Backed by Nvidia and other top investors, Emerald AI, a startup, has developed software that enables AI data centers to reduce electricity consumption during peak times without compromising performance. In a pilot with Oracle and Nvidia, energy consumption was reduced by 25% in just three hours. By rerouting or delaying noncritical tasks, the platform eases grid stress, supports renewables, and could unlock 100 gigawatts of capacity, potentially accelerating the rollout of AI infrastructure. Read the report:
In many parts of the world, including major technology hubs in the U.S., there’s a years-long wait for AI factories to come online, pending the buildout of new energy infrastructure to power them.
Emerald AI is developing an AI solution that could enable the next generation of data centers to come online sooner by tapping existing energy resources in a more flexible and strategic way.
“Traditionally, the power grid has treated data centers as inflexible — energy system operators assume that a 500-megawatt AI factory will always require access to that full amount of power,” said Varun Sivaram, founder and CEO of Emerald AI. “But in moments of need, when demands on the grid peak and supply is short, the workloads that drive AI factory energy use can now be flexible.”
That flexibility is enabled by the startup’s Emerald Conductor platform, an AI-powered system that acts as a smart mediator between the grid and a data center. In a recent field test in Phoenix, Arizona, the company and its partners demonstrated that its software can reduce the power consumption of AI workloads running on a cluster of 256 NVIDIA GPUs by 25% over three hours during a grid stress event while preserving compute service quality. […]
On May 3, a hot day in Phoenix with high air-conditioning demand, SRP’s system experienced peak demand at 6 p.m. During the test, the data center cluster reduced consumption gradually with a 15-minute ramp down, maintained the 25% power reduction over three hours, then ramped back up without exceeding its original baseline consumption (figure 1).
Figure 1: (Left panel): AI GPU cluster power consumption during SRP grid peak demand on May 3, 2025; (Right panel): Performance of AI jobs by flexibility tier. Flex 1 allows up to 10% average throughput reduction, Flex 2 up to 25% and Flex 3 up to 50% over a six-hour period.
The International Energy Agency projects that electricity demand from data centers globally could more than double by 2030. In light of the anticipated demand on the grid, the state of Texas passed a law that requires data centers to ramp down consumption or disconnect from the grid at utilities’ requests during load shed events.
“In such situations, if data centers are able to dynamically reduce their energy consumption, they might be able to avoid getting kicked off the power supply entirely,” Sivaram said. Looking ahead, Emerald AI is expanding its technology trials in Arizona and beyond — and it plans to continue working with NVIDIA to test its technology on AI factories.
“We can make data centers controllable while assuring acceptable AI performance,” Sivaram said.
“AI factories can flex when the grid is tight — and sprint when users need them to.”
Read more here.
