Signal Scanner · ARTIFICIAL INTELLIGENCE & AUTOMATION · 12 September 2026

AI Load Is Sold as Flexible Before the Flexibility Was Measured

US grid rules written this summer turn an AI data centre's willingness to be cut off into a priced tier of service. The first production measurement of how much load can actually be shed lands far below what those rules assume. Exposed: cloud operators, colocation, utilities, chip buyers and project finance.

The familiar story about AI and electricity is a shortage story: too much demand, too little generation, a race to build. Underneath it, a quieter change has been made to what an AI operator is actually buying. Between June and August 2026, US regulators created service classes in which large new loads agree to be interrupted, and ranked those loads below households in an emergency. Firmness, once an assumed property of a grid connection, has become a negotiable term with a price. Two weeks ago the first production-trace measurement of how much AI load can dependably be shed was published, and it sits well under the numbers the new rules assume. The gap is what should be watched over the next 18 months.

Signal Identification

A regulatory pivot with an unpriced measurement gap underneath it. Curtailability has moved from an engineering option to a tariff category, and the quantity being bought and sold in that category rests on modelled percentages rather than on observed fleet behaviour. The exposure is contractual, not technical: operators are committing to relief volumes that the first published measurement suggests they may not be able to deliver at the durations regulators care about.

Time horizon: 1-5 years (FERC show cause responses filed August and November 2026; PJM's Interim Resource Adequacy Service would apply from June 2027; firmness prices into contracts 2027-2029)
firmness prices in2026202720292030to 2038
Plausibility band: Medium–High
LowMediumHigh
Geographic / Jurisdictional Scope: Primary: the six US RTO and ISO regions under FERC's June 2026 orders, with PJM and SPP furthest ahead and ERCOT on its own controllable-load route. Spillover: Malaysia, Singapore, Korea, Ireland and the United Kingdom.
PrimaryPJMSPPERCOTMISO / CAISO
SpilloverMalaysiaKoreaSingaporeUKIreland
Sectors exposed:
Hyperscale cloud and AI operatorsColocation developersUtilities and grid operatorsData-centre project financePower trading and originationEnergy regulatorsWorkload orchestration softwareCorporate treasury

What's Changing

Three instruments landed inside ten weeks. On 18 June the Federal Energy Regulatory Commission issued show cause orders to all six US grid operators (FERC, 18/06/2026), ordering three new transmission products among them a “non-firm contract demand service” for “new large loads that can be curtailed when transmission capacity is constrained”, a product “curtailed before firm transmission service” (RMI, 06/07/2026). The Commission's docket asks whether facilities “that agree to be flexible and curtail usage” should clear interconnection studies inside 60 days (FERC, accessed 12/09/2026).

PJM then put a ranking on it. Its 13 August filing would cut demand from new large loads without their own generation “ahead of any action that would serve to shut off traditional consumers, including residential consumers”, and before calling on demand-response customers paid in advance to cut use (PJM, 13/08/2026). The justification is concentration: “Of 32 GW of growth in forecasted electricity demand between 2024 and 2030, 30 GW is attributed to data centers.” Operators have been willing sellers, on modelling that training and inference workloads “can offer between 18% and 55% flexibility relative to their average power consumption” (Utility Dive, 26/06/2026).

Then someone measured a real fleet. A 4 September study reconstructed “4,439 hourly power observations from a 185-day trace of 155,410 GPUs” and put immediate eligible curtailment at “12.1% of workload power and 6.35% of median facility power” (arXiv, 04/09/2026). Relief decays with event length, and a single mean-calibrated percentage “overstates these quantities by 17%, 25%, and 47%”.

What the rules assume, and what one production fleet delivered

ASSUMED FLEXIBILITY, AS A SHARE OF DATA CENTRE LOAD 0% 20% 40% 55% Modelled workloads 18% to 55% EPRI demonstrations up to 40% RMI Malaysia model cap 40% Measured fleet 6.35% of facility power AND IT DECAYS WITH EVENT LENGTH: 95% AVAILABLE RELIEF, ONE FLEET Median facility demand 55.8 MW Relief at 1 hour 2.51 MW Relief at 4 hours 2.32 MW Relief at 24 hours 1.95 MW A single mean-calibrated percentage overstates the one, four and 24 hour figures by 17%, 25% and 47%.

