Power grid operators and datacenter developers in the United States are in a bind, and energy analysts can't see an easy way out. The American power grid is old, outdated, and in desperate need of upgrades. Add in a growing number of gigawatt-scale AI datacenters demanding stable access to power that doesn’t disrupt service-level uptime agreements and things start to look even worse. Energy costs have already skyrocketed in the nation's largest energy market mainly thanks to the bit barn bonanza, leading to a new chorus of calls for datacenters to bring their own power generation if they want reliable supplies not bound by grid constraints.
According to energy analysts at Wood Mackenzie, datacenter operators have a choice to make, and neither option is great.
They can wait the five to 10 years it’ll take for grid operators to upgrade their transmission and generation capabilities to account for their demands. Or they can accept deals with power companies that supply them with power but require curtailment during peak loads, and then install their own on-site power generation to make up the difference. This is the far riskier option, but it’s one that many operators are going with. “With more than 90 GW of collocated generation in US interconnection pipelines, it is clear that the need to scale up at speed has sent many data centre developers down the riskier path,” WoodMac explained in the report. “Collocating volatile AI workloads with power generation has scarce precedent, though, and is far more difficult than most in the industry understand.” What the grid wants, the grid getsMany datacenter operators are only thinking about generator megawatts, say the analysts, leaving engineers frustrated at having to explain the technical complexities of building colocated power generators. Multiple grid operators have passed rules that, as key stakeholders understand them, could give utilities priority rights over colocated power generators during shortages, potentially forcing datacenters to reduce demand while supplying power back to the grid, WoodMac said.“This effectively makes the model unworkable,” the analysts said. “Few data centre developers would invest in baseload generation if it could not be utilised when it was most needed.”Near-instantaneous swings in AI power demand can damage reciprocating engines and gas turbines, while batteries may not respond fast enough to every spike and can degrade over time. The rapidly fluctuating power demanded by hyperscale AI datacenters could even damage the grid itself.“These loads can also cause sub-synchronous oscillations, which pose fundamental stability risk to not only local generators but also to distant ones on the transmission system,” the analysts note. “Technology providers are only beginning to come to terms with this challenge, the mitigation of which is site specific, making solutions hard to scale.”
