Who Can Switch Off the AI Europe Uses? 2/2

The EU’s call for AI gigafactories now closes on 3 December for the large sites and on 4 February 2027 for the medium ones. The original deadline was 12 November. EuroHPC moved it, in its own words, to give member states time to complete their national budget procedures and submit financial commitments; according to sources quoted by Europa Press, not a single bid had arrived by then. This is where the sleeper agent, the risk that a supplier can switch a machine off or turn it against its users, could become infrastructure. The gigafactories are sized to someone else’s yardstick. The large sites must reach a hundred thousand H100-equivalent units, the measure of a race without a finish line that Europe did not choose. Public money comes in as the anchor customer, the guaranteed buyer of computing hours, and only one billion euros of EU money is committed today.

The rules leave room for a closed loop. The call requires European control of the company, the sites and the core data, according to the tender summary by Sullivan & Cromwell. It imposes no chips, and encourages European ones only in the second phase. It also lets hyperscalers, the large US cloud providers, run their own share of capacity, which «may place those activities outside certain AIGF-level sovereignty and data-residency requirements». So European public money would de-risk sites full of US chips that host partitions run by US companies, which sell services to European customers from there. The market is European, while the margin and control of the service sit elsewhere, and the whole thing is called sovereignty because the operating company is based here. None of this is settled yet. The French AION consortium, led by Scaleway, is European, and Mistral is building data centres of its own. In Germany the Telekom and Schwarz project has been scaled back, and Handelsblatt reports, citing unnamed sources, that Hochtief is preparing a bid and talking to Vodafone. Whoever fills those hours will decide whether Europe has bought computing time or a place in the game.

My first proposal is defensive: treat every foreign machine like critical infrastructure. Whoever sells models to public administrations and essential sectors also supplies a version that can be installed on site, with no remote switch, and the «degrade or disrupt» clause of the Cloud and AI Development Act is extended from cloud services to models. Verification has to be pooled, because no single state can afford to open a model and read its internals. The AI Office in Brussels and the national AI safety institutes could become a single verification capability, with access to model internals and to samples of training data. Every administration that adopts a foreign machine keeps an alternative ready, with a plan to switch to it. The Franco-German joint paper already lists substitutability, multi-supplier strategies and exit plans among its dimensions. Making them binding would be enough.

The per-token levy (in English) is where defence turns into play. Whoever profits from selling machines in Europe pays the same rate, whatever their size, and the money funds verification as well as what Europe builds. Today users pay less than the service really costs, so the mechanism should be fixed now, before AI has a real price. Those who fear the cost will land on users are choosing deference. Trump threatens 100 per cent tariffs over digital taxes: negotiate, and do not cave.

My second proposal is active, and it starts from what Europe already knows how to do. Verification can become an industry. Europe has a tradition of certification and testing that the world has relied on for decades, from TÜV to DEKRA to Bureau Veritas. A European capability to verify models would serve anyone who has to trust a machine they did not build, and trust is something you can sell. Then there is the field where the race is still open. By the time the gigafactories are running, around mid-2028, computing will have shifted towards everyday use of models, and the conclusions of July’s Franco-German Council of Ministers name industrial AI. That is where Europe has factories, process data and customers of its own. There is also the freedom not to adopt. The mere availability of a tool does not oblige anyone to use it, and choosing where the machine is not needed is as much industrial policy as choosing where it is. Finally, there are seeds of shared infrastructure that no remote switch can reach, such as the consortium for digital commons founded by France, Germany, the Netherlands, Italy and Luxembourg, or the open-source office suites for public administrations. They are small, and they belong to someone here. Many other countries are users in the same position, from Canada to Japan to Korea, and common terms negotiated with them carry more weight than twenty-seven separate negotiations.

The real knot is pooled money, more than rules. The Draghi report put the additional investment Europe needs at 750 to 800 billion euros a year. Of the thirty billion announced for the gigafactories, one is committed. Berlin has ruled out owning one, and the call itself slipped because member states had not yet found their budget commitments. The AI Act delay was sold as help for European players. The player is missing for lack of common capital: postponing rules costs nothing, while merging budgets takes a political decision nobody wants to sign.

Whether Europe really wants autonomy of use will show in two signatures: the member states’ financial commitments to the gigafactories before 3 December, and a common budget for AI, which no government has signed yet. Refereeing takes the will to apply the rules to the very suppliers you depend on, and playing takes money put together. For now Europe is postponing both.