Sorting for Sovereignty
One word. Three layers. Three different stories.
Pardon the eruption into what could be mistaken for a tech newsletter — right now, the 'S' word is everywhere, especially in places where it has political traction. Sovereignty.
A €75M European federated cloud project led by Telefónica, called EURO-3C, launched in March with more than seventy participating organisations across thirteen countries.
Mistral, France’s leading AI company, raised $830M in debt in late March to install nearly fourteen thousand NVIDIA chips in a new data centre south of Paris, calling it part of “an independent AI stack” for Europe.
Amazon launched a European version of its cloud service in January in Germany — €7.8B of investment, structurally separate from the rest of Amazon’s cloud, run by EU-resident staff under a German subsidiary.
France has been touting the €109B in private AI investment Macron announced last year.
India is running a layered model through its IndiaAI Mission — tens of thousands of GPUs of subsidised compute, state-selected startups building domestic foundation models, and Indian-language data platforms underneath.
The UAE and Saudi Arabia are running the opposite play — gigawatt-scale hardware builds (Stargate UAE, HUMAIN) running largely American chips and models, with smaller domestic-language efforts on the side.
Same vocabulary. Different layer arithmetic.
The headlines all use the ‘S’ word, but they’re not all describing the same thing. This is what F10 — From Global Technology Integration to Sovereign Stacks — was pointing at. But the part that’s getting harder to see in the noise is that sovereignty isn’t just one layer.
It’s at least three, and they move at different speeds:
Hardware and power. Chips, data centres, electricity. Capital-intensive, increasingly real, mostly what gets the press releases.
Jurisdiction. Whose laws apply, who can subpoena what, where workloads execute. Slower and more contested. American law can still compel American-owned companies to hand over data even when that data sits on European soil — which means Amazon’s new German subsidiary, however operationally separate, is still inside the reach of Washington’s courts. Europe is trying to work around this with what it calls an “effective control” test, but the legislation that would put real teeth on that idea has just slipped from this year to 2027.
Architecture. Open versus proprietary models, federated versus hyperscale, model weights versus pay-per-call APIs. Slowest, most consequential, least covered.
These three layers can move in opposite directions inside the same product. Amazon’s European cloud is genuinely operationally separated — its own infrastructure, its own staff, its own management. That’s hardware-and-power sovereignty doing real work. But most analysts argue it doesn’t solve the jurisdictional question, because the corporate parent is still American.
Two layers, moving at very different speeds, in the same press release. The reason we keep coming back to this is operational: the organisations we most often work with — NGOs, foundations, multilaterals, cross-border programmes — are the ones who get caught when the layers separate. The cloud contract that suddenly violates a new compliance regime. The hosting arrangement that has to be unwound under deadline. The dataset that has to be moved because the jurisdiction shifted while you were paying attention to something else. That’s the cost of treating sovereignty as one thing when it’s three.
That’s the part worth holding onto. When the next “sovereign AI” headline lands — and there will be several this week — the 10F question isn’t is this real or performed? It's which layer is moving — and whether that's the layer that will land on you when the contracts have to be rewritten.
A few near-term moments where the answer gets clearer:
The EU’s AI Act starts biting in August — for the highest-risk systems, anyway, unless a pending European simplification package pushes the deadline to late 2027. Either way, thousands of deployments will eventually discover what their compliance and jurisdiction stack actually looks like under load.
Europe’s flagship cloud-and-AI legislation has just slipped 18 months. That’s a real signal. The question of how much sovereignty European law will require is now a 2027 question, not a 2026 one — and the political appetite for the harder version of the answer may not survive the wait.
A second tier of European AI mega-facilities — five “gigafactories” at around 100,000 chips each — is supposed to start moving from announcement to deployment this year, sitting on top of the fifteen smaller AI factories the EU is already standing up across member states. Whether “operational” in 2026 means running production AI workloads or ribbon-cut will be a useful tell.
The entanglement worth watching now: Hormuz has exposed how much of the tech ecosystem rests on dependencies that weren't really priced before the conflict began. The vulnerability everyone was tracking was Taiwan; the one that broke was the Gulf.
The world's most important memory-chip producers run on energy that mostly transits one strait, as Carnegie walked through. The helium that those same chipmakers depend on comes mostly from Qatar, where supply disruption has doubled spot prices.
Every sovereign-compute build, anywhere, sits on top of those dependencies. Hardware sovereignty is the layer moving fastest. It's also the layer most exposed.
If you have a clearer answer — or a fifth option — we’d like to hear it. Comment below.
Follow the Project
Beyond this newsletter, you can find 10F on:
LinkedIn — 10F Consortium — longer-form updates and discussion
Bluesky — @10fconsortium.bsky.social — signal tracking and real-time commentary
Instagram — @10fconsortium — visual summaries and forecast highlights
GitHub — 10F Consortium — full forecast archive with version history and transparency, and new tools as we develop them
Does sovereignty even matter? Yes, as a generalised misunderstanding.