OpenAI's data centers will shed 1 GW of power when the grid strains
OpenAI's answer to the data center power fight is not more supply. It is a switch — and on Monday the company put the first megawatt figure on it.
OpenAI launched the Grid Protectors Program on Monday, a national effort to design its future US compute campuses so that utilities can cut their power draw when the grid is under strain, according to the company's announcement. The first concrete number is in Georgia. Under its agreement with Georgia Power for the planned Effingham County campus, OpenAI has committed to up to 1,000 megawatts of flexible demand response — about the average electricity demand of 700,000 households, which the company points out is more than Washington, D.C. and Boston combined. For a single campus, that is roughly the size of a large gas plant's output that the grid can call on without building anything.
The money is the part that will decide whether this is a real program or a press release. OpenAI says it will pay the full infrastructure and electric-service costs required to serve Effingham, rather than having them socialized across other ratepayers. Georgia Power projects about $950 million in annual customer savings beginning in 2029 from its portfolio of large-load agreements, including this one, and estimates a benefit of at least $15 a month for a typical residential customer using 1,000 kilowatt-hours, per the utility's own release. Those are the utility's projections, not audited results, and they lean on growth materializing on schedule.
Why this lands now is not mysterious. US data center capacity is projected to exceed 100 gigawatts by 2030, according to BNEF research OpenAI cites, and 2026 is the year the electricity bill arrived in politics. Moratoriums went on ballots, state commissions started pushing back on special tariffs, and county boards discovered they could delay an interconnection queue. We followed that fight through the governor races in data center moratoriums are on the ballot in 16 governor races. A campus that can be switched off answers the central objection — that AI is taking power from households and handing the bill to them — more cheaply than building another generator.
The engineering caveat is where the headline number gets soft. Flexible demand is a range, not a fact, and three unpublished numbers decide what 1,000 megawatts actually means: how much notice the utility must give, how long the reduction lasts, and how many times a year it can be called. A battery can respond in milliseconds. A training cluster running a synchronous job across tens of thousands of accelerators cannot simply be dimmed without checkpointing work and losing throughput. OpenAI says the arrangements "will differ by location" and that it will share additional commitments as they are finalized, which is an honest way of saying the Effingham terms beyond the capacity figure are not public.
OpenAI also concedes the limit of its own idea. "Flexibility cannot replace investment in new energy supply and transmission," the announcement says. That sentence matters because demand response is often sold as a substitute for the hard, slow, unpopular work of building lines and plants. It is not. It is a way to use existing infrastructure better at the margin, and the program as described applies to campuses OpenAI builds in the future — the load already contracted elsewhere is not part of this commitment.
Read the program as a permitting strategy wearing an engineering badge. The binding constraint on AI buildout in 2026 is not capital or chips; it is a grid connection and a local vote. A lab that walks into a county hearing with a contract saying its campus can drop a gigawatt on request, and that it will pay the full cost of the wires serving it, is arguing about a different project than the one that got shouted down. Whether the argument is true depends entirely on terms nobody has published.
There is a real precedent for skepticism about utility projections, and a real one for optimism. Georgia Power's large-load framework has been through public litigation for two years, and the $950 million figure is the utility's own estimate of a portfolio effect, not a measured outcome. On the other side, the Effingham agreement is a signed contract with a regulator-approved tariff structure, which is more than most AI energy announcements can claim.
Anthropic has quietly added two sentences to Claude Code's system prompt telling it not to recreate an existing video game, and developers say the same guardrail is now refusing ordinary reverse-engineering work. The rule, confirmed by Anthropic's Addy Osmani and first surfaced in a screenshot posted on X, reads: build original games, games in a genre, and games from your own designs and assets — but do not build "a recreation of a specific existing game, even if it is renamed or called original," as Tom's Hardware reported. Osmani said the change was made because "using Claude to infringe IP is against our usage policy," and asked users to send him cases where Claude wrongly refuses work that is theirs.
The scope is narrower than the reaction suggests, and wider than Anthropic probably intended. The blocking appears tied to the combination of Claude Code and the Opus 5.5 model — developers report that switching to Opus 4.8 lets the same task through — which means the rule lives in a harness prompt, not in the model's weights or the API. That is a policy statement you can point at in an IP dispute, and a filter that can be routed around by anyone who knows which model selector to change.
The collateral damage is the interesting part. Claude Code's own GitHub repository has filled with complaints from developers whose non-game work is being refused, and Tom's Hardware collected direct reports of reverse-engineering tasks hitting the guardrail: its former editor was blocked running checks on his own software on October 31, and a Reddit user had an account suspended and reinstated after using an open-source reverse-engineering tool alongside Opus 5.5. Anthropic already runs a Cyber Verification Program that gates full security capabilities to vetted organizations, so a stricter posture on that class of request is consistent — it just also catches people doing decompilation and porting work that has been legal, public and community-run for years.
The context is a retro-gaming scene drowning in machine-generated ports and decompilations, to the point that one decompilation tracker froze its public leaderboard this year because AI forks were gaming it. Ars Technica's read is the one to keep: a prompt-level rule will not stop the copying boom, because the copying does not depend on Claude Code specifically. It only moves the work to the next model that says yes.
What to watch: whether Anthropic extends the rule to the API and claude.ai, and whether the Effingham contract's response time and duration ever get published. Both stories are about the same thing — a company writing its own limits before someone writes them for it.
If your campus can drop a gigawatt on request, should it get to skip the interconnection queue? Write to us via our contact page.
