Somewhere in the high desert of the American Southwest, a datacenter the size of a regional airport is drawing close to a gigawatt of power — roughly what a mid-sized city pulls on a hot afternoon. The operators know exactly when that draw will spike, because the inference workload is predictable down to the minute. What they’re starting to figure out is how to make that predictability useful: feeding it back into the grid, negotiating in real time with utilities, and letting the facility itself behave less like a passive consumer and more like an active participant in energy markets. The datacenter is beginning to think about its own electricity.
This is one of the quieter but more consequential developments in AI infrastructure right now. The scale of the buildout is well-documented — hundreds of billions of dollars committed, gigawatt-class campuses rising faster than utilities can permit new transmission lines, a total projected demand that analysts keep revising upward. What gets less attention is the control problem this creates, and the distinctly AI-flavored solution starting to emerge for it.
The core challenge is that grid operators hate surprises. A facility that suddenly ramps from 400 megawatts to 900 megawatts because a massive training run just kicked off is, from the grid’s perspective, nearly indistinguishable from a catastrophic industrial accident. Frequency drops, neighboring loads fluctuate, spinning reserve margins shrink. The traditional answer is to build massive on-site generation — gas turbines, banks of lithium batteries — and simply decouple from the grid as much as possible. That works, but it’s expensive, carbon-intensive, and wastes the potential value of having a highly predictable, highly controllable load sitting right there on the network.
The smarter approach is to treat the datacenter as a dispatchable resource. Operators like Constellation and NextEra are already in early conversations with hyperscalers about what this looks like in practice. The basic idea: if the grid needs demand to drop in the next four hours, the datacenter reschedules its interruptible workloads — prefill-heavy inference batches, non-urgent fine-tuning jobs, redundant checkpointing — and communicates that flexibility to the independent system operator with enough lead time to actually matter. In exchange, it gets favorable capacity pricing or priority interconnection rights.
The piece that makes this feasible is ML-based workload forecasting running inside the facility’s own operations stack. Systems being developed by several hyperscalers treat the datacenter’s future power draw as a time-series prediction problem: given the queue of pending jobs, the historical patterns of organic inference demand, and the thermal state of the cooling infrastructure, what will the facility need — and when? Forecasting accuracy in the four-to-twelve-hour window is now good enough to make meaningful commitments to grid operators. The same models that predict power draw are also starting to optimize cooling plant scheduling, pre-chilling thermal storage when power is cheap, then drawing down on that stored cold when spot prices spike. It’s demand response, but running on inference instead of human judgment.
The longer arc here is genuinely fascinating. As AI facilities grow large enough to move regional electricity markets on their own, the relationship between the grid and the compute layer inverts. The datacenter stops being something the grid has to accommodate and starts being something the grid can lean on. The computational load becomes a form of grid asset. And the AI workloads running inside — the models, the training jobs, the billions of daily inference requests — become inputs to an energy optimization loop that is itself running on AI.
There is something almost self-referential about it, and also something that points clearly forward. The next generation of gigawatt campuses won’t just consume power at scale. They’ll manage it at scale. The infrastructure that powers AI is about to start getting smarter about its own hunger.