California’s grid operators spent 2024 throwing away solar power at record scale — terawatt-hours curtailed, midday prices dipping to zero and below. The electrons were real, the panels were paid for, and nobody could use them fast enough. Meanwhile, a twenty-minute drive away, GPUs were burning 700 watts each to answer questions. Put those two facts in the same room and stay there until the implication lands: the marginal electron at noon is nearly free, and intelligence is about to be priced against it.
Run the division yourself. A modern datacenter GPU draws roughly 700W sustained. Serving a mid-size open model, that card pushes on the order of two to five million tokens an hour. At $25 per megawatt-hour — a generous solar PPA price — the electricity inside a million tokens costs between half a cent and two cents. The electrons in a long, rambling conversation cost less than the network packets carrying it. Energy was supposed to be the binding constraint on AI. For inference, the constraint is already rounding error.
The objection writes itself: if energy is nearly free, why isn’t inference nearly free? Because today’s price is capex amortization wearing an energy costume. You are not paying for the watts; you are paying down the $30,000 card, the building, the cooling, the financing. That capex follows its own brutal curve — dollars per million tokens for equivalent quality have fallen roughly a hundredfold in two years — while energy sets the floor. And the floor is sinking through zero at noon.
Curtailment economics is the mechanism that connects the two curves. Solar PPAs clear at $20–30/MWh, curtailed solar clears at nothing or less than nothing, and the gap between those prices is a standing invitation to any load flexible enough to chase it. Training runs, embedding jobs, batch inference, distillation — the industry’s background workload is almost perfectly time-shiftable. It eats exactly the electrons the grid cannot use, at exactly the hours the grid will pay you to take them.
Behind-the-meter inference is where this lands physically. The pattern already exists in miniature: hobbyists running home GPUs off rooftop solar, timing big jobs for the midday export peak their retailer barely pays for. Scale that instinct up one notch and you get edge operators siting inference next to solar farms; up another notch and you get hyperscalers signing dedicated solar PPAs for AI campuses, which is already happening. Each rung of the ladder is the same trade: intelligence manufactured from electrons nobody else wanted.
Batteries change the shape of the trade without changing its direction. Four hours of storage turns the noon glut into evening inference, which means token pricing goes time-of-day the way electricity did — expect off-peak API discounts to become explicit rate cards within a couple of years. The analogy to rooftop PV is nearly exact: dollars per watt fell roughly 90% in a decade as deployment scaled, and every forecaster who extrapolated linearly missed low. Inference cost is falling faster, from a higher start, with the same compounding structure.
Watch what breaks first. Per-seat AI pricing is arbitrary once marginal cost approaches zero — it survives only as long as buyers cannot do division. The permanent-GPU-shortage narrative splits in two: frontier training stays scarce and expensive, while commodity inference drowns in abundance; anyone quoting one price for both is selling you a story. And cloud egress quietly becomes the real tax, because moving tokens costs more than minting them — which is why the winners will manufacture intelligence as close to the electron source as physics allows.
The investment conclusion is boring in the correct way: build for abundance, not scarcity. Architect as if inference were water — metered, cheap, everywhere — rather than diamonds to be rationed behind seat licences and rate-limit theater. The grids of the 2030s will curtail staggering quantities of clean power, and every curtailed megawatt-hour is a standing offer to whoever shows up with silicon. Intelligence will trade at the marginal price of off-peak solar. Price your products accordingly, before someone else does it for you.