An old steel mill, rewired and repurposed, could soon do more useful work than a brand-new data centre. That is the bet behind Polarise, an AI infrastructure firm that has just entered the race for a slice of Europe’s data-centre boom. Its CEO, Michel Boutouil, told EU Perspectives why Europe should stop copying America’s giant AI factories and build smaller, smarter ones instead.
The European High Performance Computing Joint Undertaking (EuroHPC) AI Gigafactory tender aims to select consortia to build and operate up to seven vast sovereign compute sites. It is backed by as much as €10bn in EU and national funding and meant to mobilise at least €20bn in private investment. The initiative first envisioned single, centralised mega-sites. Each would run at well over 100 megawatts of IT load, packing tens of thousands of top-end AI processors into one location. The tender now allows a Gigafactory to be spread across several sites, or even across borders, as a distributed network of connected facilities. That pivot is what motivated Polarise to join.
Polarise is a Berlin-founded, now Düsseldorf-based AI infrastructure company with a plan to build “sovereign European AI factories.” It rejects the greenfield hyperscale playbook, slotting modular compute into existing buildings and reserving power-hungry training clusters for regions with abundant renewable energy. The company already runs compute out of Oslo and Munich, with a 65-megawatt site planned in Amberg, Bavaria.
EU Perspectives put questions to Boutouil on scale, sovereignty, the climate math of the build-out, and why he believes capital is the real hurdle for Europe’s AI future.
Why do we need to build more data centres?
AI will have a big impact and change a lot of things in the world, especially the way we work. It’s a tool, and this tool requires compute capacity, and compute capacity requires data centres. At the moment, if it scales the way it’s scaling, we don’t have enough infrastructure. We need more. But we have to be cautious about how we do it. It’s important not just to scale data centres, but to do it in a way that’s sustainable, that brings value rather than just consuming resources.
How do you see the European AI strategy at large? Should Europe be trying to compete on equal footing with the US and China, aim for strategic interdependence, or something else?
First of all, the whole EU tender is more like a drop in the ocean. If you look at where we came from, the first statement from von der Leyen versus what actually went into the tender, there’s a big difference. To be honest, it doesn’t make much difference compared to what existed before. But it’s a start, a good point from which to discuss different options.
I don’t think we’ll compete in the same way the US or China do, because they have different conditions for that. But it’s a great opportunity for Europe, because we can build something, from Europe, for the world, that’s more sovereign, that carries our values, that is more protected. It gives you a better feeling that you can trust your data to a machine. And at the end of the day it’s a machine that belongs to somebody, fed with data that can be used for good and for bad. As Europeans, we have to do something.
We missed the digitalisation and cloud trains. The cloud sector is already lost. But in AI we still have a chance. — Michel Boutouil, CEO, Polarise
There’s a chance we can leapfrog. We missed the digitalisation and cloud trains. The cloud sector is already lost. But in AI we still have a chance. There’s a misunderstanding in the market about what will actually be required. The last few years have been dominated by foundation models and training clusters, which need really big data centres and a lot of GPUs. But after training there’s AI in operation, inferencing, and that doesn’t require huge clusters. It’s a different form of consumption. And open-source models are getting better; the gap between foundation and open-source models isn’t that big anymore. We shouldn’t just look at what OpenAI has done or what Anthropic is doing. We should look at what else can be done, and how that can be interesting from a European perspective.
Are localised models part of that equation? An area where Europe could have its own place in the conversation?
More so open-source models. If you look at a company like Hugging Face, more of a developer platform, you can fine-tune models, develop models, use Chinese, American or European open-source models. Everyone is building not just foundation models but also open-source models, on Llama and so on, and those can be taken further and serve their own purpose. The world won’t consist only of frontier models from OpenAI or Anthropic; there will be a lot of other things that can be used. You just have to learn how.
How do you reconcile the climate concerns with the data-centre build-out ahead of us?
It depends on how you do it. Looking at the US and China, I see a big problem: they consume a lot of energy, a lot of water, a lot of space, and they don’t produce many jobs compared to industries of the past. But if you’re more creative about what’s actually required, there are a lot of options. You don’t have to build a data centre purely as a consumer of energy. You can build one that transforms that energy into heat, and that heat can warm homes in the community. This requires looking not just at the profit on your IT, but at the profit for the environment and for the people. That’s exactly what we do.
The difference with us is that we don’t build on greenfield sites, from scratch. We bring the data centre to where infrastructure already exists: old factories, old steel mills, which already have an energy connection and a building, and we retrofit them into AI data centres. That saves a lot of CO2, because you upgrade an old structure rather than building a new one. Then we extract the most heat we can from the IT system and transport it to a heat exchanger, where local grid or heat companies can connect and take it. We also look for places with renewable energy nearby, such as wind or solar, and build battery parks so we can capture the most energy across 24 hours, because at night we don’t have sun.
It requires more investment upfront, and a bit more thinking. It’s not like Grok, where you build something in Texas and, because you don’t have a grid, you surround it with diesel generators and let it run. That setup has an awful CO2 balance, and we try to do it better.
The Commission wants to triple data centres across Europe, which will exceed what the grid can support. Where is the ceiling for your approach, and what happens when you near it and the choice becomes “don’t build” or revert to on-site fossil fuels like the Grok instance?
