Own the Intelligence You Run

2026/09/20

In America, a heated debate has started around one issue, and it has already been politicised: whether AI is moving too fast. Europe has no frontier labs to slow down, and no serious movement to speed up. It’s not our debate.

The key debate for Europe is something else instead: when this technology runs everywhere – in administration, industry, research, defence – will we own the intelligence we run, or rent it?

Rented intelligence

Today, most European AI use is a subscription to an American API. That is dependence in its purest form: the model is trained by someone else, aligned to someone else’s interests and values, priced at someone else’s discretion, and revocable at someone else’s convenience.

We have seen this playbook already: we spent decades building an economy on imported gas because it was cheap and convenient, ignoring every warning Russia gave us – the cutoffs, Georgia, Crimea. When the large-scale invasion of Ukraine came in 2022 we finally did the right thing and broke the dependence, but in a hurry, and at enormous cost. But this cost was not the problem: becoming dependent on a hostile power was.

And let’s not pretend the warnings are absent this time. America has paused intelligence sharing to Ukraine to extract concessions, openly threatened the territorial integrity of a European state, wielded export controls and tariffs against allies, and asserted, through the CLOUD Act, jurisdiction over any data an American company holds, wherever the servers sit. America is not Russia, and it does not need to be: diverging interests are enough. Europe is already rearming, because we finally understood that we cannot rely on American protection. The same logic should be applied to AI.

We cannot afford to repeat the same mistake.

Owned intelligence

But there is a second path, and unlike the frontier race, it is open to us right now.

Open-weight models exist. They trail the frontier by months, and the gap keeps narrowing. A model that runs on our own hardware cannot be arbitrarily repriced, gated, deprecated, or aligned against our interests. Nobody can revoke it, as it is not a service, but our own infrastructure.

Frontier-class open models run today on consumer GPUs and modest servers, on open source inference stacks maintained in public.

Europe can still own its intelligence, and the best time to start doing that is now.

The threat to open source

Here is what should actually worry us: the real target of the safety rhetoric in the American debate is open source. Open models, we are told, are dangerous precisely because nobody can revoke them. Notice what that argument does: it defines the property that makes intelligence ownable as the property that makes it unacceptable.

If America restricts open-weight release, that is their loss. American law ends at America’s border. The biggest danger is Europe importing the same discourse and talking points, ending up treating open release itself as the risk and regulating the one path that is our opportunity to actually run sovereign intelligence. If that happens, dependence on American AI will no longer be a choice at all.

What this means in practice

This is what betting on open source AI means for Europe. We must:

Run it ourselves: Public administrations, universities, and companies handling sensitive data should run open models on European servers as the default. Every procurement contract that specifies “on-premise, open-weight” pulls capability, skills, and infrastructure into existence here.

Fund the open layer: European open labs and open fine-tunes, funded as strategic assets. The cost of training runs of frontier-adjacent models is in the tens of millions, not billions: a fraction of what we spend subsidising legacy industries is enough.

Build the boring parts: Models need electricity and silicon. That means energy abundance, nuclear included, and European compute capacity for training and inference, owned and operated here.

Protect it legally: Anchor the right to publish, host, and run open-weight models in European law, the way we once anchored open standards and free software in public procurement. If restrictions on open weights ever come from elsewhere, Europe will remain the place where open source AI is safe.

The key choice

None of this requires winning the frontier. The technologies that define sovereignty in each era are the ones Europe must own to be truly free, and now intelligence is one of them. We already know what it costs to run a continent on infrastructure someone else controls, and we should never repeat the mistake.

Own the intelligence you run.