The largest single buyer of AI datacenter power has volunteered to be interruptible. For everyone behind OpenAI in the queue, flexible load just moved from nice-to-have to the opening term of the interconnection negotiation.
OpenAI launched its Grid Protectors Program on November 2, committing that campuses it develops itself will reduce electricity demand when the local grid is under strain, with the terms — how much, how fast, for how long — negotiated utility by utility. The pilot is the planned 3.2 GW Effingham County, Georgia campus with Georgia Power, where OpenAI has already agreed to shed up to a gigawatt of demand on request and to pay the full cost of the grid infrastructure that serves it. Some sites may also feed the grid from on-site generation or batteries.
The policy backdrop explains the timing. FERC's June orders directed the six big grid operators to accommodate flexible and behind-the-meter large loads; the Senate permitting bill would let states require curtailment as a condition of service; Texas has paused new datacenter grid approvals until an audit of its 474 GW queue is complete. OpenAI is getting ahead of a rule that was coming anyway, and turning it into a ratepayer-protection story.
For the market, the consequence is structural. A campus that can shed 30% of its load on a hot afternoon is a different interconnection application from one that cannot — and the operators who can offer it will be the ones that get served in 2027 and 2028.
What it means if you're buying: Expect curtailment provisions in new power contracts, and price them in. Training workloads can checkpoint and pause; latency-sensitive inference cannot. If your deployment is inference-heavy, sites with on-site generation or batteries that can ride through a curtailment call are worth a premium over pure grid sites. If it is training-heavy, flexibility is a bargaining chip: offer it, and ask for the earlier energization date in return.
What to do about it: Powered sites on the marketplace that already carry this story: SPICA 285 (285 MW IPP direct supply in North Dakota, scalable to GW level), SIRIUS 1000 (hybrid grid plus behind-the-meter gas, Alabama) and ANTARES 30 (30 MW Texas colocation scaling to 130 MW, RFS spring 2027). Sellers: if your site has interruptible capability, batteries or on-site generation, put it in the listing — it is now a headline feature, not a footnote.
Program announcement · openai.com · axios.com ↗ · Summary and analysis by Ax3
