A recent piece in Foreign Policy (Europe Will Never Be an AI Superpower) rehashes all the trite wisdom of the European “AI maximalists” who essentially recommend Europe further resigns itself to being just a “middle power” and deepens its abnegation – as tech vassal and colony – in the name of getting access to the totemic “AI models” from frontier labs. What is curious is that the piece also singles out EuroStack as the epitome of European delusion and wishful thinking around “AI sovereignty”. It is clear EuroStack is moving the dial – others out there say similar things, but we are the sole target of this attack because we are making a difference and we are moving the mountain. Good.
That said, the Foreign Policy piece so profoundly misrepresents what EuroStack stands for that it deserves a response. Europeans cannot be gratuitously sold down the river in a US magazine (telling the audience what they presumably want to hear – the meme of Europe as pathetic and clueless and weak) without setting the record straight. One needs to question the motives of the writers – and the fact they praise the Europe 2031 dystopian novella, describing the authors as “respected AI researchers and investors”, is telling about the camp they inhabit.
A few points of clarification on where EuroStack stands in the “AI sovereignty” debate.
1. EuroStack (do we even need to say it) is not about cloud specifically, or each of chips, hardware, software, open source, data, AI – it is a full stack approach that cares for Europeans having a bigger role (as innovators, entrepreneurs, investors, vendors) in every component of the digital infrastructure stack. For value capture, direction of travel and growth.
2. For us, AI is a use case: cloud + software + data. Super important as we head into the AI transformation but not the sole focus. So when we talk about “tech sovereignty” we don’t automatically mean “AI sovereignty”. We talk about the stack. And there, the reality is that Europe has hundreds of tech businesses at multiple levels with innovative ideas and actual clients and sales and dozens of extraordinary ecosystems (from Stockholm to Copenhagen to Warsaw to Vienna to Zurich to Munich to Berlin to Paris to Amsterdam and more) for research, startup, development. We know the endemic problems of Europe inside out: no single market, no capital market union, focus on regulation instead of building for two decades, etc etc. We know. It’s frustrating (and why we don’t lean on Brussels and institutions generally). But does not remove the fact that Europe is a superpower. An economic superpower. NOT a tech or an AI superpower. Never said that, FP authors. With problems, challenges and all that, but we are a rich continent with huge capital pools. We are financially and technically sophisticated. We resist the narrative Europe can be written off because we are not in the same league as AI frontier labs.
3. It’s tough, on that we agree. The AI predicament of Europe is tough. The fact that the “AI revolution” that holds so much promise of massive efficiency benefits and business transformation is landing on European boards at the same time as concerns for dependencies and lack of resilience (in short, “sovereignty”) is going to make things even tougher. On the one hand, boards are just now (last three months) wrapping their heads around the “AI rush” and hiring, commissioning, clamouring for advice (a lot of improvised “consultants” out there – terrifying). At the same time, it’s also obvious that the frontier AI models from US labs come bundled with hyperscalers’ clouds. For multiple reasons (including circular finance) the top models are only distributed on US hyperscalers’ cloud. So massive AI adoption (e.g. running agentic teams) will tend to be sold by hyperscalers and their proxies (system integrators) in a bundle with their cloud and stack capabilities. This makes migration to European clouds more challenging. But it also makes clear that AI is cloud, and sovereignty in AI cannot be detached from cloud dynamics (a point well made by Frederike Kaltheuner and Leevi Saari here).
4. There are also other architectures and a view that not everything needs gigantic AI frontier LLM models premised on the same scaling laws. Does it make sense to use these to distil down more specific models for (say) industrial use cases? Does not seem to, and indeed there is a whole view of industrial AI which is premised on these different models and approaches.
5. So at EuroStack we do not talk about “sovereignty” as “AI sovereignty”. We see the links across dependencies. Of course we will not be able to go to the mat with US frontier LLM models and create a challenger. But we don’t believe all is lost for that reason.
6. In particular, we do not accept the logic that “therefore, because we will never make our own model and yet we desperately need access to these frontier models, given the vagaries of the geopolitical environment we MUST genuflect and prey for access in exchange for “doing deals”. These “deals” variously involve extracting commitments for access to these models in exchange for cheaper French nuclear surplus energy, or easier permitting to hyperscalers. A variant of this goes “this will make them dependent on us, and they will not want to renege on their obligations”. Where does one start with this? We agree there are datacentres to be built, and fast. But the idea they should be built by hyperscalers in exchange for access to the models they distribute is perverse further entrenchment of their dominance in Europe. We have no problem building datacentres: we have the technology and the funding. The issue is “demand”: so far Europe has not had enough compute demand for inference to motivate it, and uncertainty over speed of adoption is a factor for European builders – while hyperscalers have their fleets of GPUs already secured and massive internal demand from their legions of captive customers. This is the “chicken and egg” problem of Europe. But again, it does not in our view justify capitulation to “inviting hyperscalers to build for us”. Enforceability of any “promise” also strikes us as dubious in the face of an Executive Order or an instruction from the Administration.
What motivates these half baked suggestions? In our view this is Silicon Valley and hyperscalers’ campaigning, on stilts. Made more acceptable by “Europeans saying it” – but there’s any number of Europeans working as proxy lobbyists for US tech platforms, via mysterious opaque funding of their institutions.
We also find puzzling the focus of otherwise reasonable Prof Garicano in Silicon Continent on a breathless “race to adopt”. He does propose a few mitigation factors to reduce the probability that the top labs will extract all future rents; and that Europe will be “held up” in the classic move in which US players entrench their dominance (initially by competing hard for adoption) and then extract rents from captive customers. His proposed solution (laid out in a number of issues of Silicon Continent, most recently How to avoid being held up by the labs), includes “regulation to keep the model layer competitive, ensuring interoperability between closed-weight and open-weight models” in addition to building a shared open weight model somewhere behind the frontier as “insurance”. Fine with the latter but the former strikes us as hopeless to control the market power of rent extractors: we know the history of digital regulation and interoperability calls in Europe. This “race to adopt” frontier models in exchange for carpeting Europe with compute we do not own or control is what we are resisting. It makes no sense for Europe.
Finally, to the FP authors: moderate language (EuroStack stands apparently for a “grotesque overestimation of European capabilities”) and learn what we are actually saying. It’s good that you describe EuroStack as “the most prominent of the initiatives for greater European independence”. Clearly hyperscalers are worried. You accuse us of “providing a blueprint (for AI sovereignty) that is at the same time not realistic and not ambitious enough” – but we don’t provide a blueprint at all for AI sovereignty. We see it as part of the tech sovereignty piece, bound with cloud. “Its approach will yield a dangerously exposed and ultimately dependent Europe” – OUR APPROACH? You are the ones who are preaching trading access for more dependency, what you call “closer ties with US industry leaders of the current technologies”. Oh I see, this is not dependency. This piece is manipulative gaslighting, but in the end it confirms we are doing something right. The mountain is shaking.
