What if you bring the datacenter to the power?

Five ways to do it, and why one makes the most sense

Paul Kunneman, co-founder Project Enki | August 2026

The logic of energy transport has been the same for more than a century. You generate power where you can, and you carry it across the high-voltage grid to the places where people live and work.

The Netherlands has done this for over a hundred years and it works. Almost every user of the grid fits that model. Factories, offices, and households sit where the people are, and the power comes to them.

Datacenters are the first major power consumer for which that assumption begins to shift. They use a great deal of energy, and part of what they do is insensitive to distance. For most work the location still matters, if only because a datacenter has to be built and operated. Part of it does not. Training is the clearest case. A training run often takes weeks, and whether the data spends one millisecond or twenty milliseconds in transit makes no difference to the result. It is not the only case. Fine-tuning, batch inference, and large scientific computation are indifferent to distance in the same way. This is the compute you can, in principle, place where the power is rather than where the user sits.

This opens a question the industry has left unaddressed for a long time. What if you bring the compute to the energy instead of the energy to the compute?

The idea is not new. Hydropower plants in Norway were already attracting datacenters in the 1990s. Iceland has marketed its geothermal energy as cooling water and cheap power for server farms. But the question is becoming more urgent now that power grids are filling up, permits take years, and AI is driving demand for compute to levels no one foresaw five years ago. At the end of 2025, more than 15,000 companies in the Netherlands were on the waiting list for a new or heavier grid connection. In Germany, lead times for some projects run to almost ten years. At the same time, wind energy in the North Sea is regularly curtailed because the grid cannot absorb the power.

There are five serious ways to bring compute to energy. Each deserves an honest look.

Hydropower

The oldest answer to the question. Norway, Iceland, the Pacific Northwest of the United States, Canada. Hydropower is stable, predictable, and cheap. Meta's datacenter in Luleå runs largely on Swedish hydropower, combined with cold outside air.

The problem is scale and geography. The best hydropower sites are already taken or nearly full. Norway now has the same connection queues it always presented as the alternative. And hydropower sits where there are rivers and elevation differences, not where the large AI demand comes from. Transport remains a bottleneck. For Europe as a whole, it solves little.

Geothermal

Iceland is the textbook example. Energy from the earth, widely available, no transport needed. It works. Iceland has built an attractive datacenter market on precisely this condition.

But geothermal is geographically even more limited than hydropower. The places with enough geothermal potential at the right depth are scarce. Outside Iceland, Kenya, New Zealand, and a handful of other locations, it is not a serious option at scale. For European AI compute demand, it solves nothing.

Solar

The most-discussed option of the moment. Microsoft, Meta, and Google are building or planning datacenters in sun-rich regions such as Spain and the Gulf states, directly beside large solar parks. The logic holds. Solar power you cannot feed onto the grid anyway is converted directly into compute.

There are two objections. The sun does not shine around the clock, but for AI training that is less fatal than for other workloads, because a training run can pause and resume. The real problem is the utilization of your expensive hardware. The capacity factor of solar power in Europe averages between 11 and 17 percent. Over a year, a solar park therefore delivers only a small share of its maximum output, with many hours at almost nothing. The GPUs in an AI datacenter are the most expensive component and only pay for themselves if they run almost continuously. Tie them directly to solar and they sit idle most of the time. Compensating for that with battery storage is technically possible, but at the scale required the business case breaks down.

Then there is the heat. The locations with the most solar potential are also the hottest. A datacenter in the Sahara or on the Arabian Gulf has to cool at an outside temperature of 45 degrees. You gain on the energy side and lose on the cooling side. And the best solar parks lie outside the EU, in countries with different legal systems and different geopolitical interests. For anyone who wants to build sovereign European AI infrastructure, that is exactly the problem you set out to solve.

Nuclear

The energy sector's rediscovery. Small modular reactors (SMRs) are presented as the solution for energy-hungry datacenters in remote locations. Microsoft has signed a deal with Constellation Energy for nuclear power for its Azure datacenters, and Google did the same with Kairos Power.

The reality is that in 2025 these small reactors are barely operational anywhere. The first commercial projects are scheduled for 2030 or later. As a technology it is promising. As an answer to an urgent problem it does not fit yet, there is no rapid scaling pathway for SMRs as of now.

Offshore wind

And then the fifth option, which at first glance seems the least obvious. A datacenter at sea, next to a wind park. No solid ground. Waves of salt water..

Look at what is actually happening. According to Aurora Energy Research, Europe curtailed some 72 TWh of renewable power in 2024, a considerable share of it wind. Narrower counts come out lower, depending on the definition, but the pattern is the same. Green power is being thrown away on a structural basis because the grid is full. In the North Sea, wind parks are shut down (curtailed) at times because there is no room on the high-voltage grid. That energy is not harvested anymore once the turbine stops. Wind does not wait.

At the same time, Europe faces a paradox. Demand for AI compute is growing fast, but there is nowhere to build new datacenters and no power to run them. Cities such as Amsterdam, Dublin, and Frankfurt have restricted or frozen the construction of new datacenters. The grid is full, cooling water is scarce, and the permits do not come. Two industries that need each other but cannot find each other.

