Nvidia and a group of U.S. energy and data center firms are developing a system that would let AI facilities dial back their power consumption during grid stress events, a strategy the partners say could free up 100 gigawatts of electrical capacity across the country. The effort centers on a “power-flexible AI factory” reference design that treats data centers not as fixed loads but as adjustable grid participants. If the concept scales, it would represent one of the largest demand-side interventions in the history of the U.S. power grid, equivalent to the output of roughly 100 nuclear reactors.
What the Partnership Actually Involves
Emerald AI, a firm focused on grid-interactive computing, is leading the design effort alongside Nvidia and a set of energy-sector partners. The group plans to build and operate a 96-megawatt Aurora AI Factory in Manassas, Virginia, slated to open in the first half of 2026. The facility will serve as a working prototype for the reference design, which other operators could replicate at their own sites.
According to Emerald AI, the Aurora project will bundle Nvidia’s latest GPUs with advanced power electronics, on-site energy storage, and control software able to modulate consumption in near real time. Instead of treating power usage as a fixed requirement, the system is designed so that AI workloads can be shifted, slowed, or temporarily paused in response to grid signals. The partners describe this as a template that utilities, cloud providers, and independent developers could adapt to local market rules and grid conditions.
Separately, Digital Realty, one of the world’s largest data center operators, confirmed it is collaborating with Nvidia at its own Manassas campus. That work covers the design and deployment of next-generation AI infrastructure, including liquid cooling, power management, and energy efficiency tied to the Nvidia AI Factory Research Center. The geographic overlap in Northern Virginia is no accident: the region hosts the densest concentration of data centers on the planet and faces some of the most acute grid constraints.
Digital Realty’s collaboration is distinct from Emerald AI’s Aurora facility but directionally aligned. Both initiatives are premised on the idea that AI data centers cannot simply demand ever-larger blocks of firm power; they must also offer services back to the grid. By co-locating flexible load capabilities with high-density computing, the partners hope to show regulators that AI growth can be managed without triggering chronic reliability problems.
How Flexible Load Actually Works in Practice
The core idea is straightforward but technically demanding. AI training workloads, unlike web searches or streaming video, can tolerate brief pauses or slowdowns without catastrophic consequences. If a data center can reduce its draw by a meaningful fraction during a summer heat wave or a generation shortfall, the grid operator avoids rolling blackouts and the data center avoids being denied a grid connection in the first place.
In practice, flexibility relies on a stack of software and hardware controls. Orchestration systems must understand which training jobs can be deferred, which clusters can safely run at lower power states, and how to stagger checkpoints so that progress is not lost when power levels change. Power distribution units and cooling systems also have to respond quickly, since cutting server power without adjusting cooling can waste energy and limit the net benefit.
A field demonstration in Phoenix, Arizona, tested this premise at commercial scale. Researchers ran a 256-GPU AI cluster at a hyperscale cloud data center and achieved a 25% power reduction sustained for three hours during a simulated peak grid event. That result, documented in a preprint paper, is significant because it shows the curtailment is large enough to matter to a utility and long enough to cover a typical afternoon demand spike. The test maintained AI workload quality of service throughout the reduction period, which addresses a key objection from operators who worry that flexibility means lost revenue.
The Phoenix trial also highlighted operational nuances. Not all AI tasks are equally flexible: latency-sensitive inference for real-time applications has less room to slow down than batch training runs. The demonstration focused on training workloads, which are more forgiving, but the partners argue that even inference clusters can provide limited flexibility if they are overprovisioned or use intelligent queuing. The Aurora reference design is expected to codify these distinctions so utilities have a clearer sense of what they can count on during emergencies.
EPRI’s DCFlex Program Sets the Testing Framework
The partnership builds on the DCFlex initiative launched by the Electric Power Research Institute, or EPRI, an independent nonprofit that conducts research for the utility sector. DCFlex is structured as a multi-year testing program with demonstration deployments beginning in the first half of 2025 and testing running through 2027. EPRI’s stated goals for DCFlex include enhancing grid reliability, improving asset utilization, and supporting the clean energy transition.
