Let’s stop the narcissism over small differences about data centres

Europe’s AI debate is stuck on whether to completely obsess over our own full stack or be open to hyperscaler compute. Both camps need the same thing first: strong supply-side measures to make any buildout possible on European soil.

On 12 June, the US government ordered Anthropic to cut off foreign nationals from its most capable models. European companies and public bodies lost access overnight. The order was lifted less than three weeks later, and Europe had no say in either decision.

On 1 October, I moderated a closed-door discussion on the geopolitics of AI at Friends of Europe’s State of Europe high-level roundtable. Officials, politicians, investors, industry representatives and researchers agreed that Europe needs more computing power. Tempers rose over where that capacity should come from. One camp wants a full European stack, from EUV machines to the model layer. The other, as outlined in the Transformative AI Strategy for Europe, wants Europe to welcome US hyperscaler compute capacity onto European soil.

Their disagreement rests on questions nobody can settle today. How much will access to US frontier models matter for economic competitiveness and security in sectors like pharma? Would hyperscaler compute on European soil actually help secure that access? Is there a way to use US artificial intelligence (AI) models without giving precious industrial data away? Can European sovereign data centre builders actually build the compute we will need in the years ahead? Meanwhile, large parts of the policy agenda are the same for both camps.

Why compute cannot wait

AI will decide which firms, and which economies, stay competitive. A study of more than 12,000 firms by the Bank for International Settlements (BIS) and the European Investment Bank (EIB) finds that adopting AI raises the productivity of European firms by 4% on average. In September, the President of the ECB warned that in fiercely competitive sectors such as pharma, finance and defence, a small edge in AI capability can bring outsized gains and lock in a lasting lead. Every model relies on computing power supplied by data centres.

Currently, the world is short of computing power and likely to stay that way. Epoch AI estimates that the capacity to serve AI models grows about 3.4 times a year, while demand grows about tenfold. Meta expects to be supply-constrained “for the foreseeable future”, and Amazon says most of its capacity for 2027 is already reserved.

In a shortage, compute is rationed, and not only by price. Washington has made it clear in the past that US demand comes first for AI – a position shared across political lines. Even without political pressure, a provider that can fill its servers with American customers has little reason to prioritise European ones. A European firm that depends on capacity in Virginia risks being served last.

Building any data centre on European soil takes years. In Europe’s main data centre hubs, a grid connection alone takes seven to ten years, according to the International Energy Agency. By the time the need for more compute is obvious, it is too late to start and European companies risk being outcompeted.

Above all, the Cloud and AI Development Act should set an explicit target for Europe’s share of global compute, for instance 15%. A clear political commitment at that level would give local projects the backing they need to get over the line.

Measures to reduce time-to-power

On the supply side, time-to-power matters more than the cost of energy. Carnegie Endowment modelling puts the cost of each year of delay at about 5.5% of a data centre’s lifetime value. Alongside faster permitting processes, grid operators can auction spare capacity and be rewarded for speed.

Another supply-side gain would come from no longer treating every data centre as if it needs full power every hour of the year. With a flexible connection, a site connects sooner in return for accepting cuts when the grid is under strain, which it rides out on batteries or by shifting work to other sites. Modelling by Aurora Energy Research finds that if data centres accepted cuts in the most strained 2.5% of hours, the power systems of Britain, France, Germany, Iberia and the Nordics could supply an extra 21GW of AI computing capacity by 2030.

Developers must play their part to keep the data centre buildout politically viable. In Finland, Google cleared over 500 hectares of forest for two sites, and at one site kept felling for months after learning an environmental assessment was needed. Regulators have ordered the work suspended and police have been asked to investigate. Every such case undermines the public appetite Europe needs for further data centre development.

The weakest link, though, specifically for the European sovereign compute buildout, is demand. Data centres are financed on long-term contracts, and European customers rarely sign them. In August, ASML, Amadeus, Capgemini, Caisse des Dépôts and CMA CGM showed that it can be done by committing in advance to buy computing capacity that Mistral has yet to build. The largest European companies should wake up to the importance of AI and make such commitments, in collaboration with AI scale ups and institutional investors. Europe’s occupational pension funds hold almost €2.9tn and euro-area households hold about €440bn in US tech firms. Someone needs to stand up and coordinate European demand.

Beyond compute: building the rest of the stack

Even a data centre owned by a European company mostly runs on chips designed in the United States and made in Taiwan. TSMC says Europe has yet to show the demand for leading-edge chips, and ASML says it is “selling absolutely nothing in Europe” because “no chip factories are being built”. Here too, both camps agree Europe needs more of the supply chain. European chip designers such as Euclyd, Fractile and Axelera could be a starting point. Nobody has yet found the perfect policy solution that brings leading-edge chipmaking to Europe, but finding it should be an obsession for Europe.

Freud called it the narcissism of small differences: people who are close fight hardest over what separates them. Europe’s AI camps risk doing just that. Their shared agenda is far longer than their list of disagreements. They should push together for measures to reduce time-to-power and stimulate demand from the public and private sectors, so that Europe does not miss the transformative AI train.

The views expressed in this #CriticalThinking article reflect those of the author(s) and not of Friends of Europe.