A clear-headed, fact-based way through the noise.
The week sovereignty got a date on it
On 12 June 2026, a single US government directive ordered Anthropic to suspend access to its two most capable models, Fable 5 and Mythos 5, for any foreign national, inside or outside the US, including its own foreign-born employees. With a scope that broad, enforcing it selectively was never realistic, so Anthropic made the only responsible call and switched the models off for every customer worldwide. Eleven days earlier, it had confidentially filed a draft IPO registration.
None of this is new, only faster. Digital sovereignty has been a slow-burning concern in Europe for years, driven above all by the market dominance of the US hyperscalers: the quiet reality that the cloud, the tooling and increasingly the intelligence under European business all sit in someone else's hands. What's changed lately is the temperature. Under the new US administration, a long-term strategic worry has hardened into a near-term operational one, and the Anthropic directive is the sharpest signal yet.
That's exactly what makes the case instructive. This wasn't a careless vendor or a botched response. It was a well-run company making the right decision under an impossible directive. The capability still vanished overnight, for everyone, regardless of where anyone's data was hosted. Sovereignty stopped being a conference-panel word and became an operational risk with a date on it.
The government's stated concern was a jailbreak that could point Fable 5 at finding software vulnerabilities, which Anthropic disputes as no different from what rival models already do. But grant the security rationale in full and the second-order effect is unchanged: a directive that disables a US champion's flagship days before its IPO, while cutting foreign access to the most capable models, hands every non-US buyer a fresh reason to reduce its dependence on US AI. If the aim was to entrench American dominance, it may have done the opposite. Intent is debatable; the effect is what you have to manage.
Two reflexes, both wrong
The reactions worry me as much as the dependency.
The first is new. You judge technology by its passport, not its merit: reject the model, the cloud, the chip for where it's from, not whether it's the best tool for the job. It feels patriotic; it's usually just self-harm with a flag on it. One caution up front, because it's the obvious objection: judging by origin is not the same as weighing political risk. "Where is this from?" is the reflex. "What political risk does it carry, and can I live with it?" is the discipline, and it earns its own lens later.
The second is old, the one that got us here: the comfortable "it'll be fine" we've run for 10 to 15 years, the assumption that a strategy paper and a sovereign-cloud logo add up to sovereignty.
Both skip the actual work: deciding holistically, case by case, with the numbers in front of you.
A word on method: I'm deliberately keeping this non-tribal, judged on merit and risk, not allegiance. Not because the politics don't matter (they plainly do), but because political heat is exactly what tips people into one of the two reflexes. Turn the temperature down and you can actually decide. Politics still gets a seat at the table, as one lens among several, not the one that shouts loudest.
This isn't new: ask the car industry
If the Anthropic shock feels novel, it isn't. Europe's own industry lived the same lesson in hardware months earlier. In October 2025, after the Dutch government seized control of Nexperia (a Netherlands-based chipmaker owned by China's Wingtech, under US export pressure), China retaliated by blocking exports from its Chinese plants. Nexperia makes around 40% of the world's automotive transistors and diodes by volume. Carmakers had two to four weeks of inventory. Production wobbled worldwide; Honda alone flagged a hit of around 110,000 vehicles and roughly
Notice what that case is not: a story about foreign technology. Nexperia is European. The passport reflex wouldn't have saved a single production line. The failure was concentration plus no backup: one supplier for a critical part, run on just-in-time inventory, by an industry that had already been burned the exact same way in 2021. That's not bad luck. It's a managed risk left unmanaged, self-deception with a balance-sheet cost.
That's the through-line. Whether the chokepoint is a Chinese-owned Dutch fab or a US AI lab, the lesson is identical: concentrated dependence you haven't actively hedged is a bill waiting to arrive. AI just sent Europe the same invoice, in software.
The honest numbers
Slogans don't survive contact with the data. The figures below are as of mid-2026.
Compute. Europe holds only around 5% of global AI compute, against roughly 80% for the US. (Europe 2031 report, 11 June 2026)
Capital. The EU's InvestAI initiative, a target to mobilise €200B (roughly €50B of EU money, much of it repurposed, plus around €150B in hoped-for private capital), is dwarfed by a single year of US hyperscaler capex: Google, Amazon, Microsoft and Meta guided to roughly $725B for 2026, up 77% on 2025's $410B. (Q1 2026 earnings; Yahoo Finance / Tom's Hardware)
Chips. About 85% of Europe's AI GPUs are Nvidia (The Next Web), and around 80% of the world's most advanced AI chips are bought in the US (ASML's CEO, June 2026).
