“There is a persistent belief that Europe is falling behind the United States and China in the global AI race. Yet Europe will never compete on sheer scale alone. Where Europe can lead is elsewhere: in trust, frugality, and human-centred AI. Europe can be a world champion in frugal AI.” - Francesco Ferrero, Leader of the Flagship Initiative on AI and Head of the Human-Centred AI, Data and Software Research Unit at LIST
Less is more
Last year, Microsoft ran an experiment that changed how a growing number of AI researchers think about the field. Two models were tested on advanced mathematics — the qualification paper for the US Math Olympiad. The first had 671 billion parameters. The second had 14 billion. A size ratio of nearly fifty to one. The smaller model won.
Microsoft had taught the small model to reason the way the large one did, transferring thinking patterns rather than raw data. The result was a more capable system at a fraction of the infrastructure cost.
DeepMind had reached a similar conclusion three years earlier: a 70-billion-parameter model, trained on the right data, outperforms one four times its size on the same compute budget. The lesson cuts through a decade of received wisdom. How you train matters more than how much you build.
Twenty watts. No data centre required.
The timing of this shift could hardly be more convenient for Europe. In the United States, the infrastructure bet is hitting a political wall. Moratorium bills have been introduced in at least eleven states in 2026, and counting local-level pauses, more than 69 jurisdictions had active data centre restrictions as of early 2026. The reasons are consistent: power grids under strain, water use out of control, household electricity bills up by as much as 267% in areas near data centres since 2020. In 2025 alone, $156 billion worth of projects were blocked or delayed across the country.
Frugal AI sidesteps that problem entirely. The design principle comes from the most efficient computing system ever built: the human brain. 86 billion neurons, of which only a tiny fraction fire at any given moment, drawing around 20 watts in total. The insight is an engineering direction: activate only what the task requires, and nothing more.
Frugal models consume a fraction of the energy of their large-scale counterparts. They can run locally with no data ever leaving the premises. And when you couple a small model with proprietary data you would never entrust to a hyperscaler — patient records, industrial processes, confidential legal files — the result is a competitive advantage built on assets your competitors cannot access, at a cost they would not expect.
“Powerful techniques such as distillation and quantisation now make it possible to build very small models, under three billion parameters, that perform impressively on local devices. And they offer two advantages at once: frugality and sovereignty.” - Francesco Ferrero, Leader of the Flagship Initiative on AI and Head of the Human-Centred AI, Data and Software Research Unit at LIST
From benchmark to bedside
Mario Draghi's competitiveness report put Europe's structural issue plainly: world-class science, consistently failing to reach the market. The frugal AI approach addresses this failure directly, not through a new funding mechanism, but through a different cost structure.
DFKI trained a medical AI model from scratch with one hundred million parameters, seven hundred times fewer than the leading open-source alternative, on a curated corpus of one billion medical tokens. Every token selected for relevance. None scraped indiscriminately. The model was not adapted from a general-purpose system. It was built for medicine from the first line of training.
It outperformed the leading alternative on every standard clinical benchmark: MedQA, MedMCQA, and PubMedQA, the tests that measure whether a model can reason about diagnosis, treatment, and clinical evidence. That performance points to a concrete future application: a tool like the LIST AI Sandbox could produce a full MDR certification pathway.
Seven hundred times fewer parameters. Better results.
It runs on a clinical workstation that already exists in every hospital. No GPU cluster. No cloud contract. No patient data leaving the institution.
"We have spent a decade measuring AI by the wrong metric. The question was never how large a model is. It was always how much intelligence you get per unit of resource consumed. Every result that has arrived in the past three years, from DeepMind to DeepSeek to our own laboratory, points in the same direction: frugality is not a constraint on capability. It is the source of it. The institutions that internalise this first will build things their competitors cannot replicate by spending more.” - Wolfgang Maaß, Director of German Research Center for Artificial Intelligence (DFKI)
Scale was a proxy for capability. It was never the capability itself. The companies and institutions that build for frugality from the outset, treating resource constraints as design objectives rather than budget limitations, will hold a structural advantage that cannot be closed by infrastructure spending alone. Their edge will not be in the size of their systems. It will be in the precision of their thinking.
That is the race Europe is positioned to win.
Francesco Ferrero and Wolfgang Maaß are presenting this vision on the Horizon Stage at Nexus Luxembourg, 10 June 2026 at 1:45 PM.