Abhay Chakra Sadineni

Research paper · November 2026

Speed to Power

Abhay Chakra Sadineni

EnergyGrid flexibilityBattery storageAI infrastructure

Abstract

Grid operators now ask large AI campuses for three kinds of flexibility. On a scale of seconds, they want synchronised training jobs to stop swinging tens of megawatts. On a scale of hours, they want the right to curtail the campus during the few hundred hours a year when the system is stressed. And on a scale of months to years, campuses that accept curtailment may be connected sooner. This paper prices each layer for a campus built on current Blackwell hardware, using public specifications, published grid studies and current battery costs. The power-supply energy storage NVIDIA now ships, 65 joules per GPU, holds about 1 kWh across 100 MW of racks, enough to ride through a 50 MW swing for about 70 milliseconds. Swings with periods of seconds need tens of kilowatt-hours, which is little energy but a large amount of fast power. For curtailment, a widely cited study finds that 76 to 126 GW of new US load could be added if it accepts curtailment of 0.25 to 1 percent of its maximum annual energy. On GPU rents and capital costs in late 2026, the compute lost at those rates is worth about $18,000 to $167,000 per MW a year. A battery that carries half the load for two hours costs about $67,000 per MW a year. Even on the optimistic assumption that it avoids all curtailment, it pays only above curtailment of roughly 0.4 to 0.9 percent. The largest number is the third. One month of earlier operation is worth 730 full-load hours, as many as 33 years of curtailment at 0.25 percent or 8 years at 1 percent, before discounting and whatever the price of compute. If flexibility shortens the wait for a grid connection, which is plausible but not yet measured, it is mainly a way to buy time.

Cite

Abhay Chakra Sadineni (2026). Speed to Power. Research paper. https://doi.org/10.5281/zenodo.24210931 https://www.abhaychakra.com/papers/speed-to-power/