You Can Buy an AI Chip Tomorrow. The Turbine to Power It Ships in 2031.
The newest constraint on the AI buildout isn't silicon — it's a 300-ton spinning machine that only three factories on Earth know how to forge, and their order books are already full for years.
Order a turbine in 2026 vs. 2034 — same factories, very different wait
Illustrative model, not a forecast: bars are indexed (0–100), not real gigawatt counts. "Backlog" tracks the reported industry pattern of roughly tripling since 2024 then leveling off as OEMs work it down; "delivery capacity" tracks GE Vernova's and Siemens Energy's own stated capacity-expansion targets through 2028–2030. The wait-time line is derived from that gap and simplified to move smoothly — real wait times shift in steps as specific factory expansions come online. Drag the slider or press Run to move from 2026 to 2034.
The plain version
Here's the twist in the AI power story nobody saw coming two years ago: the problem isn't building enough chips, or even enough data centers. It's building enough power plants to run them.
A big AI data center can pull as much electricity as a mid-sized city. The fastest way to get that much reliable power on-site is a natural gas power plant, and the heart of a gas power plant is the turbine: a jet-engine-sized machine spinning thousands of times a minute, built from parts forged and machined to tolerances measured in microns. Only three companies on the planet make the large ones — GE Vernova, Siemens Energy, and Mitsubishi Power — and between them they run just a handful of factories capable of casting the giant metal blades and rotors involved.
Demand has exploded so fast that all three are effectively sold out for years. GE Vernova's order book jumped from 83 gigawatts of turbines at the end of 2025 to 116 gigawatts by mid-2026, and it's chasing 125 gigawatts by December. Mitsubishi Power is already booking delivery slots for 2031 through 2034. Put plainly: you can buy a state-of-the-art AI chip and have it running within months. Order the power plant to run the data center around it, and you could be waiting the better part of a decade.
It's already colliding with the real world — Texas froze new data-center grid hookups on August 3 because pending requests topped 474 gigawatts, more than five times the state's entire record power demand. The line for a turbine, not the line for a chip, is now setting the pace of the AI buildout.
The expert version
The binding constraint on new AI data-center capacity has shifted from semiconductor fabrication to heavy-frame gas-turbine manufacturing — a forged-metal industry whose production economics differ fundamentally from silicon's.
Utility-scale combined-cycle plants sized for hyperscale campuses use F-class and H-class gas turbines (roughly 300–600+ MW per unit), which require single-crystal nickel-superalloy blades cast in vacuum furnaces, precision-forged rotor discs, and combustion sections qualified through yearslong test cycles. Unlike a chip fab, where a new line can in principle be replicated by procuring more lithography tools, large-frame turbine output is gated by a small number of forging presses and single-crystal casting facilities worldwide, owned almost exclusively by three OEMs: GE Vernova, Siemens Energy, and Mitsubishi Power. Expanding forging capacity itself takes years and heavy capex, making turbine supply structurally inelastic on the 3–5 year horizon that matters for the current buildout.
The order books show it. GE Vernova's gas-turbine backlog rose from 83 GW (end of 2025) to 100 GW (Q1 2026) to 116 GW (Q2 2026), and the company is guiding to at least 125 GW under contract by year-end 2026; annualized output is ramping from roughly 20 GW/year in Q3 2026 toward 24 GW by 2028 and 30 GW by 2030. Siemens Energy's backlog stands near 69 GW, with 90–100 GW of total commitments (orders plus slot reservations) expected by the close of fiscal 2026, and quoted lead times of three-plus years. Mitsubishi Power lead times now exceed five years, with contracts being signed for 2031–2034 delivery slots.
This has turned power hardware, not compute hardware, into the pacing item for AI infrastructure — and it's propagating into grid operations. ERCOT is tracking more than 1,800 large-load interconnection requests totaling over 474 GW, more than 5x its record peak demand, which prompted Texas Governor Greg Abbott's August 3 directive pausing new data-center interconnection approvals pending a state audit covering power sourcing, on-site generation, and water use.
Why it matters for tech + supply chain: chip supply used to set the pace of the AI race. Now it's whoever locked in a turbine order first — because the wait for the power plant is longer than the wait for the chip.
Why it matters for tech + supply chain: it reframes AI capex risk — wafer capacity is a multi-fab problem with real substitution paths, but heavy-frame turbine supply is a three-vendor, forging-limited problem on a multi-year capital cycle, pulling serious strategic weight toward whoever secured turbine slots and grid interconnection earliest.