A Kill Switch of One's Own

In June, the European Commission unveiled its Technology Sovereignty Package, a stack of proposals built around a Cloud and AI Development Act and a second Chips Act, and the people selling it were unusually honest about the fear underneath. The worry, in the framing they chose themselves, is a kill switch: the possibility that the cloud running Europe's hospitals and ministries and grid operators answers to a jurisdiction on the other side of the Atlantic, and that the access could be throttled or cut from there. Executive Vice-President Henna Virkkunen said the goal was to guarantee that critical workloads do not carry a remote off switch somebody else controls. Von der Leyen said Europe cannot afford to depend on others for the technologies that keep its hospitals running and its grid stable. The Act proposes to triple EU data center capacity within five to seven years, sort cloud services into four "assurance levels", and push sensitive government workloads away from AWS, Microsoft, and Google. The Commission puts the data center build alone at around two hundred billion euros by 2036, mostly private money. Its own target date for Parliament and Council merely agreeing on the text is the fourth quarter of 2027. The kill switch, presumably, will wait.

The jurisdiction problem is still real. The CLOUD Act concerns access to data under qualifying legal process; it is not a general authority to shut down a cloud service. Sanctions, commercial decisions, and operational failures create other dependencies. On May 7, a cooling failure in one Northern Virginia Availability Zone forced servers to shut down and left customers working through a prolonged recovery. That was a localized infrastructure failure, not the whole region going dark. It still showed why the location and control of a dependency deserve separate scrutiny.

The proposal needs to be judged by what operators can do when a dependency fails.

The Commission includes an open-source strategy, and its sovereignty framework considers operational autonomy, control, resilience, and exposure to third countries, so they are going to more of the primitives involved. My concern is what happens after certification. For example, can an operator move a workload, keep its data usable, and recover without permission from the failed provider? More European capacity can reduce some dependencies, but capacity and assurance levels alone do not demonstrate that those recovery paths work.

Europe has run this experiment before recently, and it kept the receipts. GAIA-X launched in 2020 as the continent's sovereign-cloud answer to exactly this anxiety. Within a year the American hyperscalers were members; one of the French founders, Scaleway, walked out, saying it had no more time to waste on a project rotting from the inside with American Big Tech, and the European share of Europe's own cloud market kept sliding while the project that was meant to reverse the slide held conferences. It is hard to build a moat when you invite the besiegers onto the committee that designs the moat.

There is an older warning in France's Plan Calcul, launched in 1966 to reduce dependence on foreign computer suppliers. Its national champion, CII, was eventually absorbed into Honeywell-Bull in 1976. That history does not prove the current European proposal will fail. It does show that domestic ownership at launch is a poor substitute for control that survives later mergers, suppliers, and operational changes.

Sovereignty has both legal and operational parts. Workloads that can run across independent providers and your own hardware can reduce the reach of any one service-denial decision. That takes more than placing a second logo on the architecture slide. Identity, DNS, networking, data formats, and control planes can preserve the same dependency across nominally different providers. Failover has to be tested, the dependencies have to be independent, and the people operating it have to know what they can keep running when one supplier disappears.

Europe is right to examine who can interrupt its critical systems. The useful test for the package is whether it buys operators a working alternative: a workload moved, a dependency isolated, and a recovery rehearsed. Ask for that evidence alongside the capacity targets and assurance certificates.

Want to learn how intelligent data pipelines can reduce your AI costs? Check out Expanso. Or don't. Who am I to tell you what to do.

NOTE: I'm currently writing a book based on what I have seen about the real-world challenges of data preparation for machine learning, focusing on operational, compliance, and cost. I'd love to hear your thoughts!