Could AI Compute Become Interruptible Like Industrial Electricity Loads?

Photo: Mostafameraji, Wikimedia Commons, CC0

01 The idea in one exchange rate

Industrial customers have sold interruption to grids for decades: aluminum smelters, chemical plants and, more recently, cryptocurrency miners accept contracts that pay them to shed load during the hundred-odd hours a year when the grid is most stressed. Data centers were long assumed to be the opposite kind of customer — a server hall is popularly imagined as a load that can never blink. That assumption is now visibly breaking, and the reason is the arithmetic of the grid: systems are overbuilt for worst-case peaks, so a new customer willing to skip those peaks can connect to capacity that already exists.

Duke University’s modeling made the exchange rate explicit: 76 gigawatts of new load — about 10% of U.S. aggregate peak demand — could connect to the largest balancing authorities at an average annual curtailment rate of just 0.25%; 98 GW at 0.5%; 126 GW at 1%. And the capacity sits exactly where data centers want to be: roughly 18 GW in PJM, 15 GW in MISO, 10 GW each in ERCOT and SPP. The Department of Energy, for its part, reported finding no examples of grid-aware flexible data center operation — with one exception. That exception is now an industry.

Analysis grounded in the documented record, not a prediction. N43 and Hermes AI verified figures against Google’s announcements, Reuters reporting and grid-operator documentation as of September 19, 2026.

02 Google proved the load can blink

What changed the conversation from modeling to market was Google’s sequence of utility contracts. In August 2025 the company signed the first agreements delivering data center demand response by targeting machine-learning workloads specifically — with Indiana Michigan Power and the Tennessee Valley Authority — building on an earlier demonstration with Omaha Public Power District in which Google reduced ML-associated power during three grid events. By March 2026 it announced a cumulative 1 gigawatt of demand-response capacity embedded in long-term contracts with five utilities — adding Entergy Arkansas, Minnesota Power and DTE Energy — with the curtailment capacity explicitly used to connect new data centers to constrained grids faster. Google’s own caveat is as important as the milestone: capability varies by site, and there are limits to how flexible a given data center can be, since real-time services like Search and healthcare-critical cloud workloads cannot be interrupted.

The policy environment formed around the same time. A DOE report the prior year had identified essentially no flexible data center operation; within months, Indiana regulators had approved a large-load settlement, Ohio approved a similar one in July 2025, Texas passed a law requiring large data centers to reduce power during grid emergencies, and PJM launched a fast-tracked process for large-load interconnection rules while the Southwest Power Pool moved to streamline connection for facilities that commit to flexibility. EPRI’s DCFlex initiative, with Google as a founding member, is working to make demand response a formal capacity resource. Interruptibility has stopped being a curiosity and become an interconnection strategy.

03 Why AI is uniquely interruptible

The physical argument for interruptible compute is that AI load has a property profile no traditional industry offers: its work is checkpointable, deferrable and geographically shiftable. A training run saves state at regular intervals; pausing it costs time rather than product. Batch jobs — video processing, indexing, evaluation sweeps — are by definition not urgent; they can run at 2 a.m. or in a different region. Inference, the user-facing layer, is the only genuinely real-time component, and it is a minority of the power bill at training-heavy campuses. An aluminum smelter that stops loses metal and risks furnace damage; a paused training run loses, at worst, wall-clock schedule.

The counterargument is honest and must be stated: checkpointing is not free. Coordinated checkpoints of a large distributed training run can take substantial time and storage; failure recovery on a big run is a real engineering cost, not a press of pause. Long-horizon runs can hold state that is expensive to serialize; and every interruption introduces failure surface. The Duke study’s numbers were computed with curtailment capped at modest annual rates precisely because beyond some point, interruption becomes expensive rather than merely inconvenient. That cap — the tariff’s fine print — is where the market design lives.

04 Designing the tariff product

What would an interruptible-compute tariff actually look like? The pieces already visible in Google’s contracts and grid-operator rules sketch a three-product structure. Batch-interruptible: training and batch workloads commit to curtailing up to a capped number of hours per year (Duke’s scenarios imply roughly 0.25-1% of annual load, concentrated in peak events), in exchange for the deepest interconnection priority and price discounts. Scheduled deferral: non-urgent compute shifted on forecast rather than emergency signal — Google’s carbon-intelligent computing already does exactly this — rewarding predictable behavior with medium discounts. Critical-exempt inference: a firm-power carve-out for the small share of a campus that truly cannot blink, at little or no discount. The utility gets a planning-grade resource; the operator gets cheaper, faster power; the grid’s other customers get a slower rate spiral, since peak-driven infrastructure is the primary cost driver on everyone’s bills.

Who buys this? The natural customers are training-centric operators — frontier labs, GPU clouds, research campuses — whose work mix is dominated by interruptible jobs. Enterprises running latency-critical inference for customers (search, health care, finance) will mostly buy the firm product and pay for it. The interesting middle case is inference providers serving AI applications: their traffic has daily peaks that already follow load-shifting patterns, and a deferral-tolerant service tier (slower model, smaller batch, cached answers) is a plausible consumer product nobody has yet built on purpose.

05 The risks to training runs

The risk ledger for the buy side has three entries. Throughput risk: curtailed hours push out launch dates; a capped-tariff contract needs force-majeure style protections for events beyond the cap (the grid can call on you only so often). Integrity risk: interruption during distributed training can corrupt state if checkpoint discipline fails — the mitigation is engineering, but the tariff should price the storage and the checkpoint frequency, not assume them. Asymmetric-cost risk: the last week of a run whose model must ship on a fixed date is far more expensive to interrupt than the first; sophisticated buyers will want the option to firm-up specific windows — buying back firm power in the final stretch, at a premium — which is a natural derivative product for the same utilities selling the interruption.

On the sell side, the risk for grids is overestimating flexibility: a contract is only as good as the operator’s checkpointing discipline and honesty about what is truly critical. That argues for the verification infrastructure PJM is building into its large-load rules — measurable, metered curtailment, with penalties for non-performance — rather than the honor-system language of early programs. Texas’s emergency-curtailment mandate is the hard-law version of the same discipline; the tariff version simply pays for it in advance.

06 What to watch

Watch the second gigawatt: whether Google’s 1 GW milestone is followed by other hyperscalers (Microsoft, Amazon, Meta) signing named demand-response contracts, which would mark the product’s move from pilot to commodity. Watch PJM’s large-load interconnection rules when they land — they will define the template most U.S. grids copy. Watch whether curtailment caps become standardized (0.25%? 0.5%? 100 hours?) — the cap is the price, and standardization is what turns bespoke utility contracts into a market. Watch GPU-cloud uptake: training-rental firms are the natural mass buyers of deep-discount interruptible power. And watch for the first firm-up derivative — a contract letting an operator buy back firm power for a launch-critical week — because when that product appears, interruptible compute has become a real market rather than a set of pilot deals.