The AI power trade has its first hard denominator: 412.6 terawatt-hours a month, and 9–17% by 2030 is now an arithmetic test

For two decades US electricity demand was flat and nobody had to check. That ended in 2026, and August is the first month the whole question can be settled against a meter instead of a press release. The EIA put total retail sales at 412,568 GWh, up 4.6% year on year, with commercial sales — the customer class that contains data centers — rising 5.1% while industrial sales fell 1.5%.

Two details deserve more attention than the headline. Residential consumption rose 7.9% while its price rose only 3.4%; industrial customers paid 7.3% more per kilowatt-hour while buying 1.5% less power. And commercial sales of 150,597 GWh are already 36.5% of every kilowatt-hour sold to US customers, sitting within 1.8% of residential sales on a year-to-date basis — the closest those two classes have ever been.

EPRI's range — data centers at 9% to 17% of US electricity by 2030, against roughly 4.5% today — sounds like a technology forecast. It is really a forecast of the denominator. Annualising the 2,809,886 GWh of retail sales the EIA has reported for January through August 2026 puts today's US market near 4,215 TWh, of which data centers are roughly 190 TWh and everything else roughly 4,025 TWh.

Two facts in the same release point in opposite directions. Henry Hub gas averaged $2.76/MMBtu in August, down from $2.93 in July, and the gas-equivalent cost of generation fell to $22.09/MWh. Yet the average retail price rose 5.1% to 14.87 cents. The molecule is a shrinking share of the bill and the plant is a growing share.

With gas supplying 46.7% of generation but worth roughly $22 of a $148.70 retail megawatt-hour before efficiency, operating cost and capital, the AI power problem is a capital-cost problem, not a fuel-cost problem. Value accrues to whoever owns the plant, the transformers and the rate base — not to whoever owns the molecule.

That is the bull case, and it is not a straw man. GE Vernova booked $24.2B of orders in a single quarter, up 88% organically, and lifted total backlog to $176B while gas turbine backlog plus slot reservations went from 100 GW to 116 GW, with at least 125 GW targeted by year-end. The Korean suppliers show this is a global equipment shortage rather than a US bubble: Doosan Enerbility became the first Korean company to sell 380 MW gas turbines to a major US technology customer, and HD Hyundai Electric has a $721M North American data-center order in hand alongside a KRW 1.1tn agreement.

But an order book is a claim about the numerator, and the denominator moved 4.6% in the very month the data was published. If data centers land at the low end of EPRI's range, a large tranche of turbine, transformer and switchgear capacity gets delivered into a market that grew less than 5% — which is a backlog problem, not a growth problem, for the suppliers.

The third quarter of 2026 produced something the order books cannot explain. Bloom Energy posted the best quarter in its corporate history — revenue of $1,235.8M, up 138.1%, product revenue up 179.4%, GAAP gross margin of 33.8%, and full-year revenue guidance raised to $4.4B–$4.6B — and the shares fell 11.2% in the next session.

I cannot prove the denominator caused that drawdown: interest rates, AI capital-spending sentiment and sector rotation all moved at the same time. What is verifiable is the shape of it — operating results accelerated while the shares fell — and that the market began underwriting the denominator instead of the numerator. In this trade the numerator is the order book. The denominator is whether enough megawatts can be delivered, connected and financed on time.

Here is the denominator's blind spot. Bloom Energy sells solid-oxide fuel cells that generate electricity on the customer's own site, behind the meter, without a utility. That is data-center load growth that will never appear in the EIA's commercial sales line — and it is growing at 138% a year while grid-served commercial sales grow 5.1%.

It also helps explain a divergence inside the August release itself: utility-scale net generation rose 6.8% year on year while retail sales to customers rose 4.6%. Part of that gap is weather — residential sales jumped 7.9% on cooling load — and part is load served outside the retail system. Either way, EIA commercial sales are a floor on data-center consumption, not a measurement of it.

On the other side sit the counterparties. Constellation Energy signed an additional 920 MW of long-term nuclear power purchase agreements on 15- to 20-year terms starting between 2029 and 2032, lifted full-year adjusted operating earnings guidance to $11.50–$12.50 a share from a second-quarter print of $2.55, and ran its nuclear fleet at 44,160 GWh against 45,170 GWh a year earlier. Talen Energy holds a 1,920 MW agreement with Amazon Web Services from its Susquehanna plant.

The clearest downstream evidence comes from a regulated utility in a data-center state rather than from a merchant generator. The EIA found that Virginia's commercial electricity sales rose by nearly 30.0 million MWh between 2019 and 2025 — more than any state except Texas — and that the growth is largely driven by a concentration of data centers. Dominion Energy serves that load: second-quarter 2026 revenue of $4,558M, and trailing capital spending equal to 66.7% of revenue.

The cost question is the one investors keep getting wrong. Industrial customers paid 7.3% more per kilowatt-hour in August while buying 1.5% less power; residential customers paid 3.4% more. The cost of new data-center load is being spread across the rate base of the states that host it — which is exactly what Virginia's regulator has been asked to decide, and why the low end of EPRI's range is a regulatory question before it is an engineering one.

The August data neither confirms nor refutes the hyperscaler power claims. It prices them. The denominator is roughly 4,215 TWh and growing 4.6%; commercial demand is growing 5.1% and is the only customer class with a structural reason to accelerate; industrial demand is shrinking; and the marginal megawatt-hour is a gas megawatt-hour bought at a capital cost, not a fuel cost. That combination rewards the owners of scarce equipment and rate base and punishes anyone whose thesis is simply that power gets more expensive.

The second-order consequence is the one to keep an eye on. If the denominator has to grow 11% faster than the EIA's baseline to deliver the high case, then the deliverable is not more equipment — it is more transmission, more interconnection queues and more regulatory permission. Those are the bottlenecks that decide, over 2027 and 2028, whether the order books convert into generation.