The Emerald AI announcement explicitly references EPRI’s DCFlex testing as part of the validation pathway for the Aurora facility. That connection matters because EPRI’s involvement gives utilities a degree of confidence that the flexibility claims are being independently measured rather than self-reported by the data center operators themselves. Utilities have historically been skeptical of large industrial customers promising to curtail load on demand, and third-party verification is one way to overcome that resistance.
Under DCFlex, data centers will be evaluated on metrics such as response time to grid signals, magnitude and duration of achievable load reductions, and the repeatability of those responses over many events. EPRI’s role is to translate those performance characteristics into language that system operators and regulators can use in planning and market design. If the Aurora AI Factory and similar projects meet DCFlex benchmarks, they could be treated more like traditional demand response resources in capacity markets.
Why 100 Gigawatts Is a Loaded Number
The 100 GW figure cited by the partners deserves scrutiny. It does not mean 100 GW of new generation capacity will be built. Instead, the claim is that by making existing and planned AI data centers flexible enough to reduce consumption during peak periods, the effective available capacity on the grid increases by that amount. Marc Spieler, identified in the Emerald AI press materials as an Nvidia spokesperson, framed the reference design as a way to unlock grid capacity equivalent to 100 nuclear reactors.
That framing is aspirational. It assumes wide adoption of the reference design across the U.S. data center fleet and presumes that utilities and grid operators will accept flexible data center loads as a reliable resource in their planning models. Neither condition is guaranteed. Grid operators like PJM Interconnection, which manages the mid-Atlantic grid where the Aurora facility will sit, have only recently begun to grapple with the scale of data center interconnection requests. Whether they will credit demand flexibility in their capacity planning remains an open regulatory question.
The 100 GW estimate also depends on how much flexible capacity each facility can safely offer without undermining its business model. A data center that commits to shedding 30% of its load for several hours might need to overbuild hardware or accept longer training times, both of which carry costs. The partners argue that higher utilization and faster interconnection approvals will offset these trade-offs, but until more real-world deployments are studied, the net economics remain uncertain.
Still, the direction is clear. Data centers currently face multi-year waits for grid connections in constrained regions. If operators can demonstrate they will not draw full power during the hours that matter most, utilities may be willing to approve connections faster and at larger scale. That is the practical unlock the partners are betting on.
The Tension Between AI Growth and Grid Limits
The U.S. power grid was not designed for the load profile that AI training imposes. A single large AI training cluster can consume as much electricity as a small city, and the load is concentrated in a handful of regions. Northern Virginia, Dallas-Fort Worth, and Phoenix have all seen sharp increases in data center power demand, straining local transmission and generation resources.
Most proposed solutions involve building new power plants, whether natural gas, nuclear, or renewable. Those projects take years and face permitting, financing, and community opposition. Transmission upgrades, which are essential for moving power from remote wind and solar projects to data center hubs, are even harder to deliver on short timelines. Against that backdrop, demand-side measures like flexible AI factories are attractive because they can be deployed alongside new computing capacity rather than waiting for grid expansions to catch up.
Emerald AI and Nvidia are positioning flexibility not just as a reliability tool but as a competitive advantage. Developers that can show credible, verifiable curtailment capabilities may gain priority in interconnection queues or qualify for new grid services revenues. To support that vision, the partners emphasize standardized processes, including the use of shared reporting platforms and automated telemetry, so utilities can audit performance without bespoke integrations for each site.
For now, the Aurora AI Factory in Manassas will be the most visible test of whether this model can scale. If it delivers on its promise (sustained, controllable reductions in power consumption without degrading AI services), it could shift how regulators think about data centers, from passive loads to active grid resources. If it falls short, the industry may find itself pushed back toward more conventional, and slower, grid expansion strategies.
Either way, the stakes are high. AI demand is rising far faster than traditional planning processes anticipate, and communities hosting data centers are increasingly vocal about reliability and environmental impacts. Flexible AI factories will not eliminate the need for new generation and transmission, but they could buy crucial time for the grid to adapt, if utilities, regulators, and data center operators can agree on the rules of engagement.
More from Morning Overview
*This article was researched with the help of AI, with human editors creating the final content.
