The energy demands of the “bigger is better” approach are becoming increasingly difficult to sustain. Data centres already account for around 1.5% of global electricity consumption and often rely on “very unclean energy generation,” including gas-powered generators, said Francesco Ferrero, Head of the Flagship Initiative on Artificial Intelligence and of the Human-Centred AI, Data and Software Research Unit at the Luxembourg Institute of Science and Technology (List), speaking at the Nexus event in Luxembourg on 10 June 2026.
Europe’s AI challenge requires a different path
Europe’s position in the global artificial intelligence race is often perceived as one of a “laggard” compared to the US and China, noted Ferrero. Currently, only one European model sits in the top 20 rankings of intelligence analytics, and the continent lacks a presence in the GPU market, where a single American firm controls 95% of the sector.
Furthermore, Europe’s data centre capacity is significantly lower than that of its competitors, with the US holding 44% of global capacity compared to Europe’s 10%. However, researchers from the List and DFKI suggest that the current “bigger is better” approach—which simply throws more energy, data, and money at the problem—is a race Europe should not attempt to win.
The environmental impact of the status quo is increasingly untenable. Data centres currently consume 1.5% of the world’s electricity, a figure estimated to reach 3% by 2030, according to Ferrero. In some regions, like Ireland, the lack of grid capacity has forced operators to use gas generators, leading to non-green energy consumption and high water usage. This trend risks a social backlash that could stifle AI’s positive potential.
Europe bets on frugal AI over scale
As an alternative, the “Frugal AI” framework focuses on three pillars: using AI more selectively to protect human cognitive health, employing smaller models, and establishing independent metrics for energy consumption.
The benefits of smaller models are manifold. They offer “sovereignty” by allowing data to remain local rather than being exported to untrusted hyperscalers. Research indicates that smaller models trained on confidential, high-quality data can often outperform massive models that lack such specific access.
Wolfgang Maaß, director at German Research Centre for Artificial Intelligence (DFKI), offered a biological metaphor to illustrates this: the human brain, which operates on less than 20W, is the most efficient computational unit known, far outperforming the 50-tonne infrastructure required to power a whale’s larger brain.
Efficiency is also being improved through architectural changes. Early models like GPT-3 were highly inefficient, activating billions of neurons for a single token. Modern architectures have reduced this activation to less than 10%, moving closer to the 2–5% activation seen in the human brain.
In research settings, simulations of a digital human knee have shown that “efferent regulation”—only transporting essential data—creates more sustainable movement control than “full power” approaches, noted Maaß.
The implementation of frugal AI involves “full stack” innovation, including neuromorphic computing and spiking neurones that mimic biological hardware. These methods can reduce energy consumption during training and inference by 30%.
Edge AI could be Europe’s secret weapon
Additionally, the focus is shifting toward the “edge.” With 7bn smartphones and 1.5bn laptops globally, of which 96% are currently idle, there is a vast distributed resource representing roughly 50trn AI operations. By integrating technologies like WebLLM into browsers, AI can be moved directly into the “pockets of humans,” leveraging existing hardware for a more sustainable, distributed, and smart European AI ecosystem.
