The potential for flexible data centers | Catalyst

The study started from a contradiction. Utility regulators in the southeast told Norris and his co-authors in the fall that they had been told data centers are 100% inflexible, so they were planning for all of it as firm load. At the same time the Secretary of Energy Advisory Board had issued recommendations on data center flexibility and the Electric Power Research Institute had launched a flexibility initiative. Both cannot be right. The question matters because the power system is sized for rare extremes, heat waves and polar vortex cold snaps, so much of it sits idle most of the time. Across the 22 largest balancing authorities, covering roughly 95% of US load, the average load factor is 53%, meaning average consumption is about half of peak consumption. In 90% of hours more than 30% of the system goes unused. If a new load could be added that simply never showed up at those extremes, the question is how much of that idle capacity it could occupy.

The method is what makes the answer defensible, and it is also where the conditions live. Norris calibrated to each balancing authority’s realized peaks over the past nine years and required that new load never push a realized peak above them, which deliberately leaves unused the reserve margin, the cushion of spare generating capacity a system carries above its peak. That choice is conservative, since the numbers would be larger if reserve margin counted; it also means the added load does not lean on the dirtiest, least efficient units. The team originally planned to model one to five percent curtailment, one to two percent being roughly what existing demand response programs ask for. The results came back so large they ran it down to 0.5% and then 0.25%. The critical thing to understand about those numbers, which Kann first restated as a share of hours before the two of them settled on the right reading, is that they are percentages of the new load’s maximum potential annual energy consumption, not percentages of hours. The model assumes the new data centers otherwise run at 100% utilization constantly. The budget is therefore energy and not hours, and it can be spent as many shallow reductions instead of a few full shutdowns, which means curtailment touches more hours of the year than the headline percentage suggests.

That distinction is what makes the ask plausible operationally. At the 0.25% setting, curtailment of some amount would be required in about 85 hours a year on average, and in 73 of those 85 hours at least half the new load is retained, in 50 of them at least three quarters. This is mostly partial dimming, not switching off. The extreme peaks themselves average two and a half to five hours, and Norris says forecasting has improved enough to see polar vortex events coming up to two weeks out, so at least the cold-snap events are short and can be anticipated. The menu of responses is correspondingly wide: defer or front-load a training run, shift workloads between data centers in different markets, lean on on-site generation or storage, or reduce cooling, which is a large share of data center draw and is conveniently not needed at full power on the winter mornings when the southeast actually peaks. Straight reduction of operations is the last resort and the one operators like least.

The catch is that the study answers the generation question and not the deliverability question. In many places the binding constraint is not whether there is enough power but whether the wires can carry it to that spot under contingency conditions. Asked how much of the 98 gigawatts survives real transmission and distribution limits, Norris said there is no way to know without running it, then under pressure offered roughly 10% erosion as a guess he immediately labeled as just him talking. Kann was audibly skeptical it could be that small. Norris’s counter is that full deliverability is an extremely strict criterion tested against conditions that are genuinely rare, and that load and generation can be connected before those upgrades are finished, which is how ERCOT, the Texas grid operator, gets generation online faster. The second catch is institutional rather than physical. This is not simply enrolling data centers in demand response, which Norris calls a simplistic reading, because existing participants were planned as firm load and opted in later for economic reasons. Doing this properly means treating the load as flexible in the planning and interconnection study itself, which is a different act by a different set of people. And the market is arranged against it: financing counterparties and owner-operators are used to firm service as the gold standard, and hyperscalers have competitive reasons not to disclose what they can actually do. The lever most likely to break that open, in Norris’s view, is an explicit trade of flexibility for a faster grid connection.