Sources: arXiv preprint 2609.05406 (4 September 2026) for the measured fleet and decay ladder; Utility Dive (26 June 2026) for modelled workload and EPRI demonstration ranges; RMI (3 September 2026) for the Malaysia modelling cap.

Disruption Pathway

Stage one runs to mid-2027 and is paperwork. RTO responses to the show cause orders were due in August, with the abeyance route pushing the rest to November (RMI, 06/07/2026), and PJM asked FERC to accept its service inside 60 days (PJM, 13/08/2026). Curtailment terms get drafted here, before anyone has been curtailed. Stage two is the first real events, from summer 2027, when a non-firm large load is called and either performs or does not. Stage three is repricing: performance data feeds back into interconnection studies, capacity accreditation and, eventually, the cost of capital for a project whose revenue depends on uptime.

Stress concentrates in three places. The first is workload mix, because training is batchable and inference is not, so a site's flexibility falls as it shifts toward serving. The second is portfolio correlation: aggregation raises four-hour firmness from 0.38 to 0.66 and then stalls, since clusters peak together (arXiv, 04/09/2026). The third is control, with utilities asking for direct authority over the load-side breaker as the price of a faster connection (Utility Dive, 26/06/2026). Two adaptations follow. Orchestration and on-site storage get built to manufacture relief the workload will not supply. Firm and non-firm capacity separate into products with different costs of capital, and colocation leases begin specifying curtailment hours.

Why This Matters Now

This lands on boards and CFOs at cloud and colocation operators, at the utilities writing the agreements, and at the funds lending against the buildings. The decision architecture that needs revising treats power as procurement and uptime as engineering: a curtailment clause is now a revenue term, signed by people who have not measured the quantity they are selling. Operators should instrument their own fleets before committing to relief volumes, and price flexibility by duration rather than as one share of load. Regulators should require evidence of dependable relief rather than a declared figure. Taken together, the sources suggest the credible sellers of flexibility will be those who can show a trace, not those who can quote a number.

Decision-action posture for this signal: Prepare — the service classes exist and the filings are live, but nothing has been curtailed yet and PJM's rule would not bite until June 2027, so positions should be built against named filing and event triggers rather than committed against a settled price for firmness.

Counter-Argument

The strongest objection is that the measured gap is an artefact of one fleet. The trace covers a single 185-day window and a workload mix a purpose-built flexible site would not run. Sites designed for flexibility have done better in demonstration, up to 40% (Utility Dive, 26/06/2026), and capacity-expansion modelling still finds real value in the trade, with zonal shifting cutting PJM system cost by 6% in 2028 (arXiv, 20/08/2026). On that reading the rules are correctly anticipating a capability the industry will build.

That may be right and still leave the exposure intact, because the contracts are being written now against the capability's projected version. The same modelling work concludes that “realistic event-shape limits diminish the value of curtailment”, and it is event shape, not headline percentage, that a tariff calls on. A firm that signs a non-firm service on demonstration numbers and operates on production numbers has sold a volume it does not hold.

Implications

This is a durable change rather than a cyclical one, because it alters the definition of the product rather than its price. Once curtailability is a tariff class it does not revert, and the same trade is already being designed into Malaysian interconnection policy (RMI, 03/09/2026). The inflection window runs from the November 2026 compliance filings to the first curtailment season in 2027. Winners are operators with measured, instrumented flexibility and the storage to underwrite it, and the orchestration vendors who supply the evidence. Losers are single-site operators with inference-heavy workloads who took the faster connection and priced the option at zero.

Early Indicators to Monitor

Disconfirming Signals

Strategic Questions

Keywords

Data centre flexibility; large load interconnection; non-firm contract demand service; FERC show cause orders; PJM Interim Resource Adequacy Service; curtailable load; demand response; AI compute firmness; workload orchestration; capacity accreditation; grid interconnection queue; inference workloads

Bibliography

Source tiers: Tier 1, governments, regulators and intergovernmental bodies. Tier 2, think-tanks, academic institutes, major consultancies and quality data providers. Tier 3, quality journalism and specialist trade press. Tier 4, vendor, company and practitioner sources, used only as directional corroboration.


Prepared by Shaping Tomorrow: 12 September 2026