Under the hood of data center power demand

Note first what this episode deliberately does not do. Kann opens by naming the loud argument, a recent EPRI report whose high end has data centers at about 10% of US electricity by 2030 against skeptics who expect efficiency to erase most of that growth, and then sets it aside: he posits that demand will be a lot, and a lot more than historically, and says the question he wants is how it gets met. Nothing here is evidence about the size of the load; it is a conversation about mechanisms, conducted on a premise. What was known in mid-2024 was mostly utility interconnection queues, and Janous discounts those immediately. Queues carry zombie requests the same way generation queues do, so the headline figures, roughly 80 or 90 gigawatts at AEP depending on which speaker you take, are not forecasts. But the discounted number is still large: he puts real demand inside AEP’s system above 10 gigawatts almost certainly, against a total US data center fleet Kann recalls at around 20 gigawatts, which he flags may be a year or two stale. The direction of revision tells more than the level. Roughly 4 gigawatts of data center demand were anticipated for AEP by 2030 as of January; by May the figure had been revised to nearer 15. AEP draws this load for a physical reason: it runs the country’s highest transmission voltage at 765 kilovolts, and announced hyperscaler sites trace that line, because gigawatt-scale campuses need a point on the grid that can already support them.

Why campuses rather than fleets is the pivot. A cloud region is a cluster of data centers within a latency envelope, and a region at the end of the last decade meant a few hundred megawatts. AI training changed the unit of demand because a training run cannot be spread across regions, a constraint Janous flags as outside his expertise, so the requests became gigawatt-scale single sites. That same batch-process character is what makes people assume training is curtailable, and Janous thinks the assumption is overblown. Training is indeed a batch process and inference is not, but the binding constraint is economics rather than physics: the servers cost so much that utilization has to stay high, in the same way an electrolyzer or a nuclear plant does, so nobody will attach a training cluster to a wind farm and run it at a 35% capacity factor. What is plausible is avoiding significant contributions to system peak for a few hours a year, not following the sun.

That sounds like a small prize until you look from the utility’s side, which is where the episode turns. Ten gigawatts of firm round-the-clock load means ten gigawatts of new peak capacity plus a reserve margin on top, and roughly seven years to build it. Shaving peak can also dodge expensive afternoon power, but the bigger question for the data center is whether it is willing to wait seven years. A few hours a year of flexibility becomes the price of getting online. From there the options sort into three. Going off-grid is the one Janous finds least interesting, because it mostly means building a gas plant, which trades the power grid for the gas grid along with its constraints and its questionable firmness; solar and storage will not serve a site of 500 megawatts to a gigawatt, and while he is bullish on nuclear it will not deliver inside the two-to-four-year window these companies are working to. The second is the assets already on site. A data center, he says, is a power plant and a storage plant that happens to have a room of servers next to it, and those diesel generators and uninterruptible power supply batteries can be made to talk to the grid. Microsoft built a gas turbine project in Cheyenne in 2016 and a grid-interactive UPS in Dublin some years later, but the practice has normalized only in Dublin, and only because the grid operator was headed for a moratorium and was persuaded instead to require dispatchability in exchange for a connection. The third is the utility’s own side of the meter, where grid-enhancing technologies free up existing capacity and cut how much flexibility the customer has to supply.

Two things stay unresolved. The utility ask cuts against the rate-base incentive, since utilities have historically made money by building; Janous’s answer is competitive rather than regulatory, that they want this load and its economic development badly enough to move fast, and that squeezing the existing system lets them say yes before their peers do. When Kann presses on fairness, whether other ratepayers end up subsidizing this and whether other projects get crowded out, the answer moves to why utilities are motivated rather than to how the cost is allocated, and the tariff question is left open. On decarbonization Janous is more direct: the headwinds since the 2030 commitments were written, COVID, supply chain disruption, interconnection queues and AI, mean those targets are probably not achievable on the timelines set, though he notes that none of the companies has said so. Efficiency does not rescue them either, on his reasoning, because the gains get spent rather than banked. If a company secures a site that can deliver three gigawatts and chips then double in efficiency, it does not build a 1.5 gigawatt data center; it builds a three gigawatt one that is twice as powerful, because the competition is over the size of the model. Which leaves peak contribution as the most useful lever, since it is peak that pushes utilities to add fossil generation and extend the lives of old plants.