A Dutch court was asked, earlier this year, to decide whether an electricity grid is full. Not congested at rush hour, not expensive, not politically awkward. Full. In April 2026 a judge ruled that the grid around Haarlemmermeer genuinely has no more room and that TenneT does not have to connect a new data centre immediately. Constraints in that region may hold until around 2035.
That is a strange sentence to read. For anyone who has spent a career treating “the cloud” as an elastic abstraction, it’s the sentence that reorganises the map. Capacity isn’t a slider in a console. It’s a transformer, a substation, and a cable route that somebody has to actually dig.
The queue is not a paperwork problem
TenneT, the Dutch transmission operator, isn’t the villain here. It’s running a physics-bound waiting list. Its own numbers, via a Stibbe summary of the February 2026 parliamentary letter on grid congestion, put the offtake queue at 212 requests totalling 38 GW. Thirty-eight gigawatts of things people would like to plug in and can’t.
Meanwhile TenneT is investing EUR 43 billion in the Dutch onshore and offshore grid between 2025 and 2029, with roughly a thousand projects in flight, and it says openly that expansions are taking longer than planned. Permits, people, supply chains, the physical world. You can’t npm install a substation.
What the physics changed
The awkward bit for our industry is that AI workloads aren’t shaped like the workloads the grid was sized for. Training and heavy inference concentrate power per rack in ways that older web and database tiers didn’t. Racks have moved past what air can reasonably cool, which is why liquid and immersion cooling keep coming up. A data hall stops being a warehouse of hot air and becomes something closer to a plumbing job with silicon in it.
That matters upstream. When you draw more kilowatts per square metre, “we’ll just add another hall” turns into “we need another feeder from a substation that’s already spoken for.” The IEA notes that a typical AI-focused data centre draws as much power as 100,000 households, with the largest under construction drawing twenty times that. That’s not a rounding error on the grid plan.
The government made choices, mostly before this wave
The Netherlands has been making land-use and energy calls that predate the current AI push. Projects of 70MW or over 10 hectares are banned across most of the country outside designated areas, and Amsterdam has paused new data centres until 2035. A democracy is entitled to decide what scarce grid capacity is for. Housing and industry and heat pumps are also in that queue.
It’s worth being honest that most of these rules were written before ChatGPT-scale demand landed. Whether they’re the right rules for the workload we’ve now got is a live question. I don’t think it’s an obvious one.
Neighbours, water, and the counterpoint
Then there’s water, which is where the story gets uncomfortable for the industry’s usual talking points. During a drought year, the Netherlands discovered that a Microsoft data centre had used 84 million litres of drinking water. Equinix has stated that a planned Amsterdam site would need 274 million litres a year initially, rising above a billion litres when complete in 2036, with local campaigners now asking for it to be stopped.
Take that seriously as an argument, not as an obstacle. If you live near the thing and it’s drinking the same water you are, “but our workloads are important” isn’t a persuasive reply.
The genuine counterpoint exists though, and it’s not greenwashing. Recovered data centre heat is feeding district heating schemes, one of which is expected to warm more than 10,000 households. Closed-loop cooling changes the water picture materially. A well-sited, heat-exporting facility on a spare bit of grid is a genuinely different object from a speculative hyperscale pad in Haarlemmermeer.
What actually changes for the engineer choosing a region
The practical bit, for anyone picking where to run a workload: “eu-west-something” is no longer just a legal or latency question. It’s a question about a specific substation. “We’ll scale up later” quietly assumes capacity that may be queued into the 2030s. And the cheapest AI capacity, increasingly, may be the workload you don’t run at all, or that you run somewhere with slack in the grid and a heat buyer next door.
I’d watch two things over the next few years. First, whether heat reuse and flexible or interruptible grid connections stop being clever exceptions and start being the default. Second, whether the queue starts being allocated by what the capacity is for, rather than by who filed the paperwork first. Both are hard. Both are, I think, where this is going.
What I don’t know
I should be plain about this. I’m a software engineer. I’m not a grid engineer, I’m not a hydrologist, and I haven’t sat across a table from anyone at TenneT. Several of the tradeoffs in this piece: how to weigh drinking water against district heat, how to price a gigawatt-year of grid capacity against housing, whether Haarlemmermeer’s constraints really do bind to 2035, are genuinely outside my competence to settle. I find the whole thing worrying and fascinating in roughly equal measure, and I don’t think anyone shouting confidently about it from either side is being straight with you.