In Part 1, the central question was whether a SEAL 4 hyperscale cloud can be built at all, and whether it is worth it. I ended with a roadmap: six layers, from bottom to top, and on each floor the same two questions: what does the supply chain look like when you build it, and when you operate it?

We start at the bottom, with the layer that is least glamorous and most often skipped in sovereignty debates: the datacenter itself. Buildings, power, cooling and fiber. No software, no chips, but concrete, copper and water. This is probably the layer where European sovereignty is most achievable, although there are obstacles, especially physical ones. The advantage of a datacenter in itself is that it can be nationalized and thereby shielded from possible government interference from abroad. As the case of Nexperia in 2025 shows, that is diplomatically complex.

The anatomy of a hyperscale region

A hyperscale cloud does not think in datacenters, but in regions. A region is not a building but a cluster of buildings, usually divided into three availability zones. Each zone is a self contained collection of datacenters with its own power, cooling and network. An outage in one must not affect the other. The zones are far enough apart not to be taken down by the same flood, fire or power failure, usually from a few kilometers up to about 100 kilometers, but close enough to communicate over fiber with less than about 2 milliseconds round trip latency. Those 2 milliseconds are not an arbitrary number: they are the boundary below which you can replicate synchronously, meaning a database in zone 1 and its copy in zone 2 behave as one system.

Why three, and not two? Because with three you can form a majority, a quorum. If one zone fails, the other two together hold the truth and the service keeps running. Two zones give you redundancy, but no majority: in a network split, neither knows whether it is still in charge. Three is the minimum for a system that keeps itself upright without human intervention. That is also the essence of why a hyperscale region is more expensive than “a datacenter”: you effectively build the capacity one and a half to three times over.

Connection with other regions

One region, however robust, is still one place on the map. To protect against disasters that can affect an entire region, such as a prolonged regional power crisis, a flood, or a human error that spreads across the whole region, you need multiple regions. Unlike between zones, that connection is not synchronous: over hundreds of kilometers the speed of light is simply too slow for those 2 milliseconds, so you replicate asynchronously. You accept that the copy in the other region or regions lags a fraction behind, in exchange for the certainty that your data exists in multiple places.

And this is where the physical layer suddenly touches sovereignty. Hyperscalers offer options to replicate to other regions within the same jurisdiction. For data residency, that is exactly what you want: your failover site falls under the same law as your primary region. At this layer, “data stays in the EU” is therefore not a promise but an architectural choice you can point to on a map. The connections themselves, especially the fiber backbone and submarine cables, are located in Europe for Europe and can therefore, if they are not already owned by a European party, be made European.

Sovereignty, however, is also about resilience, keeping systems available under extreme circumstances. Ukraine has therefore placed many of its systems outside the country, and thus outside its explicit jurisdiction. As Johan Cruijff used to say, “every disadvantage has its advantage.”

Energy: the real bottleneck

Now the bill. A single hyperscale datacenter easily draws tens of megawatts, enough to power hundreds of thousands of households. A region with three zones and multiple buildings per zone therefore quickly reaches several hundred megawatts, and the newest AI regions are heading toward a gigawatt or more. For comparison: that is on the order of a medium sized power plant, permanently, day and night, for one region.

For the Netherlands this is not an abstraction but the bottleneck. Our high voltage grid is at or beyond its limit in large parts of the country. The grid congestion that TenneT had warned about for years is now a reality. New large consumers are in the queue, and that queue is not measured in weeks but in years. In some markets, the connection time for a large consumer has risen to as much as eight years. You can order the fastest chips in the world, but without a grid connection a datacenter can do nothing. Energy is therefore not the most expensive line in the budget, it is the line that determines whether the project happens at all, and when. It also explains why energy is so central in Dutch policy, which I will return to shortly: a hyperscaler competes for exactly the same scarce sustainable electrons as the energy transition of households and industry. At the same time, it is precisely hyperscalers that drive investments in renewable energy, and there are places where they play a role in balancing the power grid by feeding electricity back into it.

What hyperscalers do to save energy

The hyperscalers are painfully aware of that bill, if only because electricity is their largest operating cost. Efficiency is measured with PUE, Power Usage Effectiveness: the ratio between total power and the power that actually goes to the servers. An average enterprise datacenter is around 1.5 to 2.0. Over the past two decades, the hyperscalers have made all kinds of improvements to bring this down to 1.1 to 1.2. That difference, the “overhead” that goes not to computing but to cooling and losses, is exactly where the savings lie, and where a small player loses out to scale.

