Goldman Sachs published a number in March that should have stopped the room. 24x growth in token consumption by 2030. That is 120 quadrillion tokens per month. The source is agentic AI. The shift from single-prompt interactions to multi-step tasks that burn through compute like a turbine engine.
Everyone is looking at chip costs. They ought to be looking at the power plant.
The chip story is the easy story
Fortune ran the headline this week. “The AI economy could crash on mounting chip costs.” The logic is sound on the surface. A single Nvidia Blackwell GPU costs as much as a Tesla Model 3. Chip demand is outpacing fabrication capacity. A new fab costs tens of billions and takes years to build.
But chips are a solved problem technically. We know how to make more. We know how to design better ones. DeepSeek showed that algorithmic innovation can reduce compute requirements dramatically. The market will correct on silicon.
Energy is not a solvable problem the same way. The laws of thermodynamics do not respond to market signals.
The power curve nobody modeled
A single Blackwell-based cluster running at full inference load draws between 30 and 45 megawatts. That is not a server room. That is a small town. The hyperscalers, Microsoft, Google, Amazon, Meta, already consume more electricity than most European nations. By 2028, if Goldman’s token projections hold, AI inference alone will account for roughly 8 percent of global electricity consumption.
The grid is not ready. It cannot get ready in four years.
Data center buildout is constrained by power availability, not silicon supply. Virginia’s data center alley in Loudoun County is effectively tapped out. Dominion Energy stopped accepting new interconnection requests for large-scale facilities.
What a power-constrained correction looks like
The correction is a cap, not a crash. Token prices rise because the cost of the energy input spikes. Inference gets rationed. Agentic AI, the kind that runs background tasks and multi-step workflows, gets deprioritized for high-value user-facing applications.
The number that matters is not the price of a Blackwell GPU. It is the price of a megawatt-hour in Northern Virginia for 2028 forward contracts. Nobody is trading those yet. They will be.