Almost every serious AI deployment in Europe today rents its foundations from three or four American companies. The model is trained on a US hyperscaler’s cloud, served from it, and often is itself a proprietary US model accessed through an API. That is the architecture everyone has quietly accepted, and it is the one Wakeline, a Düsseldorf deep-tech startup founded in 2025, just raised €2.1 million in pre-seed funding to poke at. “We invest because we believe Europe needs its own architectures in the next AI generation,” said Jan Jeske, partner at neoteq ventures, one of the backers. That is technological sovereignty stated plainly, by a person writing the cheque, and it is the loud reason for the round.

The quiet reason is a specific technical bet, and it is more interesting than the sovereignty language around it. Wakeline’s argument is that the dominant design has a structural flaw nobody talks about because everyone shares it. Today’s models are trained on historical data, deployed, and then frozen until the next retraining cycle. Between those cycles the model does not learn, which means it is always operating on a snapshot of a world that has already moved on. For a chatbot that is tolerable. For a system forecasting something that changes by the minute, it is a permanent half-step behind reality. “Most AI investments today are bets on better models within the same architecture,” said Ansgar Schleicher, managing partner at TechVision Fonds, which led the round. Wakeline is betting on a different architecture instead.

Learning while it runs, not between shifts

The technical idea is to fold learning and deployment into one continuous process rather than two separate phases. Inspired, the founders say, by biological learning, where an organism does not stop adapting in order to retrain, the system keeps adjusting while it operates. The design consequence that matters commercially is the one that follows from it: because it is not tethered to a giant pre-trained foundation model, the platform can run independent of proprietary models and US hyperscale cloud. It is, in miniature, the same instinct behind Vienna’s wager that the winning move is a smaller model, not a bigger one. Both are European teams poking at assumptions the American labs sell as settled, from opposite directions: one shrinks the model, the other rethinks how it learns.

The first live use case is forecasting for European energy markets, where prices and loads shift by the minute and a model frozen since last quarter is worth very little. After that the company names industrial manufacturing and, more striking, neurological research, including the early detection of Parkinson’s, a domain where a system that keeps learning from new patient data rather than waiting for an annual retrain has an obvious appeal. The €2.1 million goes to platform development, go-to-market and hiring.

The founders, and the harder bet

Wakeline was founded by four people, Tim Gülke, Jan Böggering, Simon Sprünker and Merten Tiedemann, two of them carrying doctorates, which is the right shape of team for a bet that is fundamentally a research wager dressed as a company. The round was led by Aachen’s TechVision Fonds with Cologne’s neoteq ventures, a pair of regional German deep-tech investors rather than the usual Berlin or London names, which fits a company whose whole pitch is homegrown infrastructure.

Here is the part to sit with. Challenging the architecture is the rarer and harder path. Tuning a model within the established design is incremental, fundable, and low-risk; rethinking how the model learns is the kind of swing that either reorders a field or quietly fails to scale, and at pre-seed nobody knows which. That unproven-at-scale quality is the honest risk, and it should not be waved away with sovereignty rhetoric. But it is also exactly the sort of foundational, homegrown alternative that European capital spends a great deal of breath saying it wants and then, more often than not, declines to fund in favour of another wrapper on an American model. Four founders, two doctorates, an energy grid to forecast, and a thesis aimed squarely at the cloud everyone else rents: this time, at least, the capital matched the speech.