On November 2, OpenAI announced the Grid Protectors Program, a national commitment that its future compute campuses will reduce electricity demand when the grid is under strain. Working with utilities, regulators, and policymakers, the company said, it will develop arrangements for campuses it builds to cut their draw during periods of grid stress — the winter cold snap, the summer heat wave, the hours when homes and businesses need power most. The announcement is framed as a standard OpenAI wants its campuses to meet: pay the costs they create, and be flexible when the system needs them to be.
The program has exactly one quantified example so far, and it is worth examining closely, because it is both substantial and unfinished. The pilot is Project Camellia, OpenAI’s planned campus in Effingham County, Georgia, about 45 minutes outside Savannah. Under a 25-year agreement with Georgia Power, announced July 22 and cleared through the Georgia Public Service Commission’s regulatory process in August, OpenAI will pay the full infrastructure and electric-service costs required to serve the project, and Georgia Power will be able to reduce energy delivered to the facility by up to 1,000 megawatts during periods of high demand. OpenAI says that figure is equivalent to the average electricity demand of roughly 700,000 households. Set against the campus’s expected total of approximately 3,200 megawatts, the curtailable share is roughly 31 percent: a partial reduction, not a shutdown.
One caution belongs before any analysis, because it conditions everything that follows. The campus is not built. OpenAI’s July announcement said power would be delivered in phases between 2028 and 2032, and its November announcement repeats that the campus is still in development. The 1,000 megawatts is therefore a commitment about future operations, not a record of performance. No curtailment has been called, no response time has been demonstrated, and no other site’s terms have been disclosed. Until then, Grid Protectors is a policy with one detailed example attached.
What the pledge does commit to
What is on the record is not trivial. A 25-year agreement in which the customer pays the full cost of the infrastructure built to serve it, plus a 1,000-megawatt demand-response commitment, is a materially stronger bargain than the early large-load deals that fueled the “who pays” backlash of the last two years. The Georgia framework behind it matters here. Georgia Power’s August announcement put the OpenAI contract inside a broader portfolio story: projected incremental revenue from new large-load customers producing approximately $950 million per year in customer savings beginning in 2029, or $2.847 billion over 2029 through 2031, illustrated as at least $15 a month for a typical residential customer using 1,000 kilowatt-hours.
Read those savings figures with the attribution they carry. They are Georgia Power’s projections for its whole large-load portfolio and expected growth, including Project Camellia — not savings attributable to OpenAI alone, and not savings any customer will see before 2029 at the earliest. The utility frames the $15-a-month figure as an increase from its December 2025 commitment of $102 a year to $180 a year. That is a forecast about a portfolio, not a bill credit, and it depends on the large-load growth actually arriving on schedule.
There is also a transparency caveat that cuts against the program’s own logic. The underlying contract has not been made public — local reporting has described it as heavily redacted on trade-secret grounds — and the August clearance took the form of PSC staff declining to object to the filing, not a vote of the five-member commission. One commissioner, Peter Hubbard, publicly argued that an agreement of this size deserved additional scrutiny, while also calling the 1,000-megawatt flexibility commitment a positive feature that should appear in future contracts. The public knows the ceiling and the term, but not the operating language.
What the pledge does not say
That brings us to the gaps, and they are the heart of this story. The public materials — OpenAI’s November announcement, its July project announcement, and Georgia Power’s two releases — do not specify how quickly the campus must respond when Georgia Power calls for a reduction, how long a curtailment can last, how often it can be called, which workloads can be interrupted, whether OpenAI receives any separate compensation for making the capacity available, or how the emissions consequences of a curtailment would be accounted for. OpenAI’s announcement acknowledges the shape of the gap without filling it: arrangements will differ by location, and the company will work with utilities to agree on how much demand can be reduced, how quickly, and for how long.
These are not footnotes. They are the terms that determine whether a gigawatt of flexibility is a grid resource or a press release. Demand response is a product defined by its operating parameters. A campus that can shed 1,000 megawatts within minutes, sustain the reduction for hours, and do so repeatedly through a multi-day heat wave is a fundamentally different asset from one that can shed the same megawatts with a day’s notice, for one hour, a handful of times a year. The first displaces generation investment; the second trims a peak. The public materials do not say which one Georgia is getting, and the difference between them is the entire value proposition.
The workload question is the least discussed and possibly the most consequential. Training runs can, in principle, pause and resume; inference serving live users cannot. OpenAI says its newer compute facilities are “much different than data centers built even a few years ago, allowing us to be flexible in our power demand.” But flexibility is not uniform across computing work. The public materials do not say which portion of the campus’s load is interruptible, or what happens to curtailed work — whether it is deferred, shifted to another campus, or absorbed as lost throughput. If curtailed computing migrates to a campus in another state, the grid relief in Georgia is real but the system-wide load picture is merely relocated. Nothing in the public record addresses that accounting.
Compensation is similarly unaddressed. In organized demand-response programs, participants are typically paid for availability, for performance, or both, and the payment structure shapes the incentive to actually show up. Here, the public record describes the flexibility as part of the service agreement, not as a separately priced product. But without terms, outsiders cannot tell whether the 1,000 megawatts is a firm obligation with penalties for non-performance or a best-efforts arrangement. A ceiling without an enforcement mechanism is a hope, not a hedge.