Offshore wind has a property the other four options lack. The energy is generated in precisely the locations where you can place a datacenter without bothering anyone. No local residents. No drinking-water use, because seawater cools the servers. No permitting battle on land. And a large part of the energy the datacenter consumes would otherwise have been curtailed. In those hours the turbine would have been shut down.

The technology is present in a way the other options still lack. Offshore platforms have been built for decades for oil, gas, and wind. Modular datacenters you can ship already exist. Subsea cables already connect platforms to the mainland. Seawater cooling is not an experiment. Google has cooled its datacenter in Hamina with water from the Gulf of Finland since 2011. The only thing that is new is combining all these parts in one location.

This is also precisely the direction we are exploring at Project Enki. Together with Vattenfall and ABB, we are examining the feasibility of an offshore AI datacenter in the North Sea, connected to existing wind parks. The project is in an exploratory phase.

The principle is simple. In some hours a wind park produces power the grid cannot receive. A datacenter next to the turbine takes that power. The turbines keep running, energy that would otherwise have been curtailed goes to AI workloads, and the data travels back to the customer on the mainland through a subsea cable.

What this changes

What makes this interesting goes beyond the technology alone. It changes the economic logic of offshore wind.

Wind parks that are curtailed earn nothing in the curtailed hours. That is a loss for the operator, for the investor, and for the energy transition. A datacenter next to the turbine changes that. It gives existing wind parks a second revenue stream in the hours when the grid is full. It makes new wind parks easier to finance, because the business case depends less on grid capacity. And it makes the energy transition itself more stable.

For the AI sector it is the mirror image. No queue, no drinking water, no permitting battle, no objections from local residents. Project Enki focuses on the compute that does not need to sit next to the user. Training first of all, but also fine-tuning, batch inference, and other work that runs in the background rather than in real time. What these share is that they are hungry for power and indifferent to a few extra milliseconds of distance. Real-time inference, gaming, and trading need a server close to the user. This kind of work does not. It needs power, as cheap and as green as possible.

And much of this work can pause and resume. Training is the clearest case, because a run can checkpoint and pick up where it left off. That is why the variability of wind is manageable here in a way it would not be for real-time work. This is compute that does not punish you for running on a source that rises and falls.

The cost of idle hardware

There is a catch under all of this, and it is financial rather than technical. In an AI datacenter the GPUs are the part you cannot afford to leave idle. They are expensive, they lose value whether they run or not, and they have to be paid for in full regardless of how often they compute. The business case does not turn on the price of power. It turns on how much of the time those GPUs are working.

This is the real reason cheap energy on its own is not enough. A cluster that computes thirty percent of the time is not a third of a business. It is close to no business at all, because the hardware costs the same either way. Free power on an idle GPU still loses money. Utilization is the constraint that decides whether a location works, and it quietly disqualifies most of the cheap-power options above. Solar fails here first. So does any design that ties expensive hardware to an intermittent source and hopes for the best.

Offshore wind clears this bar, but not on its capacity factor alone. A North Sea wind park delivers around 40 to 50 percent of its rated output averaged over the year, far more than solar and still short of what expensive hardware needs. The number that matters is not that average. It is how often the wind sits above the level the datacenter is built for. If you build the datacenter to match the wind park's full peak, it runs at full output only when the wind is strong, which is rare. If you build it to a fraction of that peak, a quarter or a third, the wind clears that lower level almost all of the time. A wind park reaches full output only now and then. It delivers a base level nearly always. You give up the peaks, which the datacenter could not have absorbed anyway, and in return the power you built for is there whenever you need it.

That still leaves the hours when the wind is genuinely low. Here the honest answer is that the datacenter draws power from land. This sounds like a contradiction and it is not. The grid supplies a top-up, not the base load. The bulk of the energy remains curtailed wind that would otherwise have been thrown away. The top-up covers a minority of hours, and it runs over the wind park's own cable to shore, which sits idle in exactly those hours because the turbines are barely producing. No new connection is needed and no multi-year place in the queue. A load that leans on the grid only when it has room to spare is the flexible demand a network operator is actively asking for.

There is one honest cost, and it is timing. A long, still, cold spell is exactly when power across the market is scarcest and most expensive, so the top-up in those hours is not cheap. But it is a small share of the year, and paying for it protects the utilization that makes the whole cluster viable. The result is a wind-majority site, sized on purpose, with a modest top-up from land for the hours the wind cannot cover. That is what keeps the GPUs working often enough to make the economics real, while most of the energy stays green, cheap, and otherwise wasted.

There is also a broader argument. Europe has decades of expertise in offshore engineering and one of the strongest maritime industries in the world. It has the wind parks. It has the knowledge. It has a growing market of companies and governments deliberately looking for infrastructure that meets European standards, under European law, on European soil.

Conclusion

The other four ways to bring compute to energy are each bounded. By geography, by network, by timing. Offshore wind in the North Sea combines what the others each lack on their own. Surplus that has nowhere to go. A location where no one is in the way. Technology that already exists. And a market that is waiting.

The question was: what if you bring the datacenter to the power?
In the North Sea, the answer is already waiting.

Paul Kunneman is co-founder of Project Enki, a Dutch company building AI data centers at sea, directly next to offshore wind farms. Project Enki works with Vattenfall and ABB on Europe's first offshore AI infrastructure.

projectenki.com | contact@projectenki.com