Cledion
Back to home

The grid is the bottleneck. The fix is software.

April 12, 20264 min read
Data center cooling infrastructure

For most of grid history, large electricity loads were inflexible. Factories ran on schedules. Smelters ran continuously. Heating and cooling tracked the weather, not the market. Demand-response programs existed but worked at the margin, tied to backup generators and thermostat setpoints. The grid planned for the peaks and built the wires to match.

AI changes that. Training is interruptible (checkpoint and resume). Workloads are shiftable in time (run tonight instead of this afternoon) and in space (run in Illinois instead of Virginia). GPUs have a continuous power dimmer through dynamic voltage and frequency scaling. The very thing driving the power crisis is the most flexible large load that has ever been connected to a grid.

The technology question is closed

Phoenix, May 2025: a 256-GPU cluster reduced power 25% for three hours during grid stress. Software only. No hardware swap. No batteries. All service-level agreements held. The peer-reviewed paper hit 4.52% RMSE on power prediction. Follow-up demos in Chicago, Manassas, and London confirmed the same result at scale and across markets.

EPRI's DCFlex consortium now has nine demo sites across the US and Europe. The technology to make AI workloads flexible exists, has been proven, and is being deployed.

The market question is open

FERC ordered PJM to create flexible interconnection in December 2025. Texas SB6 requires real-time telemetry for large-load flexibility, with rules finalizing December 31, 2026. The DOE has used emergency authority twice in 2026 to activate data center backup generation during grid stress. Regulators are pulling on this lever harder every quarter.

What is missing is the connective tissue: the methodology that says this reduction counts, for this capacity credit, in this market. Without that, flexibility commitments are promises a grid operator cannot plan on. The technology shipped. The trust did not.

More in the next post on what that trust gap looks like in practice. Get in touch if your work intersects this.