Why AI Data Centers Might Soon Cut Power at Busy Times
NVIDIA, Google and startup Emerald AI started a new group on September 16, 2026 to make AI data centers use power in a more flexible way. Engadget The group is called the AI Energy Management Alliance, or AEMA.
Flexible has a clear meaning here. Member data centers would move computing work to quieter times, use batteries kept on site, cut use during busy hours, and help in grid emergencies. NVIDIA calls it a first-of-its-kind coalition for data centers that can change electricity use on demand. NVIDIA
The request is direct. In exchange for flexibility that can be checked, AEMA wants power companies, regulators and policymakers to give approved AI data centers faster and larger power hookups. Another report puts it simply as faster connection for sites that can show they can move or cut load. Yahoo Finance
Waiting for power is the problem. Emerald AI CEO Varun Sivaram said new U.S. data centers can wait ten years or more for a grid connection. Large cloud operators and startups now treat a spot in line as a business limit. Fortune
AEMA spans chips, cloud platforms, power gear and grid controls. Emerald AI, a startup, provides the software that lowers power use during peak grid events while still meeting performance needs for running jobs. Axios
This builds on an earlier test. In March 2026, NVIDIA and Emerald AI worked with large energy companies to try out flexible AI factories as grid assets. That work tested whether big AI loads could provide grid services instead of pulling steady, flat power. NVIDIA News
The grid pressure behind the pitch
Small cuts can lower bills. A 2026 Duke study found a 1% to 2% cut in data center peak use can lower electricity rates by 0.5% to 2.8% and protect reliability. That finding gives utilities a number to use when buying flexibility instead of building only for the busiest moment, a bit like easing rush-hour traffic instead of only adding lanes. Utility Dive
Rules are already shifting. New rules in the PJM area, the grid that serves much of the eastern U.S., let combined data center and power plant projects connect faster, though reporting says the design helps new on-site gas plants more than renewables. Reuters In Washington, the U.S. House was set to take up a bill on September 11, 2026 to curb electricity cost rises tied to data centers. Equipment maker Delta told Reuters that demand for AI power, cooling and data center gear is still a growth engine. Reuters
The broader context here is a trade. A data center that agrees to cut back when asked can lower the highest shared load, delay costly wire upgrades, and answer worries that households will pay for AI growth.
What flexibility actually requires
This idea is not new. Aluminum smelters and big cold-storage warehouses have long taken cheaper power under interruptible tariffs, or deals that let the grid interrupt them. AI is faster and more exact. AI computer groups can swing tens of megawatts in seconds, and job schedulers can pause, move or slow training work more easily than a factory line can stop. Running live answers for users, called inference, is less flexible because it must meet strict latency SLOs, or fast response-time targets. That makes checking critical.
Trust is the hard part. A faster hookup needs a shared definition of a cutback event, meters that show response inside the dispatch window, and scheduler links that save checkpoint state, or saved progress, and keep speed promises. Batteries can cover short events, but longer cuts still mean delayed tokens, delayed training steps or moved inference work. Utilities will want firm, checkable cuts. Cloud customers will want steady performance.
In my view, proof is where this plan wins or stalls. If AEMA can set a common way to show how much load moved, how fast, for how long, and with what effect on service, regulators can change the waiting line without risking enough supply to meet demand. Done well, bigger headline connections could sit behind a smaller guaranteed amount, fitting more computing on current wires while new lines are built. I have watched computing do this before, turning better scheduling into room to grow, and that history leaves room for hope about what this opens up.
