The first load to be shed

On 13 August, PJM Interconnection asked federal regulators for permission to switch off data centres before it switches off anybody else.

The instrument is called Interim Resource Adequacy Service, and the logic is stated plainly. A new large load — end-use demand with a cumulative peak of at least fifty megawatts at a single electrical site, entering service after 1 June 2027 — either brings its own generation or accepts that during a capacity shortage it will be curtailed prior to the deployment of Pre-Emergency Load Management. Ahead, in other words, of the demand-response programmes that would normally be the grid's first move, and a long way ahead of anything that would dim a house. PJM proposes to maintain a registry of who these loads are, where they sit, and whether they brought their own supply.

PJM was not improvising. On 18 June the Federal Energy Regulatory Commission issued show-cause orders to all six of the grid operators it regulates — PJM, MISO, SPP, CAISO, ISO New England and NYISO — giving each sixty days to justify its existing rules for connecting large loads or to say what it would change instead. That clock ran out on 17 August. FERC's own account of what it wanted is not ambiguous: the orders promote flexible transmission service, which it defines as non-firm service to a load "willing and able to limit withdrawals from the grid," in every region. CAISO's straw proposal, published on 12 August, adopts the same fifty-megawatt threshold and offers two flexible interconnection services. SPP has a price-adaptive load product moving through its stakeholder process. The binding tariff filings arrive in November.

As energy policy, this is sensible and overdue. The interconnection queue is the real constraint on AI's physical expansion, network upgrades take years, and a load that can stand down for a few hours is genuinely cheaper to serve than one that cannot. Nobody should want the alternative, which is either a four-year wait or a grid built to a peak it meets twelve times a year.

But look at what is actually being sold, because it is not electricity at a discount. The trade is speed for interruptibility, and it changes what power is inside a software architecture. For as long as the discipline has existed, electricity has been a failure domain: a thing that is either present or absent, engineered around with dual feeds, generators, availability zones and the working assumption that its absence is rare, local, and nobody's decision. Flexible interconnection makes it something else entirely — a resource rationed by a third party you have no contract with, on a signal your system cannot observe, in hours chosen by weather. That is not a failure to survive. It is a scheduling constraint, and no architecture in production expresses it as one.

The obvious objection is that interruptible service is not new. Aluminium smelters have lived on curtailable tariffs for decades, and nobody wrote an editorial about it. True — and the analogy breaks in exactly the place that matters here. A smelter is one workload, on one site, owned by the entity that signs the tariff. The party accepting the interruption is the party that bears it, and it knows precisely what it gave up. A data centre is thousands of workloads belonging to people who have never heard of Docket EL26-67. The interconnection agreement is signed by an energy and real-estate function; the consequence lands on a service owner who was not consulted, in a runbook they did not write, against an SLO they published to somebody else. That gap between where the commitment is made and where it is felt is the oldest architectural failure there is, and this is a large new instance of it.

It is not hypothetical. Google said in March that it has now integrated a gigawatt of demand response into its long-term energy contracts with US utilities, and was explicit about the lever: limiting or shifting a portion of the machine-learning workloads running in its data centres. That works because training is the most interruptible large workload the industry has ever built. It checkpoints. It has no user waiting. Delaying it by four hours costs money and annoys a research team, and that is the whole blast radius. The flexibility case was made on training's properties, which are unusually generous.

The numbers make the deal look cheap, and they are good numbers. The Duke Nicholas Institute's 2025 study of curtailment-enabled headroom found that the existing US system could absorb roughly 76 GW of new load at an average annual curtailment rate of 0.25% of maximum uptime, 98 GW at 0.5%, and 126 GW at 1.0%, across balancing authorities covering most of the country's peak demand. A quarter of one per cent of a year is about twenty-two hours. Twenty-two hours is less unavailability than most teams actually deliver, and it comes with notice.

The average is the wrong statistic, though, and it is wrong in a way architects should recognise instantly. Curtailment is not distributed randomly across the year. It happens on the continent's worst days — a heat dome, a winter morning, a long still evening — which is to say it happens to everyone at once. NERC's 2026 summer assessment flags elevated shortfall risk in New England, the Pacific Northwest and SaskPower under above-normal conditions, and names prolonged extreme heat as the primary threat. Weather of that kind does not respect the boundary between PJM and MISO. Meanwhile, the standard architectural answer to correlated risk — put it in another region — assumes those regions fail independently. Against a hazard whose trigger is a continental weather system, that assumption is close to worthless. We would be adopting, deliberately and at scale, the least diversifiable form of unavailability we have ever accepted, and paying for a redundancy strategy that does not address it.

Then there is the workload the flexibility case was not made on. Inference does not checkpoint, and a curtailed request is a failed one. The research response to that is already visible, and it is worth reading closely: a June paper on arXiv proposes quantization-enabled demand response, treating model precision as a dispatchable parameter — switching model instances and routing requests to lower-precision configurations when the grid asks for load reduction. The reported result is a 34.3% cut in data-centre operating cost, and the phrase used for it is "without curtailing served token volume." Every request is answered. Every request is answered by a smaller model.

Sit with what that implies operationally. The load is shed in the quality of the output. No error rate moves, no latency budget is breached, no page fires. We have availability SLOs, latency SLOs and error budgets; we have nothing at all that says the answers were worse between four and eight on Tuesday because the grid was tight, and no mechanism by which that fact would reach the team whose product depends on it. An unmeasured dimension is where the cost will be booked, because that is always where cost is booked.

The strongest version of the counterargument is not that curtailment is rare. It is that it is contractual — you know the terms in advance, which is more than can be said for any outage you have actually suffered, and a known bounded degradation is a legitimate engineering trade, priced and chosen. I think that is right, and it is why the flexible tier will and should be taken. But it is an argument for making the terms explicit, not for ignoring them. There is a sharper caution from within the energy debate itself: critics of how the Duke headroom figure is being deployed point out that system-level flexibility is not a site-level entitlement, that flexibility only counts if its availability coincides with both the load's peak and the system's peak, and that interconnection studies want firm, obligated reduction rather than voluntary participation. That criticism cuts against the optimists, and it cuts against me too — it argues that the curtailment commitments finally written into tariffs will be firmer and less forgiving than the ones being discussed now.

What is really returning here is capacity planning, in a form the cloud spent fifteen years abstracting away. The units are physical, the scheduler is a public body, and the interface is a phone call to a facility. Elasticity was always a claim about someone else's problem; flexible interconnection quietly hands part of that problem back, at the level of the building, to tenants who cannot see it. The architectural work is not difficult, but it is unfamiliar: knowing whether your capacity sits behind a flexible interconnection, asking your provider the question, and assigning every workload a curtailment tier the way you already assign it a recovery objective. Most organisations have no vocabulary for that. Signing the agreement assigns the tier anyway.

For thirty years the industry's founding promise was that computing would become a utility. It is finally happening, and not in the sense anyone meant. Utilities are metered, tiered, and rationed under stress, and they decide the order in which customers go dark. We asked for compute to behave like electricity. It has started to.