The constraint moved from chips to power
For two years the AI story was about chips and models. Whoever held the most GPUs and the best frontier model was assumed to hold the advantage. That framing is now out of date. The binding limit in Europe is electricity and the physical grid that delivers it. In Frankfurt, the continent's largest data center hub, these facilities already account for up to 40 percent of the city's total power demand. There are 126 data centers operating there, with a dozen more approved, and the local grid operator has said plainly that large new projects have little chance of connection before the mid-2030s.
This is not a single-city problem. Data centers consumed roughly 4 percent of Germany's electricity in 2024 and are projected to reach around 10 percent by 2037. Across the main European hubs, known as the FLAP-D markets of Frankfurt, London, Amsterdam, Paris, and Dublin, the wait for a large grid connection now runs 7 to 10 years, and up to 13 years in the most congested locations. Ireland has a de facto moratorium on new Dublin data centers until 2028. The Netherlands and Germany have effectively closed the door on new large grid connections until at least 2030. Some operators have already paused planned investments in markets where power is too scarce or too expensive.
Why this is a strategy problem, not a facilities problem
A grid connection queue of 7 to 10 years is longer than most corporate planning horizons. If your AI roadmap quietly assumes that compute capacity will simply be available when you need it, that assumption is now exposed. Capacity, physical location, and the terms of your power supply increasingly decide what you can actually run, and when. This is no longer a question for the facilities team. It is a question for whoever owns the AI strategy.
The cost picture is changing too. From January 2027, German data centers must cover 100 percent of their electricity demand with renewables, after a 50 percent threshold that has applied since January 2024 under the Energy Efficiency Act. Energy procurement becomes a compliance and cost question rather than an operational footnote. The comfortable assumption that compute is a cheap, elastic token hides a physical electricity bill that someone has to pay, and that bill is rising on both sides of the Atlantic.
What disciplined companies are doing now
The companies handling this well have stopped treating power as someone else's problem. They ask where their AI workloads physically run, what the grid and energy situation is in those locations, and how exposed their providers are to connection queues and price spikes. They favor providers with secured power and a clear renewable sourcing plan, and they read the energy terms of a contract as carefully as the price per token.
They also right-size. Not every workload needs frontier compute, and efficient models used selectively cut both the energy bill and the exposure. They plan capacity on the real timeline of the grid, not the marketing timeline of the latest model launch. And they watch the politics, because regulatory pressure is rising. The United States Congress is debating a Ratepayer Protection Act that would force data center builders to pay for grid upgrades, several states are weighing moratoriums, and public opposition to new sites is growing. Energy is becoming the contested center of the AI build-out, not a detail at the edge of it.
Read next: Your Cloud Dependency Is Now a Regulated Risk. Most Companies Cannot Even See Theirs. · Why Is AI Raising the Price of Hardware You Never Bought?