Weekly Dose of AI Sovereignty
Weekly Dose of AI Sovereignty
I keep coming back to a single moment from this week. Not the biggest headline, not the one with the biggest number attached. It was the Anthropic export control reversal. For a few days, access to Claude Fable 5 and Mythos was restricted by US export rules. Then, under pressure, Anthropic quietly lifted the curbs. Anthropic lifts export curbs as AI sovereignty fears grow
The curbs are gone. The damage is not.
What that episode proved, and what I think we should stop underplaying, is that access to cutting-edge AI models can be switched off by a foreign government with barely a press release. The models came back online, but the lesson did not. Sovereignty is no longer a policy aspiration. It is a risk management problem, and every institution and nation is now pricing it into their decisions.
---
The Switch That Could Not Be Unseen
Before Fable 5, AI sovereignty was a nice phrase to put in strategy decks. After it, it became a boardroom question: *What happens on a Tuesday when a US regulatory decision takes your model offline?*
Austria was the first to turn this into a concrete political proposal. The Austrian digitalisation secretary urged the EU Commission to negotiate hosting Anthropic's models on European soil. Austria wants Europe to host Anthropic after US AI curbs You could read that as diplomatic overreach, or you could read it as a small country finally grasping that its intelligence infrastructure sits on someone else's switch.
The US response was telling. Under Secretary Jacob Helberg warned against what he called wasteful duplication of tech stacks, arguing instead for innovation sovereignty through global contribution. US warns against costly AI sovereignty push, calls India indispensable partner Frankly, reading those remarks feels like a homeowner advising you not to build your own plumbing while reminding you he holds the pipes. There is nothing inherently wrong with the sentiment about cooperation. What gives it away is who gets to write the terms.
---
Everyone Is Moving, Everyone Is Stumbling
What followed was a wave of commitments, some impressive, some half-baked, all pointing the same direction.
South Korea announced it would channel fifty trillion won in semiconductor tax revenue into sovereign AI compute. South Korea Plans to Invest 50 Trillion Won in Semiconductor Tax Revenue to Develop 'Sovereign AI' That is a nation-state throwing real money at independence. In the UK, HSBC, Lloyds, and NatWest formed a coalition to build Lumen Sovereign, a frontier model designed to run outside US cloud. HSBC, Lloyds and NatWest to Develop UK's First Sovereign AI Model Sector-level collaboration, not just government posturing.
Europe is moving too. The EU readied a euro 140 million push for sovereign AI servers, EU Readies €140 Million Push to Spin Up Sovereign AI Servers and the new Cloud and AI Development Act introduces a four-tier sovereignty framework for government data, with the highest tiers requiring EU ownership of supply chains that US providers cannot meet under the CLOUD Act. Europe's new tech-sovereignty plan doesn't ban U.S. cloud giants Portugal launched Amalia, its first open-source foundation model, built with EU recovery funds. Portugal debuts first open-source AI model as Europe pushes for tech sovereignty
But here is the uncomfortable bit. The gap between policy ambition and corporate behaviour remains enormous. A recent piece at Racoonteur noted that despite everything, European enterprises still overwhelmingly favour US cloud and model providers. Despite Fable 5 warning, European firms resist AI sovereignty And there are absurd edge cases too: the Swedish startup Eustella promising European sovereignty while running on Chinese foundation models, data residency dressed up as strategic autonomy. Eustella promises European AI sovereignty, your data in the EU, the models from China Sovereignty cannot be a cosmetic claim. If your thinking happens on someone else's model, you do not own the outcome.
---
The Token Economy Is the Real Vulnerability
What strikes me most this week is how Palantir has crystallised the argument that most organisations have been too polite to make. Their nine-point manifesto explicitly targets the token-based consumption model used by OpenAI and Anthropic. Read Palantir's 9-point manifesto that decries tokenmaxxing and trumpets 'AI sovereignty' The word tokenmaxxing is deliberate. It reframes the debate: every token you buy from a frontier lab is not just an API call. It is a transfer of competitive intelligence and strategic dependency.
Alex Karp took this further on CNBC, arguing that the token pricing model extracts enterprise advantage while creating dangerous national security dependencies. Palantir CEO Alex Karp Challenges AI Pricing Models as Stock Climbs I do not need to agree with everything Palantir sells to acknowledge that Karp has identified a real structural problem. The convenience of renting intelligence comes with a cost most procurement teams are not asking their boards about: the slow transfer of institutional knowledge into systems they cannot audit, modify, or take offline on their own terms.
Palantir's stock responded to an expanded NVIDIA deal for sovereign AI models for US government agencies, jumping nine per cent. Palantir Jumps 9% on NVIDIA Sovereign-AI Deal, Palo Alto Networks Climbs 4% The market is beginning to price sovereignty risk. That matters. When capital starts following the thesis, the thesis stops being theoretical.
---
What Comes Next
The UN's AI panel put it plainly this week: genuine sovereignty requires domestic compute capacity and governance frameworks, not just policy rhetoric. AI sovereignty hinges on domestic compute, governance capacity: UN AI panel Australia probably cannot build frontier models from scratch without prohibitive costs. Australia shouldn't try to build its own frontier AI. Here's why India's sovereign AI ambitions face real compute and talent constraints that no policy document can conjure away. India's Sovereign AI Policy has a critical blind spot no one is talking about
None of this means sovereignty is impossible. It means it is going to be uneven, expensive, and frustratingly slow. Some nations will get there. Others will manage partial independence. Many will keep renting intelligence while telling themselves they are sovereign.
What I think is inevitable is that the Anthropic episode will be remembered as the moment the illusion broke. The models came back, but trust did not. Institutions are now building their own infrastructure, forming coalitions, and rewriting procurement rules. Some of it is genuine. Some of it is theatre. But the fact that everyone is talking about it at all means the world has moved on from the era when outsourcing your thinking to a US API was just how things worked.
The question is no longer whether AI sovereignty matters. The question is who gets there first, and who pretends they have arrived when they have not.