100 GW hiding in plain sight

The US power grid is built for the worst hour of the worst day of the year. For the other 99% of the year, enormous generation and transmission capacity sits idle. The Nicholas Institute at Duke quantified the gap in 2025: if large loads accepted curtailment for just 0.25% of uptime (about 22 hours per year), 76 gigawatts of new load could be absorbed without building anything new. At 1% curtailment, the number climbs to 126 GW.

For scale: a large nuclear reactor is about 1 GW. Twenty-two hours of flexibility unlocks the equivalent of 76 new reactors worth of capacity that already physically exists.

The fix is not generation. It is coordination.

The grid does not need more steel in the ground. It needs the loads connected to it to be willing, occasionally, to use slightly less. AI training is interruptible. AI workloads are shiftable in time and in space. GPUs have a power dimmer switch. The single biggest driver of the power crisis is also the most flexible large load ever connected to a grid.

Phoenix proved it live in May 2025: a 256-GPU cluster reduced power by 25% for three hours during grid stress, in software, with all service-level agreements maintained. Follow-up demos in Chicago, Manassas, and London confirmed it works across markets and conditions.

So why is it not happening yet?

Because a grid operator cannot plan on a promise. The grid is a life-safety system. Without independent proof that a curtailment event actually happened, no operator will grant cheaper, faster interconnection on the strength of a flexibility commitment alone.

The doing is solved. The proving is not. The proving is what is worth building.

More on this in the next post. If your work touches grid flexibility, capacity accreditation, or data center interconnection, get in touch.