The fight this pledge enters
To understand why OpenAI chose this moment to nationalize a Georgia contract clause, it helps to name the fight the pledge enters. The large-load debate has always had two questions, often conflated. The first is who pays: whether gigawatt-scale infrastructure is funded by those loads or socialized across existing customers. The second is who flexes: whether a load of that size must bend when the system is stressed, or can demand firm service at any hour.
Georgia has, by its own framework, largely answered the first question for this project: OpenAI pays full costs, with financial assurances and long-term contracts designed to protect other customers if the project is delayed, reduced, or canceled. Under Georgia PSC rules first approved in January 2025, the costs of serving a new load of this size are not supposed to fall on existing customers. Grid Protectors is OpenAI’s attempt to answer the second question voluntarily, campus by campus, before regulators answer it by rule. That distinction — voluntary offer versus enforceable standard — is where the announcement’s significance and its limits both live.
The regulatory backdrop makes the timing legible. On June 18, 2026, the Federal Energy Regulatory Commission issued six show-cause orders, one to each of the regional grid operators it regulates and their transmission owners, directing them to justify or reform tariff provisions governing how data centers and other large loads interconnect. The orders defined large load, in preliminary terms, as a new commercial or industrial customer with peak demand of 50 megawatts or more, interconnecting above 69 kilovolts, and not part of a co-location arrangement. Among the five reform categories was transmission service for large, flexible loads, including potentially new service products built around that capability. Revised filings are due in mid-November.
Note the jurisdictional wrinkle. Georgia Power operates outside the organized markets the June orders address; Georgia is a vertically integrated utility state, not an RTO footprint. The FERC orders will shape the rules for flexible loads in PJM, MISO, CAISO, and the other organized regions, but they do not directly govern the Effingham contract. That makes the Georgia deal both a model and an exception: a bilateral arrangement negotiated under state rules, now offered as the template for a national program that will land, site by site, in very different regulatory terrain. What counts as satisfactory flexibility in Georgia may not satisfy a grid operator in PJM, where accreditation, dispatch protocols, and performance penalties are set by market rules rather than by contract negotiation.
Flexibility as the third option
This desk has followed the large-load power question through two earlier answers. The first was bring-your-own-power: turbines, SMRs, and fuel cells at the fence, generating supply on the customer’s timeline. The second, documented in late October, was Google’s bring-your-own-capacity deal with Voltus: funding 1,070 additional megawatts of aggregated flexibility — batteries, thermostats, industrial curtailment — across PJM through 2040, buying demand-side capacity assembled from thousands of existing sites rather than building generation.
Grid Protectors is a third option, and the contrast with both is instructive. Unlike bring-your-own-power, it does not add supply; it subtracts demand at chosen hours. Unlike Google’s Voltus deal, it does not aggregate other people’s flexibility; it commits the campus’s own load, in a single bilateral agreement, at a site that does not yet exist. The OpenAI pledge is a design commitment for a future facility: flexibility engineered into the campus rather than bolted on afterward, which is arguably the more durable approach — if the engineering and the contract language match. But design intent is not operating history, and the first real test of either model is a scarcity event that has not happened yet.
OpenAI itself supplies the most important limiting sentence in its own announcement, and it deserves to be quoted plainly: “Flexibility cannot replace investment in new energy supply and transmission.” A 1,000-megawatt curtailment right reduces the generation Georgia Power must build for future growth at the margin; it does not eliminate the need for the transmission and generation that a 3,200-megawatt load, phased in over four years, will require in ordinary hours. Flexibility shaves peaks. It does not carry baseload.
What to watch
Four things will determine whether Grid Protectors becomes a template or a trophy. The first is publication: whether the operating terms of the Georgia agreement — response time, duration, frequency, workload scope, compensation, and non-performance consequences — become public, either through the semiannual portfolio reports Georgia Power agreed to publish or through a future disclosure. Until they do, every assessment of the program rests on a ceiling, not a specification.
The second is the FERC process. The grid operators’ revised filings in the June show-cause proceedings, due in mid-November, will show whether flexible large loads get a standardized service product in organized markets or remain a matter of bilateral negotiation. If a standard emerges, OpenAI’s future campuses in RTO regions will be judged against it, and the Georgia terms will be only one data point.
The third is replication. OpenAI has named no second site and no second utility. The program’s credibility as a national effort, rather than a Georgia announcement with national branding, depends on the next commitment, and on whether its terms are disclosed more fully than the first one’s. Watch for whether future agreements publish the parameters this one did not.
The fourth, and the one that ultimately settles the argument, is performance — but it is years away. The earliest phases of Project Camellia are not expected to draw power until 2028. The first summer peak or winter storm in which Georgia Power actually calls the 1,000 megawatts, and the campus actually delivers it, will be the first real data point in this program’s history. Until that hour, the honest description of Grid Protectors is the one its own announcement invites: an ambition, attached to a contract, awaiting its first test.
That is not a dismissal. A developer of OpenAI’s scale announcing that flexibility will be a core part of how it builds campuses — and accepting a 1,000-megawatt curtailment obligation at its first site — shifts the negotiation baseline for every large-load deal that follows. Developers who previously treated firm, uninterruptible service as the default will now be asked why they cannot offer what OpenAI offered. But a baseline is not a standard, and a voluntary offer is not a rule. The “who pays, who flexes” fight is still being settled, in commission orders and tariff filings, by institutions with the power to make the answer binding. Grid Protectors is a strong opening bid in that negotiation. It is not yet the settlement.