Nvidia and Oracle study shows AI data centers can cut power by 30% during grid stress

A series of field trials led by Emerald AI, in partnership with Nvidia, Oracle Cloud Infrastructure, Salt River Project, and the Electric Power Research Institute, found that AI data centers can dynamically reduce their power consumption by more than 30% during periods of grid stress. The trick: scheduling and shifting workloads based on real-time signals from the electrical grid.

Two trials, two continents, one takeaway

The first trial took place on May 3, 2025, at an Oracle Cloud data center in Phoenix, Arizona. A cluster of 256 Nvidia GPUs achieved a 25% reduction in power consumption, sustained over three hours. The reduction didn’t come at the cost of performance. AI workloads continued running without compromising service quality or throughput.

A follow-up trial in December 2025 pushed the results further. Conducted at a Nebius data center in London with 96 Nvidia Blackwell Ultra GPUs, the second test achieved a load reduction exceeding 30%, reaching approximately 35% at its peak. The response time was striking: power draw decreased within just 30 seconds of receiving a grid signal from National Grid. That reduction was then sustained for up to 10 hours.

The software orchestrating this flexibility is called Emerald Conductor, a platform developed by Emerald AI. It dynamically shifts or pauses tasks across GPU clusters without human intervention.

From fixed load to grid shock absorber

The concept being tested falls under EPRI’s DCFlex initiative, which explores how data centers can function as dynamic grid resources rather than static energy sinks. Salt River Project, the utility serving the Phoenix area, participated in the Arizona trial, while National Grid was involved in the UK test.

According to Duke University estimates cited in the research, even modest flexibility from data centers could accommodate roughly 100 GW of additional data center load on the US grid, without constructing a single new power plant.

What this means for the market

If data centers can reduce peak power consumption by 25-35% on demand, utilities and grid operators gain a new tool for managing peak demand. Instead of firing up expensive peaker plants or risking brownouts during heat waves, they can send a signal to participating data centers and get near-instant relief.

Data center operators themselves stand to benefit from demand response programs, which typically pay participants for reducing load during stress events. If you can cut power by a third without affecting your customers, you’ve just created a revenue stream from what used to be pure cost.

The risk to watch is whether these results hold at scale. A 256-GPU cluster in Phoenix and a 96-GPU cluster in London are meaningful proof points, but the hyperscale data centers being planned by Microsoft, Google, and Amazon involve tens of thousands of GPUs. Whether workload orchestration can achieve similar flexibility ratios at that scale, without degrading the AI services customers are paying premium prices for, remains an open question.

Disclosure: This article was edited by Editorial Team. For more information on how we create and review content, see our Editorial Policy.