Data centers and peak‑demand constraints

Eleven quiet months

A data center in late July has run without incident since the previous summer. Then a heat dome settles over the region and stays. Inside, cooling systems climb toward their design limits. Outside, the grid operator declares an emergency and asks its largest customers to shed load. For one week, the capacity the facility sold and the capacity it can deliver diverge.

The loss belongs to a building. The operator has sold continuous availability under contract; the facility draws grid power and rejects heat, and both constraints tighten on the same few days.

The load is arriving faster than the grid

NERC’s January 2026 long-term assessment is a severe reliability warning. Summer peak demand across the North American bulk system is now forecast to rise 224 GW over the next ten years, 69% more than the prior assessment, and NERC names new data centers as the main source of that load. Its description of the system is unchanged in kind: enough capacity on a normal day, shortfalls on an extreme one.^[2]

American data centers now draw an estimated 177 to 192 TWh of electricity a year, about 4 to 5% of the national total. EPRI, in its 2026 Powering Intelligence study, projects 380 to 790 TWh by 2030, between 9 and 17% of all US electricity, and Lawrence Berkeley National Laboratory puts its reference case at 649 TWh.^[1][5] In Virginia the share already exceeds 25%, a load pocket running well ahead of the wires that serve it.^[1]

The reliability problem became an operating fact in May. NERC issued a Level 3 Essential Action Alert, its highest, after computational loads (AI training, crypto mining, conventional data centers) dropped off the system or oscillated in seconds, leaving operators no room to respond. Registered entities had to act by August 3. NERC’s own finding from the prior warning was that entities lacked processes to address this class of load.^[3] The facility is a customer that might be asked to shed, and a source of grid instability on the same peak days when cooling margin is thinnest.

Heat hits both sides

Each additional degree raises the cooling load inside the facility. The same weather burns through the reserve margin on the grid outside it. Both failure paths hit on the same afternoons, so an operator prices them as one risk.

On-site power problems caused 45% of the most recent impactful data center outages in Uptime Institute’s 2026 survey, down from 54% a year earlier but still far ahead of any other cause. Uptime attributes the new pressure points to worsening grid constraints and high-density workloads. A facility that loses cooling survives on thermal ride-through, a margin measured in minutes that keeps getting thinner as racks grow denser.^[4]

Operators and tenants

The operator has sold availability: colocation and cloud contracts specify uptime, and a breach costs it service credits against revenue already booked, then refunds, renegotiated terms and, at the top of the range, a customer that moves. Weather and grid conditions the operator does not control trigger that liability.

The same week costs money on the other side of the ledger. Workloads the facility cannot host are revenue not earned. Backup generation runs on fuel bought at short notice, against maintenance intervals and run-hour permits. The operator procures power for the hours it does hold into a scarcity price. A demand-response commitment sold to the grid as an option turns into an interruption the day the grid operator calls it. The operator also derates saleable capacity, because the megawatts a building can market follow what it can cool on the worst day.

Tenants absorb throttling decisions they did not make, including software businesses three contracts from the meter that never see the building.

Listed hedges and the indemnity

Power futures and heat-rate positions settle against hub prices, which covers the cost of energy consumed and leaves the cost of capacity the operator could not sell. CME lists weather futures and options on temperature indices, including cooling degree days at named US cities, and they are a real hedge for temperature-driven cost. Their settlement reference is a regional index computed from an airport station, a different reading from the conditions at a specific intake or the point at which a chiller plant runs out of margin.

Property and equipment policies respond to physical damage and mechanical breakdown. Business interruption pays measured loss after a deductible, a waiting period and an adjustment in which the insured proves what the lost days were worth. Curtailment accepted under a demand-response tariff, or capacity derated to stay inside thermal limits, involves neither damage nor breakdown, so it sits outside those triggers even when the financial loss is plain.

The leftover basis is index temperature against facility-level conditions, a hub price against an availability obligation, and an indemnity determined after adjustment against a loss that lands inside one identifiable week.

Pricing the week that matters

Third parties record the drivers of this loss on a fixed schedule. The National Weather Service publishes station temperatures and heat advisories. Grid operators publish conservative-operations notices, emergency declarations and the dispatch of demand-response programs, each timestamped. Utilities file disturbances above the reporting threshold with the Department of Energy on Form OE-417. Any of those supplies an objective variable: a station maximum above a stated threshold for a stated number of consecutive days, or a declared emergency event in a named balancing authority inside a defined window.

Any of those can anchor a contract with terms fixed before the season: a stated trigger, a stated premium, a stated payout, settled against the published record. The leftover basis is the distance between the trigger and the facility, a term to negotiate up front.

Discrete structures these contracts bilaterally with institutions whose year turns on a handful of days. If that July week sits unhedged on your balance sheet, tell us where the exposure lives.