The grid that powers a fifth of America just hit the panic button, and AI is why.
The Summary
- PJM Interconnection, the grid operator for 65 million people across 13 states, is seeking 15 gigawatts of new power capacity in an emergency procurement to prevent blackouts driven by AI data center demand.
- 15 GW is roughly equivalent to 15 large nuclear plants, enough to power 11 million homes, and this is being framed as an emergency, not a long-term planning exercise.
- The AI infrastructure buildout is outpacing grid capacity faster than anyone modeled, and the people who keep the lights on are now in reactive mode.
The Signal
PJM runs the electricity market from Illinois to New Jersey. It's not a utility. It's the matchmaker between power plants and the wires that deliver electricity to your house. When PJM calls something an emergency, it means the math stopped working. The AI boom isn't just consuming power. It's consuming it faster than new generation can come online.
Fifteen gigawatts is not a rounding error. For context:
- That's more capacity than the entire state of Virginia currently uses at peak demand
- It's equivalent to adding 15 large nuclear reactors, or roughly 50,000 utility-scale wind turbines
- PJM's total capacity today is around 180 GW, so this is an 8% jump in a single procurement
"The AI infrastructure buildout is outpacing grid capacity faster than anyone modeled."
The emergency framing matters. Normal grid planning cycles run 3-5 years out. You forecast load growth, approve new plants, wait for them to get built. This process assumes incremental, predictable demand. Data centers used to fit that model. A Facebook or Google facility might add 50-100 megawatts over a couple years. Manageable.
AI training clusters don't work that way. A single frontier model training run can pull 100+ megawatts for months. Inference at scale, the thing that happens when millions of people use ChatGPT or Claude simultaneously, adds continuous baseload demand that doesn't sleep. Multiply that across every company now rushing to deploy agents, and you get load growth that looks more like turning on a city than plugging in a few buildings.
The grid wasn't built for this. Most of PJM's territory still runs on coal and natural gas plants built in the 1970s and 80s. Nuclear provides steady baseload, but we haven't finished a new reactor in the region in decades. Renewables are growing, but they're intermittent, and battery storage at the scale needed to smooth out solar and wind gaps is still expensive and scarce.
Key dynamics at play:
- Data center developers are signing power deals faster than utilities can build generation
- AI workloads have different load profiles than traditional computing, more constant, less flexible
- Regulatory approval for new plants, especially nuclear or large gas, takes years PJM doesn't have
The Implication
If you're building in the agent space, power availability is now a first-order constraint, not an afterthought. Where you deploy infrastructure matters as much as what you deploy. Regions with surplus generation capacity or fast permitting for new plants will attract the next wave of AI buildout. Regions that can't keep up will price out or simply reject new data center projects.
Watch PJM's procurement process. If 15 GW gets absorbed quickly, expect more emergency rounds. If it struggles to find bidders, the bottleneck isn't demand for power, it's the ability to build it fast enough. Either way, energy is now a core input cost for the agent economy, and it's going up.