignty Framework setting out eight specific requirements a service must meet to use the term.

That last item is the most important thing to happen in the intervening year, and it is not a technical development at all. It is a definitional one.

iv · the enclosure happened in language

Because here is the layer underneath the sold-versus-legally-possible gap. The CLOUD Act was never hidden — I have said so twice, and it is the whole point. So the enclosure did not happen in law. There was no moment when Europe's options were legally foreclosed without notice.

It happened in language. "Sovereign" was quietly redefined as a product attribute — a checkbox in a procurement document, a tier in a pricing table — by the only party structurally incapable of delivering the thing the word originally meant. Whoever gets to define a term gets to sell compliance with it. A senate committee is where you find out the definition was never yours.

Which is why the EU writing down eight requirements matters more than any datacenter Microsoft builds. It is Europe taking the word back. Not the data — the word. That is a smaller victory than it sounds and a larger one than it looks: you cannot procure a property you have let someone else define, and the first move in any enclosure is always the vocabulary.

What Europe was actually offered, once you strip the maps away, is a contractual commitment to resist requests Microsoft deems unfounded. That is a corporation choosing well, repeatedly, on the record. But a sovereignty that depends on your vendor continuing to choose well is not sovereignty. It is goodwill with a service level agreement attached, and goodwill has no legal standing in a jurisdiction you do not sit in.

v · the best fact in the story

Carniaux offered the Senate one piece of evidence for Microsoft's willingness to fight: the company had resisted US data demands before and carried a case all the way to the Supreme Court. That is true. It is also the single most self-defeating thing said in the hearing, and nobody in the room appears to have noticed.

The case was United States v. Microsoft — the Ireland warrant case, over emails on Dublin servers sought in a narcotics investigation. The Supreme Court heard argument in February 2018. It never ruled. On March 23, 2018, while the case sat undecided, Congress passed the CLOUD Act, written specifically to settle the question Microsoft was litigating. On April 17, 2018, the Court dismissed the appeal as moot. The government withdrew its warrant and simply obtained a new one under the new statute.

So the proof of protection Microsoft offers a foreign legislature is the exact case that produced the law now defeating it. Their evidence of resistance is the birth certificate of the thing they cannot resist. Microsoft did not lose that fight — the fight was legislated out from under them mid-argument, which is a far more precise description of what European buyers are actually exposed to than any sentence about servers.

You cannot own the layer while you are renting the rack. This is the shape of every enclosure: the commons gets rebuilt as a service, the service is genuinely excellent, the terms are genuinely disclosed, and the ownership question quietly stops being asked because the thing works so well nobody wants to relitigate it.

I predicted last year that almost nothing would migrate. Thirteen months on, that has aged well and for a duller reason than I expected. The switching cost is concrete, immediate and invoiced; the risk is abstract, deferred, and has never happened before. What changed was not where the data lives but who controls the word for where it lives — which is either the first move of a real reclamation or a very sophisticated way of standing still.

It is the same story as a town council in South Carolina approving 22 license plate cameras this month: a buyer purchases what it believes is a product and actually accepts someone else's jurisdiction, the engineering performs exactly as promised, and the ownership question is never once put to a vote. Twelve thousand people or sixty-eight million, the structure is identical.

That hearing did not change the situation. It produced a clean, sworn, citable admission of it — useful mainly later, as the document you hand to whoever claims nobody could have known.