A London Data Center Cut Its Power 40 Percent in Under a Minute. OpenAI Just Wrote That Trick Into a 1,000-Megawatt Promise in Georgia.

In December, inside a Nebius data center on the outskirts of London, a bank of 96 Nvidia Blackwell Ultra GPUs did something data centers are built never to do: it throttled itself. Researchers ran more than 200 simulated grid emergencies against the cluster over five days. Software called Emerald Conductor, built by a startup named Emerald AI, cut the facility's power draw by as much as 40% each time, in under a minute, without dropping a workload anyone cared about. In the fastest test, the cluster shed 30% of its load in under 40 seconds after researchers simulated the sudden loss of a major power plant elsewhere on the grid. National Grid, Emerald AI, EPRI, Nebius, and Nvidia ran the trial together and published the results in March. Every one of the 200-plus scenarios succeeded.

That's a real, measured, third-party-witnessed result, and it matters more than the London setting suggests. The entire argument for building gigawatt-scale AI campuses next to constrained grids rests on a bet that the load is flexible enough to bend instead of break something. Until this trial, that bet was mostly theoretical. Nvidia's Josh Parker put the pitch plainly: this kind of infrastructure can act as a "grid-aware asset," modulating its own demand in real time to support grid stability, rather than simply drawing whatever it wants whenever it wants it. National Grid says it's now using the trial's data to shape the technical standards for how future AI data centers connect to the British grid. Nvidia is already running a larger version of the same idea at Aurora, a nearly 100-megawatt "power-flexible" AI factory it operates in Virginia.

A London lab result becomes a Georgia contract clause

On November 2nd, OpenAI tried to turn that kind of result into a binding commitment on this side of the Atlantic. The company announced the Grid Protectors Program, a standing policy for its own future compute campuses to cut their power use when the grid they're sitting on is under strain. Georgia is where it starts. The pilot is Project Camellia, OpenAI's planned campus in Effingham County, about 45 minutes outside Savannah, which Georgia Power expects will eventually need roughly 3,200 megawatts, built out in phases between 2028 and 2032.

The mechanism is a 25-year supply agreement, already cleared by the Georgia Public Service Commission, that lets Georgia Power cut up to 1,000 megawatts of power delivered to the site during periods of high demand. It's worth being precise about what that is and isn't: it's demand response, a reduction in what Camellia draws, not a new source of power flowing back out to anyone's home. OpenAI has separately committed to cover the full infrastructure and electric-service cost of serving the site, and Georgia Power has told regulators that its large-load contracts — Camellia among them — are projected to save the average residential customer using 1,000 kilowatt-hours a month about $15 a month starting with the 2029 rate case, roughly $950 million a year in customer savings across its whole large-load portfolio.

This is the same Camellia contract this blog has covered before as a fight over transparency: the underlying agreement was filed with every page redacted as a trade secret, and the bill-relief number only grew from roughly $8.50 to $15 a month after advocacy groups and a sitting U.S. senator pushed back publicly before PSC staff signed off on August 26th. What's new on November 2nd isn't the contract. It's OpenAI formally branding the curtailment clause as a program, and saying — without yet specifying terms — that future campuses elsewhere will get some version of the same arrangement, worked out case by case with local utilities and regulators.

Two different kinds of proof, running together

Here's where a lot of the coverage of both stories blurs together, and where it shouldn't. The London trial is a completed, independently run, published technical demonstration: a real cluster, under real software control, tested against more than 200 scenarios, with every result reported, good or bad. The Georgia contract is a legal promise about a campus that doesn't exist yet. Project Camellia is still years from drawing its first phase of power, let alone 3,200 megawatts of it. The 1,000-megawatt curtailment commitment has never been exercised, because there's nothing built yet to curtail. One of these is evidence. The other is a bet, backed by a 25-year contract and a regulator's signature, that the evidence will transfer.

It probably will, at least partly — the physics of throttling a GPU cluster doesn't change much between London and coastal Georgia, and Nvidia's own hardware sits at the center of both stories. But "probably will" is exactly the gap GridPermit's evidence ladder is built to flag. A vendor's demonstration, a utility's regulatory filing, and a signed, exercised interconnection record documenting an actual emergency event are three different rungs, not one fact repeated three times. Reporting on OpenAI's announcement that cites the UK trial as proof the Georgia pledge will work is skipping a rung. The honest version of this story is that nobody will know whether Camellia's 1,000 megawatts of promised flexibility holds up until PJM or Georgia's own grid operator calls on it during a real event, years from now, under conditions nobody fully controls.

There are open questions in the contract language itself that matter more than the announcement lets on: what conditions actually trigger Georgia Power's right to curtail, how long a reduction can last, and what happens if a grid emergency hits at the exact hour Camellia's own workload can't be throttled without real cost. OpenAI has acknowledged, to its credit, that flexibility of this kind doesn't replace the need for new power plants and new transmission lines — it buys time and headroom, not capacity. Future campuses' terms, the company says, will be negotiated separately with each utility and regulator, which means Georgia's numbers tell you very little about what a Grid Protectors deal will look like in Ohio, Texas, or wherever OpenAI builds next.

I think the useful shift here isn't the megawatt figure, it's the direction. Most of the fights this blog covers are defensive: a town trying to stop a project, a commission trying to claw back ratepayer protection after a contract's already signed, a utility trying to prove its cost allocation won't stick residents with the bill. A demand-flexibility pledge is the first time a data center developer has shown up offering to be smaller on command instead of just bigger whenever it wants. That's worth taking seriously. It's also worth remembering that a signed promise and a proven one are not the same document, and the only thing that turns one into the other is an actual emergency, measured and reported with the same rigor the London researchers used — not a press release timed to a program launch.