AI data centers can cut grid power by 35% without disrupting workloads
A landmark power-flexibility trial in London shows how AI infrastructure can stabilize electricity networks and lower consumer energy costs
Artificial intelligence (AI) data centers can flex their power consumption in real time without disrupting computing workloads, a capability that could transform them from grid burdens into stabilizing assets for electricity networks.
A consortium of technology and energy companies has demonstrated for the first time in the UK that an AI data center can respond to grid signals within 30 seconds, cutting its power draw by more than a third and sustaining reduced consumption for hours when renewable generation is low.
“When we get the signal — middle of the night, lightning strikes — we’re able to reduce power within 30 seconds by about 35%,” said Varun Sivaram, chief executive of Emerald AI. “Later this year, we will have taken this technology from lab bench invention to commercialization at a commercial data center in just two years.”
“If you can throw more electrons at a fixed-cost system and you don’t need to build more infrastructure, then the rates come down for everyone else,” said Steve Smith, president of National Grid Partners. “You bring them on, they’re paying towards your fixed costs, and therefore it’s good for rate payers.”
The trial, jointly conducted by Emerald AI, National Grid, Nvidia and Nebius, ran for five days in December 2025 at Nebius’s new data center in London, according to a press release issued on March 2. It tested Emerald Conductor, Emerald AI’s software platform, on a cluster of 96 Nvidia Blackwell Ultra GPUs. National Grid sent more than 200 simulated grid stress events to the site without advance warning, including sudden generation losses, demand spikes and extended periods of low wind.
The system passed every test. It cut power by up to 40% during peak-smoothing events, including half-time surges during major football matches. It shed more than 30% of its load in roughly 30 seconds during a simulated system stress event, and followed load-reduction requests for up to 10 hours, all while protecting critical AI workloads running at full throughput.
The trial’s significance extends beyond a single facility. As the UK prepares for more than 6 gigawatts of data center deployments on the grid by 2030, the companies estimate that rolling out this technology across the sector could add more than 2 GW of flexible capacity back to the grid when needed, easing connection queues and supporting the integration of low-carbon generation.
The announcement is the latest in a series of moves by major AI infrastructure companies to position data centers as grid-friendly assets rather than sources of strain. National Grid and Nvidia are both investors in Emerald AI, which was founded in November 2024.
Later this year, Nvidia and Emerald AI plan to launch the Aurora AI factory in Manassas, Virginia, a 100-megawatt facility set to become the world’s first fully power-flexible AI data center.
Smith said the key to unlocking this flexibility was persuading data center operators that not all workloads are equally urgent. What initially seemed like an insurmountable constraint turned out to be a solvable engineering problem once operators were shown the physics of electric power networks and the costs of inflexibility.
Three levers, one brain
The panel on decarbonizing data and AI’s energy footprint took place at the 12th annual Sustainability Week, organized by Economist Enterprise in London. The discussion was moderated by Vijay Vaitheeswaran, global energy and climate innovation editor of The Economist, and supported by Nebius and National Grid.
Alongside Smith and Sivaram, the panelists were Josh Parker, head of sustainability at Nvidia; Anuja Ratnayake, emerging technologies executive at Electric Power Research Institute (EPRI); and Daria Mukhortova, head of sustainability at Nebius.
Sivaram described three methods Emerald AI uses to achieve power flexibility without compromising service levels:
Pausing or slowing low-priority workloads, such as a fine-tuning run not needed until the following week.
Migrating workloads between data centers. In a prior demonstration with EPRI, Emerald moved a workload from Virginia to Chicago across two Oracle data centers at a latency penalty of just 10 milliseconds, imperceptible to the end user.
Deploying on-site batteries, generators or fuel cells to absorb or release power independently of compute loads.
“There are autonomous AI agents, the secret sauce, that are talking directly to Nvidia GPUs and to the job scheduler, understanding autonomously that they have to meet the lightning strike while protecting the integrity of the workloads,” Sivaram said. “That’s an optimization problem no human can solve, but AI can.”
Nvidia’s interest in grid flexibility goes beyond its investment in Emerald AI. Flexible data centers can be connected to the grid faster, a real competitive advantage in a supply-constrained energy market. Google’s experience in Michigan and other Midwestern states showed that offering demand flexibility can secure faster grid connection approvals. Nvidia has also developed templates to help its customers adopt flexibility practices.
“Load growth historically hasn’t been a bad thing in terms of economic development, but also in terms of sustainability, especially when we’re on the threshold of clean energy being more economical than traditional fossil fuels,” Parker said.
Flexible AI demand can absorb surplus renewable generation when it is available, raising the value of variable energy sources and making data centers a critical component of virtual power plant ecosystems, where AI helps match variable supply with variable demand.
AI’s peak load problem
Ratnayake said EPRI’s engagement on AI data center flexibility stems from a near-universal pain point among its 500-plus members across 45 countries. The traditional timeline for connecting a large new industrial load to the grid is seven to ten years, but data centers need access within two to three years.
“Almost every one of our members is seeing a pain point right now from the growth of large loads that is making them pause and go: how can we do better?” she said. “We think of it as creating a common language between these two major industries. When somebody says ‘be flexible,’ we know what flexibility means.”
On March 23, EPRI launched Flex MOSAIC, a uniform flexibility classification framework developed through its DCFlex initiative in collaboration with more than 65 utilities, system operators, regulators, hyperscalers and technology providers.
Announced at CERAWeek in Houston, the framework categorizes grid needs into standardized classes, allowing data center designers to select configurations that meet specific local grid requirements and accelerate connection approvals.
A second workstream focuses on interconnection requirements specific to AI factories, a category of facilities whose energy consumption profiles differ sharply from those of conventional data centers built before 2023.
“The new AI factories have a very different profile shape, which is driving additional requirement sets that we need to get into the interconnection requirement documents,” Ratnayake said.
Mukhortova offered a concrete example of how AI infrastructure can contribute to local energy systems rather than simply draw from them. Nebius owns a data center in Finland where a heat recovery module, built into the cooling infrastructure, captures server waste heat and feeds it into the municipal district heating system.
In 2025, the arrangement reduced local heating prices by 10% and avoided 4,000 tonnes of CO2-equivalent emissions, equivalent to removing more than 900 gasoline-powered vehicles from the road.
“AI infrastructure does not necessarily have to be just a consumer. It also needs to be mindful about the grid constraints,” Mukhortova said.
The question of whether AI’s energy demand will overwhelm grids divided the panel.
Ratnayake described what she called the “chunky peanut butter versus smooth peanut butter effect”: overall electricity demand growth has been relatively flat for a century, but AI data centers are concentrating that growth into specific industries over a compressed 15-year window, creating the appearance of a crisis.
Smith said AI’s downstream efficiency gains in logistics, air conditioning, materials science, and energy production could make it a net negative for total demand over time, once model training is complete.
“We don’t have an energy problem per se. We have a peak load problem,” Sivaram said. “These AI data centers can actually be great citizens that lower your bills, make the world more sustainable, and make it possible for us to have innovation in a very community-friendly way. They should be the heroes, not the villains of your community.”
The path to that outcome, panelists agreed, runs through tighter coordination between AI companies, utilities, chipmakers and regulators, and a willingness to treat grid flexibility not as a constraint but as a commercial opportunity.