When the Utility Calls

What a data center is to the grid has changed. For most of its history, a data center was the ideal utility customer: large, steady and inflexible by design. It asked for firm power and promised never to move. AI changes the second half of that bargain. Some AI workloads can move, and the grid has begun to pay for that movement.

The clearest example comes from Google. In August 2025, Google announced demand response agreements with Indiana Michigan Power and the Tennessee Valley Authority. Under them, certain Google data centers reduce the power used by machine learning workloads when the utility asks, during periods of grid stress. Google said this was the first time it was delivering data center demand response by targeting machine learning workloads specifically. The Indiana Michigan Power arrangement is tied to a data center in Fort Wayne, Indiana. Google also said the approach followed a demonstration with the Omaha Public Power District in which it reduced machine learning power demand during three grid events.

The idea has a longer history inside Google. In 2020, the company described a carbon-intelligent computing platform that shifts flexible tasks, such as batch processing and some model training, to hours when cleaner power is available within the same data center. The 2025 agreements apply the same mechanism to a different signal. Instead of carbon intensity, the input is grid stress. Instead of an internal preference, the output is a commitment to a utility.

The mechanism relies on a distinction between workloads. Serving a live user is time-sensitive. Training a model or running a large batch job often is not, because it can pause at a checkpoint and resume. That property, discussed in the module on utilization, is what makes flexibility possible. A training job that can be checkpointed can be slowed or stopped for a few hours at a known cost.

A second demonstration adds numbers. Emerald AI, a startup backed by NVIDIA, worked with Oracle Cloud Infrastructure, NVIDIA, the Arizona utility Salt River Project and the Electric Power Research Institute, or EPRI, on a test in Phoenix. The results, published in mid-2025 as part of EPRI's DCFlex initiative, reported that a GPU cluster, meaning a cluster of graphics processing units, reduced its power use by 25 percent for three hours during a period of peak demand on May 3, 2025, ramping down over about fifteen minutes and back up without exceeding its prior baseline. It did this by orchestrating which jobs slowed, which paused and which continued.

Why would a utility value this? Researchers at Duke University's Nicholas Institute answered that in a February 2025 report titled Rethinking Load Growth. They estimated that about 76 gigawatts of new load could be added to the United States system if those loads accepted curtailment for an average of a quarter of one percent of their maximum annual uptime. The logic is that grids are built for a few peak hours a year. A load that steps aside during those hours uses capacity that already exists.

Policy is moving in the same direction, less gently. Texas enacted Senate Bill 6 in June 2025. It applies to large loads of 75 megawatts and above and requires that loads interconnecting after the end of 2025 install equipment allowing the grid operator, the Electric Reliability Council of Texas, or ERCOT, to curtail them during firm load-shed emergencies. In Texas, flexibility is becoming a condition of connection rather than a service for sale.

Now connect this to the control plane. A demand response commitment is only as good as the software that honors it. The facility receives a signal from the utility. That signal must reach the scheduler, which decides which jobs to slow or pause, in what order, with what checkpointing. The accelerators then reduce their power, and the electrical monitoring must confirm the reduction to the utility. Every layer from the grid to the chip participates. NVIDIA's DSX Flex component, named in its March 2026 reference design, is an attempt to productize exactly that chain.

Faraday's read is that grid-responsiveness will become a siting advantage before it becomes a revenue line. The money from demand response programs is modest relative to the value of computing. The value of connecting sooner, in a region where firm capacity is scarce, is large. An operator whose control plane can prove flexibility may reach power that an inflexible operator cannot.