Cloud. European providers' share of their own market fell from 29% in 2017 to about 15% by 2022, flat ever since, while the three US hyperscalers now hold roughly 70%. (Synergy Research Group, July 2025)
Models. Europe has serious labs, Mistral above all, but none yet in the Fable/Mythos capability class. And Mistral trained its flagships on US cloud (CoreWeave), on Nvidia H100s. Even our champion stands on US infrastructure. (Mistral coverage, 2026)
And yet Europe is not powerless, which is exactly why the decoupling reflex is wrong. The single most indispensable chokepoint in the entire AI supply chain is European: ASML, the Dutch EUV-lithography monopoly every advanced chip from TSMC and Nvidia depends on, and Europe's most valuable public company. Sovereignty isn't binary. It's a portfolio of leverage, and we hold some of the best cards in the deck. The harder truth is how rarely we play them in concert: ASML's chokepoint, RISC-V, real regulatory weight and an open-source base are far stronger as a coordinated bet than as scattered ones.
Brussels moved, and the vendors pushed back
Europe's institutions have moved, too. On 3 June 2026, the Commission's Technological Sovereignty Package, comprising Chips Act 2.0, the Cloud and AI Development Act (CADA) and an EU Open Source Strategy, reframed dependency as a strategic vulnerability to be legislated against, not a market quirk to tolerate. It even grades sovereignty now: an assessment scale from SEAL-0 to SEAL-4 (the top rung means a full EU supply chain, chips to software), while CADA's binding text scores cloud providers on four "Union assurance levels" and makes "Union added value" a criterion in public procurement.
That's the backdrop for what the hyperscalers did next. This month, Google Cloud published its case for "open digital sovereignty" (in part a direct pushback against CADA's geographic criteria). It's a serious, well-built offer, not a marketing slide.
The menu is tiered: a public cloud with strict European data boundaries; independently, partner-operated regional clouds (S3NS in France, qualified to SecNumCloud 3.2, France's highest bar; T-Systems and Thales in Germany); and fully air-gapped deployments for the most sensitive workloads. On top: an External Key Manager so you hold your encryption keys outside Google's infrastructure, no data-transfer exit fees, open models (Gemma) and open-source foundations to limit lock-in. Google's argument: sovereignty should mean technical, verifiable control and choice, not rigid geographic mandates.
Take it seriously: dismissing it on reflex is exactly the passport trap. But don't swallow it whole. For all the genuine engineering, the piece is notably quiet on the one question the Anthropic case just made unavoidable: US extraterritorial reach (the CLOUD Act, FISA). Customer-held keys and EU-operated partners raise the bar, but the argument strengthens compliance and data residency more than it settles who ultimately controls the switch. Real progress. Not yet a full answer.
That gap, between "compliant and resident" and "actually in control", is the whole game. Which is why you need a test, not a reflex.
Open source, all the way down to the silicon
There's a more fundamental answer to the kill-switch problem than any vendor's sovereignty tier, and Google's own pitch quietly leans on it: open source.
The logic is simple. A model whose weights you hold can't be switched off remotely the way Fable 5 was, though a licence can still be withdrawn, and you still need compute to run it. Within those limits, open-weight models, Europe's own Mistral among them (its smaller models under a genuine open-source licence, its flagships under a more restrictive one), can be self-hosted on infrastructure you control, fine-tuned on your data, and kept running. The same holds down the stack: Linux, Kubernetes and PostgreSQL are open foundations, portable by design, with no single vendor owning the exit. For the model and software layers, open source is the clearest path from renting capability to owning it. It is, almost literally, the architecture you earn rather than the vendor you buy.
