AI Now warns EU AI push cements US hyperscaler dependence

The more Brussels spends to boost AI adoption and sovereign compute, the more the money routes back to Amazon Web Services, Microsoft Azure and Google Cloud. That is the paradox at the heart of a new essay in TechPolicy.Press by Frederike Kaltheuner and Leevi Saari of the AI Now Institute, which treats Europe's various AI policy initiatives as versions of the same mistake: funding training capacity and enterprise uptake without addressing where inference actually lives.

The mechanism is straightforward. The proprietary frontier models European enterprises want are effectively bundled with the three US hyperscalers, so demand for AI adoption flows back through them. The headline case is DeepL, the Cologne-based translation company that had built its public narrative around GDPR compliance and proprietary models developed at a renewable-powered data centre in Sweden. In April 2026 it informed customers that it was entering a partnership with Amazon Web Services and would no longer process data exclusively on its own servers. SAP's 'sovereign' OpenAI for Germany, similarly, also runs on Microsoft's Azure cloud. The European clouds most often invoked as domestic alternatives are a step behind at the model layer. IONOS's largest, Llama 3.1 405B, last led the field in mid-2024. Scaleway, described in the essay as furthest ahead, offered a model that was near the frontier in late 2025. Hetzner is competitive on price, and Schwarz Digits, the digital arm of the Schwarz Group behind Lidl, wants to become a European hyperscaler.

Kaltheuner and Saari also flag European specialist AI infrastructure players including Dutch Nebius and UK NScale as at risk of being pulled into the same pattern rather than offering an escape from it.

The honest caveats sit inside the DeepL example itself. Global enterprise customers want the same service everywhere, AWS provides low-latency real-time performance that no European cloud can match today, and moving into real-time voice translation from a data centre in Southern Sweden had become unfeasible; the commercial logic for switching is real, not manufactured. What the reporting does not give you is a costed alternative Brussels could procure at frontier quality tomorrow, or a euro figure for how much sovereign AI spending is already sitting inside the hyperscaler bundle. If procurement stays fixated on training compute while inference stays bundled with US proprietary models, the sovereignty budget quietly becomes a hyperscaler subsidy, and the European providers named as alternatives will keep serving last year's open weights on this year's expensive GPUs.