As the rest of the world doubles down on brute-force AI, Europe has a unique opportunity: rethink the hardware from the ground up and win through efficiency rather than scale, argues Bernardo Kastrup.
Europe’s quest for AI sovereignty often sounds like a shopping list for expensive and elusive assets: leading-edge fabs, domestic GPUs, massive AI models and compute clusters to rival those of the US and China. But there’s another, more feasible path – one that builds on Europe’s strengths and existing semiconductor infrastructure.
The current AI boom is powered by brute force. The chips at the heart of it weren’t even designed for AI. They’re graphics processors, built for rendering video games, now repurposed for large-scale inference and training. This makes the entire ecosystem catastrophically inefficient. Energy use, system cost and cooling demands have ballooned to absurd levels.
Here lies Europe’s opportunity. Our so-called disadvantage in chip manufacturing is only relevant if we assume the current AI hardware paradigm is the endgame. But it isn’t. If we design chips and systems from the ground up for the actual demands of AI, we can achieve orders-of-magnitude improvements in efficiency, cost, energy and size.
At that point, we no longer need the most advanced manufacturing technology. Our existing fabs, enhanced with relatively modest short-term upgrades, become good enough, as long as we’re prepared to trade in some of the efficiency gains for the use of less-than-leading-edge nodes. After taking the hit, we might still come out on top, or at least match the computing power obtained by brute force.
This is a message Europe’s policymakers should take seriously. Rather than investing billions chasing an unlikely equivalence with Asian and US fabs, Europe can gain strategic autonomy by creating an alternative supply chain: design, fabricate and deploy AI systems tailored to sovereignty applications, small physical footprints, high security constraints and low power budgets, such as on-premises AI data centers. Areas like government, corporate, military and financial systems are all waiting for such solutions. These are sectors where the need for sovereignty is highest and the impact of Europe’s manufacturing lag minimal, due to lower cost-sensitivity and volumes than consumer hyperscale chatbots.
As such, the scenario outlined here offers a compelling alternative to the prevailing narrative that Europe only needs to develop its own homegrown AI software models to ensure their alignment with European values and regulations. Software sovereignty is impossible without hardware sovereignty, as alignment can only be achieved by training and fine-tuning the models in our own hardware. Any serious attempt to align AI models with European values requires hardware we control – hardware we can audit, certify and trust. Relying on compute sourced entirely from geopolitical rivals or commercial monopolies puts us in a position of structural dependence. And when the time comes to train or refine our own models, we may find that access isn’t guaranteed.
Today’s AI hardware isn’t smart. It’s general-purpose, bloated, power-hungry. We can do better. Not by mimicking Nvidia, but by designing dedicated architectures for AI workloads, from the ground up. This is how we break the cycle of dependency. Not by catching up to Taiwan’s fabs, but by changing the game entirely.
This opinion piece is based on an essay published on IAI.