Finnacl Research

R3 · free executive summary

Power for AI

The Electricity Trade — Who Wins the Constraint?

Power for AI: The Electricity Trade

Executive summary — free edition

Finnacl Research · September 2026 · The full report covers 12 core companies across five layers, 30 more names, eleven funds, two historical analogues built from the original filings, and a quarterly-scored early-warning scorecard.


The answer

The winners are whoever owns the things that take longest to build. The losers will be whoever builds against the forecast instead of the contract.

The binding constraint on the AI buildout is no longer chips — it is electricity, and specifically the sequence required to power a large new site: a grid connection (three to seven years in the main US hubs), the transformer between the grid and the rack (two and a half to five years), and the turbine or reactor behind it (five to seven years for heavy gas; every small reactor is 2030 or later). Money cannot compress any of that. So money is buying position in the queue instead — twenty-year nuclear contracts, paid turbine slot reservations, and gas plants built on site.

Three things follow.

  1. The scarcity is real and priced as such. One regional capacity market has cleared at its legal maximum three years running. GE Vernova has 116 gigawatts of gas turbines under contract and is already selling delivery slots for 2031. Turbine prices are on a path to nearly triple their 2019 level; transformers are up 77% and take up to five years.
  2. But the demand forecast is inflated at the margin — and that is where the losses will be. Utilities have been asked for 1,066 gigawatts of data-centre power and can commit to roughly 298; the difference is largely the same projects filed with several utilities at once. Of ~90 gigawatts of on-site generation announced, about 2 gigawatts is actually running — a ratio of 45 to 1.
  3. The market has priced the order book, not the output — and that is the exact trade that inverted in 2001.

The valuation finding, in one line

Median forward earnings multiple by layer: contracted generation 12.9× · regulated utilities 18.0× · builders 26.1× · equipment 27.3× · nuclear fuel 66.5×. The closer a company sits to the electron, the cheaper it is. The generator selling twenty-year power to the best-capitalised buyers on earth trades at under 13 times earnings; the company taking reservations for turbines it will deliver in 2031 trades at 45.

The history, from the original filings

In the 2001 merchant-power bust, General Electric's own annual reports show large gas turbine sales falling from 362 units to 122 in two years, with $1.5 billion of customer contract termination fees — while its orders rose in the first year of the decline. Order books break last. And Calpine's filing shows it cancelled 87 gas turbines and 44 steam turbines in a single quarter, releasing $3.4 billion of commitments for a $207 million charge: about six cents on the dollar. That is the historical price of walking away from what the industry now calls a slot reservation.

What would change our mind

Eight markers, re-scored and published every quarter. Today: two red (the interconnection queue and equipment lead times — both supporting the thesis), one green (contracted power agreements), five amber. One has already moved against us: in August, Texas paused new data-centre development and the federal energy agency cut its Texas 2027 load-growth forecast from 14% to 6% — the first official downward revision of this cycle. We publish that rather than explain it away.

The uncomfortable result we print anyway

Over the past twelve months, the one fund concentrating the merchant-generation thesis this report finds cheapest lost 1.85% while the plain utility index gained 5.65%. Being right about the constraint has not yet been the same as being paid for it.

What's in the full report

The lead-time ladder · the value-chain map · five constraint chapters with the numbers · 12 core-company profiles ordered by valuation layer · the two historical analogues in full · the fund purity analysis (one company sits in seven of eleven funds' top ten; the merchant thesis has almost no index sponsorship at all) · exposure routes by scenario · the client Q&A · full sourcing on every figure.


Disclosures. Impersonal research, identical for every purchaser and tailored to no one; not individualised advice, and Finnacl Research is not an investment adviser (publisher's exclusion, Advisers Act §202(a)(11)(D)). The tiers differ in seats and redistribution rights, not in content. No statement about what any reader should hold, no recommended allocation, no forecast. The author holds no position in any company covered in this report. No issuer payment, no advertising, no trading against published views, and no trade in any security named here in the thirty days before or after publication. Full disclosures in §13 of the report.

That was the whole finding, not a teaser. The full report adds the evidence behind it: every figure with its source and date, the Tier-1 snapshots, a plain-language version you are licensed to hand to a client, and the quarterly re-scores for as long as the edition is live.