Saturday, November 21, 2026 AI in the real economy Get the Brief

Gas turbine shortage: what AI data centers now pay per kilowatt of power

GE Vernova holds 122 GW of gas equipment under contract against 20 GW a year of output. We compared four gas power deals: existing plants trade near $730 to $1,174 per kW, while Anthropic's reported generator deal works out to about $2,000.

Gas turbine shortage: what AI data centers now pay per kilowatt of power
Image: GE Vernova

GE Vernova now holds 122 gigawatts of gas power equipment under contract or slot reservation. At the 20 GW a year its factories can build today, that is about six years of output. That single ratio explains why AI companies are paying more for a kilowatt of gas power than they ever planned to, and why the cheapest megawatt in America right now is one that already exists.

This week gave us three fresh price points. The Financial Times reported that Anthropic will pay $3.4 billion for 1.7 GW of gas generators by taking over orders placed by OnePWR Solutions. NRG agreed to sell 4.9 GW of operating gas plants to LS Power for $4 billion. And Reuters reported that data center developers are hunting used turbines and retired jet engines because new units take about six years to arrive. Put those deals side by side and a clear pattern shows up.

The claim

The binding constraint on AI power is no longer money or even grid permission. It is turbine factory output. Gas equipment makers have sold years of future production, so buyers are now paying in one of three currencies: a premium for a place in line, a premium for equipment someone else already ordered, or a lower price for old plants that are already spinning. Existing gas plants are changing hands at roughly 35% to 60% of what Anthropic’s reported deal implies per kilowatt.

The evidence

The order book outran the factory

GE Vernova is the clearest window into the turbine queue because it reports the number every quarter. In its third quarter 2026 release (October 28), the company said it signed 11 GW of new gas equipment contracts in the quarter, made up of 8 GW of slot reservation agreements and 3 GW of firm orders. Gas Power equipment backlog plus slot reservations grew from 116 GW to 122 GW, and the company now expects to reach at least 130 GW by the end of 2026. Total company backlog hit $197 billion. Orders were $29 billion, up 96% organically, and revenue was $12.6 billion, up 27%.

For context, that same measure stood at 83 GW at the end of 2025, according to GE Vernova’s fourth quarter 2025 release. If it hits 130 GW, the queue will have grown by 47 GW in a single year.

Now compare that with supply. In its second quarter 2026 release, GE Vernova said it was on track to deliver 20 GW of annual gas turbine output in the third quarter of 2026, 24 GW in 2028, and that it is taking actions to produce 30 GW in 2030.

GE Vernova gas queue GW under contract or reserved Annual build rate used Years of output in the queue
End of 2025 83 20 GW 4.2
End of Q3 2026 122 20 GW 6.1
Target, end of 2026 130 24 GW (2028 rate) 5.4

Source: GE Vernova Q4 2025, Q2 2026 and Q3 2026 earnings releases. Years of output is Benderson Media’s calculation (queue divided by annual build rate). It is a rough guide: the queue includes reservations that may not all convert to orders, and the company also ships other gas equipment.

Even with the factory running at the 2028 rate, the year end queue would be more than five years of production. The queue grew by 47 GW in 2026 while the factory builds 20 GW a year, so the line is getting longer faster than it is being served.

The rest of the industry looks the same. Wood Mackenzie reported in April that global gas turbine orders sat at 110 GW at the end of 2025 against manufacturing capacity of 60 to 70 GW a year, with six year lead times and order books sold through 2027. It put the rise in gas turbine prices at 195% since 2019 and expects prices to reach $600 per kW by the end of 2027. That figure covers turbine equipment only, not a finished power plant.

The Reuters report on November 9, as summarized by Startup Fortune, described the same squeeze: orders above 100 GW, 60 to 70 GW of annual manufacturing capacity, and about six years on average for a new unit in the US. Wood Mackenzie analyst Aurora Tenorio told Reuters that delivery talks are now “firmly focused on the 2030 to 2031 window,” according to TFTC’s account of the same report.

What a kilowatt costs, deal by deal

Here is the comparison that matters. These are four gas power transactions announced or closed in the last three months, with the price divided by the capacity.

Deal Date Capacity Price Price per kW What the buyer gets
Vistra buys Cogentrix (10 plants) Closed Oct 30, 2026 About 5,500 MW About $4.0 billion (net) About $730 Operating gas plants
LS Power buys six NRG plants Announced Nov 9, 2026 4.9 GW $4 billion About $816 Operating gas plants
Constellation buys Rhode Island State Energy Center from Shell Announced Sept 10, 2026 Up to 609 MW $715 million About $1,174 One operating combined cycle plant
Anthropic takes over OnePWR’s Caterpillar generator orders Reported Nov 8, 2026 1.7 GW $3.4 billion About $2,000 Generators already in the queue, plus engineering and construction

Sources: Vistra closing release and Vistra deal announcement (Vistra states about $730 per kW, net of about $0.7 billion of expected tax benefits); NRG release; Constellation release; Financial Times report as summarized by The Next Web. Price per kW for NRG, Constellation and Anthropic is Benderson Media’s calculation (price divided by capacity). Constellation also states a price of about $580 million net of first year tax benefits, about $952 per kW.

Read the table from top to bottom and the price per kilowatt nearly triples. The top three rows are operating power plants. The bottom row is generators plus the engineering and construction to put them to work, and Anthropic is paying the most per kilowatt for it. The companies declined to comment to the FT, so the exact split between equipment and construction is not public.

