AI Data Centres Can Shed Up to 55% of Load Without Disrupting Services, EPRI Research Finds
New modeling challenges the assumption that data centres are inflexible grid loads, potentially opening a significant balancing resource for grid operators facing surging AI power demand.
AI training and inference workloads can reduce power draw by between 18% and 55% relative to their average consumption while still meeting quality-of-service requirements, according to new modeling published by the Electric Power Research Institute on Friday (2026-06-26).4
The finding matters for grid operators across the UK and Europe managing an accelerating surge in data centre demand. Power planners have generally treated large data centre facilities as inflexible baseload rather than dispatchable resource — a classification that, if revised, would shift the economics of balancing variable renewable generation.2
EPRI's DCFlex Initiative, which aims to standardise data centre designs and utility flexibility programmes, has already demonstrated reductions reaching 40% in controlled tests. Future AI infrastructure built with flexibility designed in from the outset can expand that range "materially," according to the initiative's emerging technologies executive, Anuja Ratnayake.4
Two utility-scale tests illustrate the practical boundary. At Arizona's Salt River Project, Emerald AI gradually ramped an AI workload up 25% over a three-hour peak period in a peer-reviewed trial. The same company then reduced power consumption 20% at a data centre served by Portland General Electric in Oregon during simulated weather emergency scenarios, with results reported in March. Neither test involved service disruption.4
The distinction between training workloads — which can be paused, throttled or rescheduled — and inference workloads, which serve live requests, matters for how much flexibility can be offered at what notice. Training carries the most headroom. Inference is tighter but not zero. The 18% to 55% range reflects that spread across workload types.4
The demand backdrop makes the research commercially significant. Global data centre electricity consumption reached an estimated 415 terawatt-hours in 2024 and is projected under IEA scenarios to hit 945 TWh by 2030. AI-specific workloads are accelerating that growth: Alex de Vries-Gao of Digiconomist estimated that AI systems already account for around 20% of total data centre power, with that share potentially reaching 49% of consumption by end-2025, excluding cryptocurrency mining.1,3
The IEA's 2026 update on energy and AI estimated that by 2027, an individual server rack in an advanced data centre could carry peak power demand equivalent to 65 households. At campus scale, a single large facility can draw several hundred megawatts.3
For UK and European grid operators weighing how to handle this demand class, the implication is that data centre flexibility programmes — if standardised and contracted — could function as a balancing resource rather than pure demand increment. UK power markets are integrating growing volumes of variable renewable generation; large industrial loads that can modulate predictably would reduce the cost of that balancing task.2
Whether operators follow through depends on commercial incentive. Data centre operators have historically resisted flexibility commitments out of concern for latency-sensitive contracts and the reputational risk of any service degradation. The EPRI work provides a technical baseline demonstrating that controlled load reduction is achievable without disrupting the workloads that matter most.4
Speed of build complicates the picture. US startup Crusoe Energy has secured 4.5 gigawatts of gas-powered capacity for data centre infrastructure, with OpenAI among potential customers through its Stargate joint venture. Commitments at that scale lock in consumption patterns before flexibility programmes exist to shape them.1
The test over the next 12 to 18 months is whether EPRI's FlexMosaic standardisation effort can move from controlled trials at individual utilities to grid-service programmes that data centre operators will sign. For UK and European grid operators watching the American work, the technical case for data centre flexibility is now established. The commercial case is not.