And the same instinct now reaches the silicon itself. RISC-V, an open, royalty-free chip instruction set that no company or country controls (its foundation sits, deliberately, in Switzerland, above export-control politics), has become Europe's most credible long-term bet for chip autonomy. It's not a science project: RISC-V passed roughly 25% of the global processor market in early 2026 (largely embedded and edge cores), and the EU is funding it directly through DARE, the largest chip-development project it has ever backed (about €240M, 38 partners), explicitly to cut reliance on American and British silicon, right down to a RISC-V AI accelerator for inference. It won't dethrone Nvidia tomorrow: an open instruction set is not a finished accelerator, the software ecosystem still trails CUDA, and Europe's earlier Arm-based attempt shows how hard this is. But it's the hardware analogue of open weights: a standard you can build on without anyone's permission. And notice where that road ends. In fabrication, where Europe doesn't have a gap but an ace: ASML.
But this is why open source is a part of the answer, not the whole. Those open weights still train and serve on Nvidia silicon and, usually, hyperscale compute; self-hosting means taking on the ops, security and scaling yourself; the best models remain closed, so choosing open often means accepting a capability gap for a while; and "open" doesn't automatically mean "European": much of the strongest open-weight work is American or Chinese. So treat open source as a precision tool, not a flag: it's how you win the control-and-substitutability layers (the kill-switch risk above all), while the compute underneath still has to be solved a different way.
Five lenses, not one
Before you keep, replace, or wrap any part of your stack, run it through five lenses, and never let the last one decide on its own:
1. Technology fit. Does it do the job best, today? A worse tool chosen for its flag is a tax your customers pay. (Nadella's own discipline applies: don't use frontier models for non-frontier problems; match the tool to the task.)
2. Ecosystem fit. Can you own your learning loop around it: integrate, substitute, avoid lock-in? This is where sovereignty actually lives. As Satya Nadella argued this month, the priority is "building a frontier ecosystem, not just a frontier model": owning the loop that compounds into IP no vendor can sell you. He calls it the "key test of your control and sovereignty." A stack you can swap a model out of without losing your company's hard-won expertise is sovereign in the way that matters. It's the Anthropic lesson, too: the firms left standing were the ones that could fail over.
3. Cost fit. True total cost of ownership, including the cost of the dependency itself: switching cost, concentration risk, externally-set pricing, and the regulatory tail-risk the Anthropic case just priced into every frontier-model decision. And, honestly, the cost of building a sovereign alternative: enormous, and rarely yours to carry alone.
4. Target-market fit. Who do you serve, and under which rules? A Mittelstand supplier under NIS2, GDPR and the EU AI Act has different sovereignty requirements than a global SaaS. Build for how your market actually operates (and, per HBR's latest on how people really use AI, how humans actually use these tools), not for an ideology.
5. Political fit. Yes, politics belongs here, but as a measured risk, not a reflex. Whose jurisdiction governs this dependency, and who can throttle, compel or cut it off? What's your exposure to export controls, sanctions, or a directive like Anthropic's, and does it sit within your risk tolerance and your customers' requirements? This is the risk discipline flagged earlier, not the passport reflex: you weigh the jurisdiction, you don't worship or shun the flag. Weigh it with the other four; don't let it overrule them by instinct.
Get these five roughly right, together, and you'll make good calls without a flag in sight. (There are others too, like talent, security, IP and exit-readiness, but these five catch most of it.)
A quick worked example. Say you're choosing the model behind a customer-facing product. Technology fit might point to a closed US frontier model, because it's simply better. Ecosystem and political fit then ask: could you fail over to an open-weight model if access were cut, and how exposed are you to a CLOUD-Act-style order given who your customers are? For a regulated European buyer, the answer is often neither "ban it" nor "bet the company on it": run the frontier API for the non-critical majority, and keep a self-hostable open-weight model warm for the part you cannot afford to lose. Not all-or-nothing. A portfolio.
And the doom-version misses an upside: as the WEF noted recently, agentic AI is collapsing the cost of building. The "super-individual" founder, one person plus a swarm of AI agents, is real (I built a full app solo this way). That lowers the barrier for European builders to create, not just consume. Dependency is the constraint; leverage is the opportunity. Both are true at once.
The strongest counterargument
The sharpest pushback comes from Pieter Garicano and Simon Grimm: stop trying to substitute at all. Dependency, they argue, is historically normal (Europe has leaned on US software for decades and done fine), and building sovereign frontier models is economically irrational when, on their figures, Mistral's revenue was under 1% of Anthropic's (a gap that's narrowing as Mistral nears
They're largely right, which is exactly why "build a sovereign everything" is the wrong reading of this piece. Don't substitute for substitution's sake; on most layers, integrating with the best global tools is the rational call. That's question 1, and it's option 1 done honestly.
