Artificial intelligence infrastructure growth is creating an unexpected economic burden on the US power grid: the need to maintain expensive, slow-ramping generation units on standby during off-peak hours to accommodate rapid datacenter load swings. A newly published research study modeled datacenter behavior on a 5,000-bus electrical grid to quantify this effect, revealing costs that planners had largely overlooked.
The core problem stems from the mismatch between how fast AI workloads can change and how quickly traditional generation sources respond. Modern datacenters can ramp their electricity consumption sharply as compute jobs start or stop. Coal plants, nuclear reactors, and other conventional generators cannot match this speed without inefficiency. To prevent grid instability, operators must pre-position expensive units at partial load, ready to increase output rapidly when datacenters spike their demand—even if those generators sit idle during low periods.
Under peak system conditions paired with fast datacenter ramping, simulations showed that slow generation units reached full capacity 100% of the time, with average marginal system costs rising 8%. More critically, these elevated costs persist during off-peak windows when datacenters are dormant, as utilities maintain expensive generation capacity as insurance against the next ramp.
This hidden cost structure challenges conventional grid planning, which traditionally assumed predictable load patterns. Datacenters break that assumption by introducing highly variable, rapid-response demand that existing generation fleets struggle to serve economically.
The implications are significant for both grid operators and policymakers. As AI infrastructure continues expanding, locating new datacenters requires careful analysis of local grid flexibility. Regions with fast-ramping resources—such as natural gas peakers, battery storage, or demand response—can better absorb datacenter volatility. Conversely, grid systems dominated by slow-ramping coal or nuclear generation may face disproportionate cost increases. Future grid investment must therefore prioritize flexible capacity alongside traditional baseload expansion.