The comparison is not apples to apples, and that is the point. An older combined cycle plant on the grid and a new fleet of reciprocating gensets at a data center campus are different assets. What the table shows is what each buyer was willing to give up. The power companies are paying for existing iron at a discount to replacement cost. The AI company is paying a premium to skip years in line.

Everyone is buying the line, not the machine

The Anthropic deal is the most visible example, not the only one. Tom’s Hardware notes the OnePWR orders were placed with Caterpillar roughly two years ago. That is exactly what made them valuable: whoever placed an order in 2024 now owns something that cannot be bought new for delivery before the end of the decade.

Others are going around the heavy turbine queue entirely:

  • Refurbished jet engine cores. Crusoe contracted ProEnergy for 13 PE6000 aeroderivative gas turbine generator sets, 50 MW each, 650 MW in total, for delivery by summer 2027, according to ProEnergy’s April release. The PE6000 is built around overhauled CF6 engine cores.
  • Reciprocating engines in bulk. Caterpillar announced in January that American Intelligence and Power will deploy 2 GW of dedicated power at its Monarch Compute Campus in West Virginia using Cat G3516 natural gas gensets, with deliveries from September 2026 through August 2027 (Caterpillar release).
  • Gas engines from Europe. Wärtsilä has booked about 3 GW of US data center orders for gas engine units since spring 2025, per Reuters as reported by TFTC. Its energy business president summed up the market as “lead times, lead times, lead times.”
  • Behind the meter at scale. Omdia projects data centers could deploy 35 GW of behind the meter power by 2030, according to Power Engineering.

The common thread: smaller, faster units that can be bought from a different factory, even at a worse heat rate. Speed to power beats efficiency when a data hall full of GPUs is sitting idle.

What it means on the ground

The turbine factory is now part of the AI supply chain. Chips get the headlines, but a GPU cluster with no power is a very expensive warehouse. The constraint has moved down the chain to castings, forgings, hot section blades and the specialized machine shops that finish them. Tom’s Hardware reports that some turbine blades take 60 to 90 weeks to make.

Brandon’s Take: From the tooling side, this is the part people outside manufacturing underestimate. You cannot add turbine capacity by writing a bigger check. Every new production line needs fixtures, molds and dies for investment casting, qualified suppliers and trained people, and each of those has its own lead time stacked on top of the next. When GE Vernova says it is working toward 30 GW in 2030, that four year runway reflects how long it takes to build and qualify the tooling and the supplier base behind a precision part that runs at those temperatures. The companies that win the next three years will be the ones that locked in suppliers early, not the ones with the most cash today.

Old plants just got repriced. Vistra paid about $730 per kW for modern operating plants. NRG is selling at about $816. Wood Mackenzie expects turbine equipment alone to hit $600 per kW by 2027, before the cost of the rest of a plant. When replacement cost keeps climbing and replacement time is six years, existing plants look cheap by comparison. Expect more owners to buy, sell and recontract gas fleets, and expect long term contracts with data center buyers to be the reason.

Buyers will trade efficiency for time. Aeroderivatives, reciprocating gensets and refurbished jet cores burn more fuel per megawatt hour than a big combined cycle plant. Data center developers are accepting that because the alternative is waiting. That shifts work toward packagers, rental fleets, overhaul shops and gas engine makers, and it puts more weight on local gas supply and air permits.

Procurement now means slot management. GE Vernova sold 8 GW of slot reservations in one quarter, more than its 3 GW of firm orders. Anthropic bought someone else’s slots. The practical lesson for anyone buying long lead industrial equipment is the same one tooling buyers learn early: the reservation is the asset, and the paperwork that secures it is worth as much attention as the spec.

What to watch

  1. Siemens Energy and Mitsubishi Heavy results. Both, along with GE Vernova, have said AI driven orders exceed their production capacity, per TFTC. Their next reports will show whether their queues look like GE Vernova’s.
  2. GE Vernova hitting 130 GW. The company guided to at least 130 GW by year end. Watch how much of the increase is slot reservations versus firm orders. Reservations can be cancelled or resold; orders are harder to walk away from.
  3. The NRG and LS Power close. The sale is targeted to close in the second quarter of 2027 and needs HSR, FERC and New York PSC approval. A long review would show how regulators view gas fleets changing hands during a demand boom.
  4. More order book transfers. If Anthropic paid about $2,000 per kW for someone else’s queue position, expect other holders of 2024 and 2025 orders to test what their slots are worth.
  5. Air permits. Fast onsite generation still needs permits. Challenges to temporary turbine fleets, like the one Tom’s Hardware describes against xAI’s Mississippi site, could slow the workaround.

The short version: AI is paying for gas power with time, not just money. The price of a kilowatt now depends mostly on when you can have it.

Sources

This is commentary on public information, not investment advice.

Get The Benderson Brief in your inbox every weekday morning. Follow @bendersonmedia on X for the daily charts.

Brandon Henderson

Brandon Henderson founded Benderson Media. He has spent more than 15 years in custom injection molding, tooling and global manufacturing, working with factories in China and Vietnam on programs for aerospace and consumer brands. He is a journeyman tool and die maker, CPSP certified, and a Six Sigma Green Belt. Benderson Media covers AI in the real economy: the factories, chips, power and money behind it.

Get The Benderson Brief

Five stories on robots, chips, power and AI money every weekday morning. Free. One click to leave.

Every figure in this piece links to its source. See our editorial standards. Nothing here is investment advice.

Keep reading