A transmission fault in Northern Virginia caused roughly 3,100 MW of data center load to disappear from the grid on July 22. Backup systems caught the facilities. Grid operators saw a voltage disturbance stretching from Washington to Chicago. The event gave the data center industry's grid promises an operating test.
In a September 1 Utility Dive opinion, Drew Baglino of Heron Power, Vlad Troy of NVIDIA, Michael Polsky of Invenergy and Varun Sivaram of Emerald AI set three duties for large compute loads: ride through grid disturbances, reduce demand during stressed hours and add generation, storage or transmission capacity.
The framework is directionally right. Its execution runs through the cooling plant. A GPU workload can pause in seconds. Heat stored in cold plates, coolant, piping and facility water continues moving. Pumps need minimum flow. Chillers and dry coolers have ramp limits. Restart sequences can create a second load step after the grid event passes.
That makes cooling controls part of the grid product. An operator promising 50 MW of curtailment must separate IT load from the thermal tail, then validate how quickly the full facility meter falls. Battery systems can bridge GPU ride-through and support pumps during the transition. The building management system, CDU controls and workload scheduler need one response plan.
The authors cite Duke University research estimating that curtailing large loads for less than 0.5 percent of annual hours could accommodate about 100 GW of new demand. That estimate depends on controllable load behaving as modeled. A data center cannot sell the same megawatt twice by counting IT curtailment while leaving the cooling response unmeasured.
On June 18, FERC ordered all six federally regulated regional grid operators to justify or revise large-load connection rules. The orders cover flexible-load transmission service, cost protection, telemetry and generation serving nearby loads. Grid operators now have a federal instruction to distinguish promised flexibility from dispatchable flexibility.
Cooling teams should enter tariff discussions early. Meter the IT and mechanical loads separately. Document pump minimums, thermal ride-through, chilled-water storage and restart behavior. Test the sequence under the same summer conditions used for capacity planning.
Good grid citizenship will become a commissioning requirement. The operator that can show a verified facility load curve will receive more credit than the operator carrying a slide about flexible GPUs. Cooling controls decide which one gets paid.