They achieve this in the usual ways: cooling with outside air, free cooling, when the climate allows it, which is one reason why the Netherlands is an interesting place for a datacenter; hardware that can tolerate higher temperatures so less cooling is needed; and AI that optimizes cooling in real time. Here, scale again works in their favor. An improvement of 0.01 in PUE delivers enormous savings at the scale of hundreds of megawatts.

Water and cooling

Cooling is energy’s twin sister, and mainly requires water. Classic datacenter cooling evaporates water to remove heat. That is good for the electricity bill, but it consumes millions of liters per year per location, often ordinary drinking water. Given increasing water scarcity, this is a growing social problem, and the industry is therefore shifting rapidly. Microsoft, for example, began rolling out a closed cooling system at chip level in mid 2024 that makes evaporative water entirely unnecessary. It is filled once during construction and then circulates, saving millions of liters per location each year. Other approaches include air cooling and cooling with non drinkable, rain or recycled water.

And then there is residual heat. Removed heat does not have to simply disappear into the air. Residual heat can be used, for example, to heat greenhouses or homes through a connection to a district heating network. For the latter, however, a datacenter must be close to the city, where spatial pressure and therefore land costs are highest. Here too, advantages and disadvantages have to be weighed against each other.

The supply chain of a datacenter

Back to the two questions from the roadmap: what does the chain look like when building, and when operating?

When building, you need a surprisingly long list of heavy, critical components: transformers and medium voltage switchgear to convert and distribute grid power, emergency generators, usually diesel, for when the grid fails, batteries to bridge the seconds until those generators are running, and cooling systems and pipework. And of course also concrete, steel, copper and fiber.

And this is the bright spot for the sovereignty story. Unlike chips or advanced software, this is a chain that Europe largely controls itself. Transformers and switchgear come from European players such as Siemens Energy, Hitachi Energy, Schneider Electric and ABB. Emergency generators come mainly from Germany’s MTU, and cooling technology also comes mostly from European companies such as STULZ, Munters and Schneider. Concrete and steel are in principle local. The big exception is batteries. The lithium ion cells required come mainly from Asia, CATL, LG, Samsung SDI and Panasonic. Some alternatives exist and European production is emerging, but this remains Europe’s weak point in an otherwise well covered chain.

That will have to be scaled up further to reduce supply problems. An AFCOM study indicates that 94% of datacenter operators face supply problems, with generators, transformers and switchgear as the biggest bottlenecks. Lead times for large emergency generators have risen to between 72 and 104 weeks. There is therefore much to gain here.

When operating, the dependency shifts from things to flows: electricity, diesel fuel for the generators, water, spare parts and the technicians who keep it running. None of those flows is a direct sovereignty problem, although parts of that chain may currently fall outside Europe.

Permits and physical reality

Alongside the technical reality, there is also a process that must be completed to obtain the necessary permits. And that also has a political component.

In February 2022, the Dutch government adopted a preparatory decision that restricted the establishment of hyperscale datacenters throughout the country. The motivation was explicit and reveals exactly the tension from the energy section: hyperscale datacenters place “a disproportionately large claim on the available sustainable energy in relation to their social and or economic added value”. The national government took control and set admission criteria for permits. The best known case, the planned mega site near Zeewolde, became a national dividing line and got stuck. Not a technical problem, but a social and administrative one.

On top of that decision come the familiar Dutch files: the nitrogen crisis that delays almost every major building permit, grid congestion that can postpone a connection for years, spatial pressure in a crowded country, and water extraction that meets local resistance. Each of these files can halt a project by itself. Together they form a permitting process that takes longer and is more uncertain than ordering any long lead time component. For a truly European hyperscaler, this need not be as much of a problem. After all, there are more countries where datacenters can be located. But that brings us back to one of the earlier elements: the network connection. Building datacenters in places without a digital highway is of little use; laying tens of kilometers of cable, over multiple routes for redundancy, also requires permits and the like. Apart from the work that needs to be done.

So, is it worth it?

Back to the main question: can a SEAL 4 datacenter layer be built? On the sovereignty axis, the answer is positive. The building, the power infrastructure, the cooling and the fiber can largely be designed, delivered and operated within the EU. The only real non EU dependency of significance is the battery cell, and even that is not a showstopper.

So the fact that not many sovereign hyperscale datacenters are being built actually has little to do with the datacenter layer in itself. It is the interaction with the other layers plus the political and financial barriers. No one wants to invest billions if there is no predictable return. The American hyperscalers, by contrast, have the money themselves and, because they control all layers, the return is predictable. That does not change the fact that there are all kinds of parties building datacenters, among other things to lease them to hyperscalers. And if those companies are not European, they could be made European in a crisis.