Weekly Dose of AI Sovereignty

Weekly Dose of AI Sovereignty

Something shifted in how I look at AI sovereignty this week. I have been following this beat long enough to spot the cadence: governments announce programmes, startups raise rounds, conferences feature panels with the word "sovereign" in the title. Usually I treat it as background noise, a policy fashion that burns itself out by the next earnings quarter. Then Anthropic was ordered to disable access to Fable 5 and Mythos 5 overnight, and every European government that was quietly running those models in production suddenly had a problem they could not tweet about. The emergency EU-US talks that followed were less a diplomatic event and more a wake-up call delivered in the form of a service outage.

That is when it stopped feeling abstract. Sovereignty is not about pride or political theatre. It is about not having your infrastructure pulled from under you because a foreign administrator signed a form.

The moment access became the issue

The EU-US talks over Anthropic's access cuts exposed something most organisations had ignored. They assumed contracts and commercial relationships would hold. They assumed allies would be treated as allies. Neither assumption survived contact with export control policy. The US government drew a line at passports, not partnerships, and suddenly European ministries and enterprises were locked out of models they had integrated into their operations.

What strikes me is how quickly this rippled outward. The UAE was told to reassess its entire AI strategy after the same passport-based restrictions wiped out overnight access for non-US nationals. South Korea found a senior US State Department official publicly labelling its data localisation requirements as discriminatory barriers. The US is now openly pushing back against UN-led sovereignty initiatives, promoting its own "Pax Silica" alliance instead, which framed digital sovereignty as a problem rather than a right.

The pattern is clear. Access to AI capability is now a geopolitical lever, and anyone running on foreign models is sitting on a switch they do not control.

War, money, and the new floor for ambition

This is not just Europe's problem. The stories coming out this week read like a global inventory of who is preparing and who is not.

Israel committed thirty billion dollars to a National AI Programme that includes a hundred thousand GPU compute infrastructure, a national quantum computer, and dedicated institutes for cyber AI and deepfake defence. That is not a research grant. That is a national security budget line, and it treats AI infrastructure the same way it treats air defence systems.

Ukraine is doing something even more stark. Kyivstar and VEON signed an MOU with Ukraine's Ministry of Economy to build a sovereign AI data centre inside the country. They announced it at the Ukraine Recovery Conference, which tells you everything about the framing. This is not about competitive advantage. This is about maintaining basic digital capability while under active attack. If a country at war considers AI compute sovereignty to be part of its reconstruction, then peacetime nations have no excuse for treating it as optional.

Canada is moving too, with HIVE Digital's Buzz HPC joining Bell and Cohere for a Canadian AI factory that provides domestic compute for model training. Sweden has Brookfield partnering with Telia and KTH on sovereign infrastructure. Japan is building a trilateral coalition with France and India to reduce dependency on both American and Chinese stacks.

Meanwhile, Africa is approaching the problem from a completely different angle. MTN is pursuing "data embassy" architecture that keeps African data under African control while still accessing global AI infrastructure. It is an elegant workaround for a continent that lacks domestic compute capacity but refuses to hand over data governance to foreign hyperscalers.

The gap between ambition and reality

Here is where it gets uncomfortable. Europe holds roughly five per cent of global AI compute. Britain's position as a self-styled AI pole collapses entirely when you realise no British institution has standing to contest an American administrative decision. Australia risks trading water and power for digital dependence as it rushes to build data centres without addressing the underlying dependency problem. Russia wants sovereignty but has a chip problem that sanctions have made nearly impossible to solve.

Even Europe's best sovereignty plays are still half-measures. The SUSE and Openchip partnership to build a European stack on RISC-V accelerators is genuinely interesting, but it is years from production readiness. The EU's bet on Domyn's EUROPA consortium to build a four hundred billion parameter open-source model is the most concrete play yet, but a single startup with access to 2.5 per cent of EuroHPC capacity is a very thin hedge.

I keep coming back to what a Taipei Times editorial put quite plainly. Sovereignty is about options, not ownership. No nation except maybe China can genuinely build everything in house. The question is whether you have enough diversity of access that one political decision in Washington or Beijing does not knock your whole system offline.

What comes next

The stories from VivaTech were telling. The conference that used to be all about growth narratives and partnership announcements was dominated by cybersecurity fears and sovereignty questions. Eighteen thousand people showed up and the conversation was no longer about whether AI would transform industries. It was about who controls the infrastructure that makes AI possible, and what happens when that control is weaponised.

What I think we are watching is the end of the assumption that commercial relationships are stable. The Anthropic ban proved that models can be revoked by administrative order. Export controls can be tightened overnight. Alliances do not protect you when the policy instrument is a passport check.

The nations that get this right will not be the ones that build the biggest GPU clusters. They will be the ones that build the most diverse supply chains, the ones that invest in open-source alternatives, the ones that treat algorithmic efficiency as a strategic priority rather than an engineering nice-to-have. The shift toward small-data efficiency driven by localised compute mandates is already showing how constraints force innovation.

For everyone else, the bill is coming. And it arrives in the form of a Tuesday morning when your models stop responding and no one in your country has the authority to make them start again.