This comes back to my initial point: people are fixated on monster data centres that train large models. Training does require a huge cluster. But you don’t need one in every country. You can build training clusters where the energy is generated. Look at Scandinavia: 99.9 per cent renewable energy, and a lot of it. Training clusters don’t need low latency, because you do everything in one place. So if you think strategically about where a training cluster belongs, it should be up in the north. In the rest of Europe you build smaller clusters for AI in operation, inferencing, and services directly for the customer.
The second point is that we don’t actually know how much AI training will be developed in future, because a lot of frontier models are already extremely capable. But do we need those capabilities for everything? Do we need very high-parameter models for every task? I don’t think so. It’s a special case that needs that kind of super-intelligence. For most tasks, translating an email from English to German, maybe half of current usage, you don’t. This is also about education: not all AI is the same, and not all models are the same. We need to understand how to use it efficiently. That’s what we’re trying to do, not just build a data centre, but the full stack, an inferencing platform.
You mentioned a need to educate the public on the types of models they should use.
Yes. With a platform like ours (we call it Drive), you don’t just give the customer one model or one use case. You give them a blueprint for which model to use for what. You might create an agent that decides which model is best for a given task, because not every task is the same. An agent can understand: this task is easy, give it to a GLM; this one is demanding coding work, it needs an Opus. That will come, and it will make things look very different from today.
I also think the growth rates projected for data centres are partly sales propaganda from the chip producers, because they want to sell their stuff. New chips will come, more efficiencies will come, and a lot of things may not go the way the current market narrative suggests. So yes, we need data centres, but the right ones for the right purposes.
You supported the EU adjusting its tender requirements, which initially called for massive single sites. Is that change what incentivised Polarise to bid? And why has so much of the early private-sector support fallen away?
When the first tender came out, we ran the numbers and found it would require a 350-megawatt data centre. There aren’t enough sites in Europe for that. It wasn’t clear how much would be subsidised or how, or what software stack would sit on top. So we said it made no sense tactically, and we said so publicly: if you look at what would actually be delivered, we don’t see the rationale. We also pointed out the assumptions were based on already-outdated chips. That will always be the case, because a year later, when the tender is delivered, there’s a new chipset, so everything specified beforehand is obsolete.
Then some experts spoke to the Commission, and the revised version went in a direction that’s technically feasible and sensible, because they put a software standard on top, not just the data centre and chips, but how to consume it and reach different customers. Reaching an SMB is different from reaching an AI lab, which knows how to work with bare metal. That made sense to me.
But the downside, and this answers the second part, is that the subsidies aren’t attractive, and that won’t drive the decision. That’s why players who were cheering for it dropped out. The offtake the Commission is bringing is a drop in the ocean; it doesn’t even cover the data-centre cost. We’re talking billions to build the data centre, and billions again for the chips.
My criticism is: why not address the real issue in the AI business, which is capital? Capital drives the whole market. The players with access to it, the CoreWeaves of the world, dominate, because they can raise real risk capital and invest it. We need more backstops from the European Union. I don’t need them to create my offtake; if I have a suitable offering I can sell it, and there is demand. But to finance the infrastructure I need bankable offtake, and that’s not easy. You need a Google, a tier-one customer signing a four- or five-year contract. To deliver to the customers who actually need it, you need backstops. Instead of putting €10bn into offtake, the EU should provide the backstop, so we can raise capital from the market, build for the right customers, and let them consume the right AI. Sovereignty is a big political topic, and customers want it. They understand they need to control their data and the AI they use.
So we changed our mind about the tender because of the technical aspects. On the money side it’s not attractive, but it’s still some money, and it helps with recognition. We’re doing this anyway, so we don’t have to change anything to take part. We can bring what we’re already doing to the tender.
You’re building sovereign European AI, yet part of your stack is American, via Nvidia. How do you define “sovereign”?
I don’t see an issue with that. Some of the discussion around chips is nonsense, because we can’t and won’t catch up on chip design and manufacturing from Taiwan. And it’s not only Nvidia in the chain. You can’t produce any chip without Carl Zeiss, a European company; without Zeiss there’s no chip fabrication. You can’t produce any chip without Taiwan, because the fabs are there. And then America assembles everything, builds the software, designs the chips, and delivers to us.
A Mercedes is still a German car, a sovereign German company, but the parts come from China, from all over. — Michel Boutouil, CEO, Polarise
Look at a car. A Mercedes is still a German car, a sovereign German company, but the parts come from China, from all over. It’s a global world, and that’s fine. For me, sovereignty means control. If I own the IT boards, I buy them from Nvidia. I install them. They’re mine. The software layers come from all over, licensed here and there, but I control it. As long as I control it, and no third party from an outside jurisdiction can force me to shut off a client, I’m safe. That’s the sovereignty we need. Yes, some global deliverables could be restricted, but that’s something we can’t change at the moment.
We could try to build our own chip manufacturing, but it would take enormous time and investment, and European investors are very risk-averse. We saw that with Intel in Germany. So it’s nonsense to fixate on what we can’t change. What we can do is make an offering governed by European legislation and sell it into the worldwide market. American companies are now asking about our offering, because they’re afraid they won’t be able to sell their products into Europe or elsewhere. That’s because of a potential kill switch or spy switch from Trump. That’s why it’s so important not to reduce the conversation to a nonsense point, and to focus on what we can do to avoid being controlled by third parties.
