US regulators and grid operators are calling on data centers to cut their power use when the grid is overloaded, saying it is the only way to connect large-load customers as quickly as they are requesting.
Officials and industry executives in Houston for an energy conference last week agreed the grid can not be expanded fast enough to meet unprecedented load-growth forecasts if each data center must be served its full allocation of power every hour of the year. But panel discussions and interviews showed a gap between the flexibility regulators say is possible and what data-center operators say they can deliver right now.
"Finding a way to manage this demand in a flexible way is how we are going to get through this transition," PJM chief operating officer Stu Bresler said at CERAWeek by S&P Global in Houston, Texas. Without a way to curtail load during tight conditions, large new customers could face wait times stretching over a decade to connect on to the grid, he said.
Although data centers request power connections sized for their maximum load, they do not necessarily operate at that level around-the-clock. Meanwhile, the grid only faces severe strain for a few dozen hours a year, usually during extreme weather events, a window that the Electric Power Research Institute (EPRI) says amounts to less than 1pc of the time. Regulators argue that if data centers agreed to reduce their draw during those peak hours for the grid, the system could absorb far more demand without immediate infrastructure expansion. Officials pointed to a study by Duke University showing that if customers in PJM, the nation's largest grid operator, cut load by half a percentage point for roughly two days, 18GW of new load could be absorbed without adding physical infrastructure to the system.
Those cuts will not materialize without clear financial incentives, said the conference attendees. One way to incentivize such behavior is to pay large load customers who agree to use less power at the grid's request, the same way generators are compensated for supply, said Federal Energy Regulatory Commission (FERC) commissioner Judy Chang.
"We really should focus on the demand side and include large data centers' ability to curtail," said Chang. "We're not quite getting the incentives yet right."
Switching off
Some artificial intelligence (AI)-training workloads can be paused or rescheduled but that flexibility varies by site and depends on permits, on-site equipment and the type of computing being performed, said Google's global head of data center energy, Amanda Peterson Corio. "There are many flexibility options, but not every data center can respond the same way," she said.
AI-focused "compute factories" tend to be more adaptable because training jobs can be paused, while enterprise data-center workloads are far less flexible, said EPRI vice president of strategy David Porter. The group's work includes launching Flex Mosaic, a classification system meant to help utilities and data-center operators define what each site can realistically offer.
Large customers who can will respond if markets put a value on reducing load, but the industry still needs clear rules for how much load can be cut and under what price signals, said Peterson Corio.
Some data-center operators pushed back against the idea that most facilities can deliver the type of reductions regulators are envisioning. The vast majority of Amazon Work Spaces' (AWS) sites cannot lower their load on demand because the cloud and enterprise workloads they host cannot be interrupted without affecting customers, said AWS vice president Kerry Person.
The only practical way AWS and others like it can reduce their draw on the grid is by switching to on-site backup generators, the majority of which burn fossil fuels because battery storage technology has not yet reached the scale of demand. US rules prohibit that approach, however, because environmental permits limit backup generators to emergency use only. "That's a regulatory issue," Person said.
For now, most large-scale examples of this behavior remain confined to pilot programs. EPRI's DC Flex initiative is running field demonstrations at operating data centers to test how much load can be reduced without harming performance and whether facilities can provide services such as frequency support.
If the industry cannot agree on workable rules for cutting load during peak hours, large users will continue to pursue their own behind-the-meter generation rather than wait years for a grid connection — a shift executives warned would raise costs, lock in more fossil-fuel use and leave the shared power system worse off.
