The data center of the past treated electricity as a tap you turned on. The data center of the next few years is going to treat it as a supply chain it has to secure far in advance. The biggest shift is that operators are now buying power the way they buy silicon, with contracts signed years ahead.
No single change sums it up. What is happening is a combination of new sourcing and new physics, both aimed at the same target. Get reliable low-carbon power into the building, then do not waste it once it is there.
Buying power the way you buy chips
The old method of meeting clean energy goals was to buy renewable energy credits and call it a day. That is falling away. Operators want physical power, not paper, and they want it to match when their workloads actually run.
That is the driver behind the wave of long-term contracts between data center operators and nuclear plants. Nuclear runs at high capacity around the clock, which is exactly what AI workloads want. A single large plant can supply tens of megawatts to a hyperscale campus, and it produces no carbon to offset. In places where the grid cannot deliver, it is the most reliable way to lock in supply.
Renewables are part of the same deal, but with an important wrinkle. Solar and wind are cheaper per megawatt-hour than anything else, but they are not available around the clock. So the modern contract is increasingly a bundle, storage plus renewables plus a firm source, shaped to deliver when the data center needs it. A data center that can shift some workloads to match solar hours gets cheaper and greener power.
Cooling was the sanded-off edge of the energy bill
For a long time cooling was treated as a utility problem and a cost center. It accounted for a large slice of a data center’s total power draw, and nobody wanted to think too hard about it. That has changed, partly because the density of AI racks makes air cooling physically insufficient.
New chips cannot be air-cooled at full load. The heat is concentrated too tightly. So the industry has moved toward liquid cooling, direct-to-chip and immersion designs. Liquid is far more efficient at moving heat away from silicon than air. This cuts the fan energy that air-cooled facilities burn, and it lets racks run denser, meaning more compute in the same floorspace. Operators also chase a lower PUE, the ratio of total facility power to the power the IT equipment actually uses. A facility sitting near 1.5 wastes a third of its power on overhead. Liquid-cooled designs push that number well below 1.2, which turns into real watts saved across a big campus.
The savings are not just on the cooling side. Heat that used to go up a chimney can now be reused. Some facilities pipe waste heat into district heating systems or nearby buildings, which recovers a portion of the energy that would otherwise be lost. The people planning the cooling of an AI facility are now part of the same conversation as the people buying the power, and that was not true two years ago.
Availability is the prize
The force behind all of this is the same one. Every operator without a guaranteed supply faces a future where capacity is rationed. The ones who signed firm supply and built efficient cooling do not. They get to run more workloads, and they get to run them more cheaply.
The result is a fundamental change in how data centers are planned. A site is judged by the strength of its power contract and the efficiency of its cooling, both of which are now strategic decisions made years in advance, not afterthoughts. Power procurement has moved from a back-office function to the center of the conversation about where the next generation of computing actually gets built. Learn more about the International Energy Agency.