Of the 100 GW of gas equipment GE Vernova carried under order and reservation at the first quarter of 2026, 56 GW were slot reservations the customer can defer or cancel. The market capitalises the whole book as revenue. It is an option written by the manufacturer, and the customer holds it.
That single disclosure is the reason for this piece, and it generalises well beyond one company. Across the chain that supplies electricity to artificial intelligence — turbine makers, transformer makers, independent power producers, grid operators — the volume being priced into equities today is a mixture of contracted revenue and of intentions that are free to abandon. The two are not distinguished in the multiples.
What follows separates them. We begin with the part nobody disputes.
The demand is not the debate
None of this is an argument that artificial intelligence does not need electricity. It does, in quantities that are documented rather than narrated. Global data centre power demand goes from 104 GW in 2025 to 132 GW this year and an estimated 290 GW by 2030 — roughly the entire generating fleet of a large industrialised country, dedicated to one category of building. Consumption reaches 565 TWh in 2026 against 415 TWh in 2024. The four US hyperscalers have guided to more than $650bn of AI infrastructure capex in 2026 alone.
Exhibit 1
Global data centre power demand, GW
The composition matters more than the level. AI-optimised servers account for 31% of data centre power in 2026 and overtake conventional servers in 2027; AI-focused capacity has been growing at 2.94 times the rate of the sector that contains it. This is not a rising tide, it is a wave inside the tide, and sector averages understate it. Geography concentrates the strain: the United States hosts roughly 45% of global AI capacity by power draw, US IT load is projected to move from about 80 GW in 2025 to about 150 GW by 2028, and Ireland's data centres already exceed 20% of national electricity demand.
The bottleneck moved from silicon to steel
For three years the binding constraint was Nvidia. It is no longer. Chips can now be bought and not switched on — because the cabinet that would energise them arrives in 2029.
Substation transformers of 5–50 MVA take 75 to 110 weeks. Generator step-up units above 50 MVA reach 150. On large high-voltage transformers the Tier-1 makers — ABB, Siemens Energy, Hitachi Energy, GE Vernova — quote 48 to 60 months, with some lines already taking orders for 2030–31. In 2020 the same equipment arrived in about a year. Heavy-duty gas turbine slots are four to eight years out.
Exhibit 2
Delivery lead times for critical electrical equipment, months
Capacity is being added, and the sums are not small: Hitachi Energy is building the largest US large-power transformer plant in Virginia, Siemens Energy is spending $150m in Charlotte, Eaton $340m in South Carolina. All three start production in 2027 or later, and announced expansions take three to five years to produce output — which means the 2026–28 window is fixed regardless of what anyone orders today.
The commercial consequence is already visible. Bloomberg reported in April that more than half of the US data centres planned for 2026 face delay or cancellation, and that only about a third of the 12–16 GW slated for the year was actually under construction. The gating item is switchgear, not capital.
So the demand is real and the scarcity is real. Both are in the price. The question this piece answers is narrower, and more useful to anyone holding these shares: how much of the volume now capitalised is cancellable, and who still gets paid if it is cancelled?