But the Anthropic shutdown exposed the one place that logic breaks. "Accept dependency, it's normal" assumes the dependency stays available. It just didn't: overnight, by directive, for everyone. That's a failure mode ordinary vendor reliance never had. So the case for open weights and RISC-V isn't that they'll win the frontier; Garicano and Grimm are right that they won't. It's narrower and more achievable: continuity. No single point of politically-triggered, catastrophic failure. That's not protectionism. It's a backup generator.
The choice we keep dodging
Here's what no strategy paper says out loud. Across three layers, semiconductors (Nvidia and the advanced fabs), frontier GenAI (Anthropic, OpenAI, Google) and hyperscale cloud (Microsoft, AWS, Google), Europe cannot, today, substitute at the scale its economy runs on. That's not defeatism. It's the starting position.
From there, two honest options:
Reduce the dependencies that matter most, and pay the real price: capital at hyperscaler scale, a decade of industrial-policy patience, a capability gap accepted along the way, concentrated bets (the ASML playbook) over spreading thin, and open source where it genuinely buys you control.
Or keep telling ourselves the comfortable story, that a logo and a press release equal sovereignty, and wake up in 2031 with the bill. Europe 2031 is blunt: the window to choose is closing fast, with the last actionable moment as soon as this summer.
What we don't get is the third option everyone reaches for: decoupling-by-reflex. Ripping out the best tools on principle while changing none of the structural dependency buys the worst of both worlds: weaker products and no more sovereignty.
The bottom line
Sovereignty isn't a flag you wave, a vendor you buy, or a tool you ban. It's a portfolio of honest, holistic decisions across technology, ecosystem, cost, market and politics, made with clear eyes about where we're genuinely strong (ASML, open standards, regulation, talent, the learning loop) and where we're genuinely dependent (compute, frontier models, hyperscale cloud).
My own view, for what it's worth: there is no single "sovereign" to aim for, and the answer is an ecosystem, not a fortress. Critical infrastructure and defence demand a different level entirely from consumer electronics or most of healthcare, so the goal is never maximum sovereignty everywhere; it's the right form of it for this organisation, this use case, this risk. And the way to reach that at scale is an open ecosystem where anyone willing to contribute can take part and no one is shut out by their passport. That keeps access to the best the world builds while letting each company and institution earn the sovereignty that actually fits it. Sovereignty as participation, not isolation.
The builders and boards who treat it that way will compound an advantage. The ones who treat it as a slogan, in either direction, will pay for it.
And more important than any of it: this debate has to be had, honestly, in the open. But Europe's signature failure has never been too little discussion; it's discussion as a substitute for decision. So I won't pretend one more think-piece, this one included, moves the needle. What moves it is operators and boards running these five lenses on their own stack, this quarter, and choosing.
Not less ambition. Less self-deception, and the nerve to play the cards we already hold.
Sources
Anthropic, Statement on the Fable 5 / Mythos 5 directive (12 June 2026)
European Commission, Strengthening Europe's Tech Sovereignty (3 June 2026): Tech Sovereignty Package, CADA, EU Open Source Strategy
Nexperia crisis: Dutch government takes control (Al Jazeera, Oct 2025); resolution/de-escalation (Nov 2025); automotive impact (AutoNews / company filings)
Google Cloud, Europe's path to open digital sovereignty (June 2026)
Europe 2031, What getting AI wrong means for us (11 June 2026)
Satya Nadella, A frontier without an ecosystem is not stable (snscratchpad, June 2026)
Silicon Continent, Nineteen thoughts on AI and Europe (Garicano & Grimm, June 2026)
WEF, How agentic AI could reshape what it means to be a founder (May 2026)
HBR, How People Are Really Using AI in 2026 (June 2026)
Supporting data: EuroHPC/DARE; Mistral coverage (Raconteur / CoreWeave, 2026); The Next Web (European AI GPU dependence); ASML CEO on US buying ~80% of advanced AI chips (June 2026); 2026 hyperscaler capex ~$725B (Q1 2026 earnings); Synergy Research Group (European cloud market share, July 2025)
First published on LinkedIn on 23 June 2026. Read and comment there