The Gas Turbine Shortage Just Became AI’s Biggest Constraint

The Gas Turbine Shortage Just Became AI’s Biggest Constraint

Michael Kern

Sun, August 23, 2026 at 12:00 AM GMT+3 10 min read

• GS +3.73%

Order a heavy-duty gas turbine from GE Vernova today and it won't arrive until 2031. That's the company's actual production schedule, confirmed on its July 22 earnings call, and it's the fact sitting underneath every AI data center power plan announced in the last two years. It comes up far less often than the announcements do.

Goldman Sachs put hard numbers on the demand side in May. U.S. data center power demand climbs from 31 gigawatts in 2025 to 41 GW this year and 66 GW in 2027. Year-over-year capacity additions accelerate from 8.5 GW actually realized last year to 13.6 GW scheduled for 2026 and 36.3 GW scheduled for 2027. By then data centers would take 8.5% of total U.S. peak summer demand, up from 4.1% today.

The equipment meant to generate that power is on a completely different clock.

The Big Three Are Booked Solid Into the 2030s

GE Vernova closed the second quarter with 116 GW of gas power equipment backlog and slot reservation agreements, up from 100 GW three months earlier and 83 GW at the end of 2025. It expects at least 125 GW under contract by December. CEO Scott Strazik told analysts the company is taking reservations for 2031 delivery and should be more than halfway contracted for that year by the end of 2026. Its production plan: roughly 20 GW annualized this quarter, 24 GW by 2028, and a push toward 30 GW by 2030.

Siemens Energy ended its fiscal third quarter on June 30 with a 69 GW gas turbine backlog after booking 15 GW and shipping six. Lead times run three years or more. CEO Christian Bruch told analysts the addressable market could reach 120 GW a year, roughly half of it American.

Mitsubishi Heavy Industries reported a 35 GW large-frame backlog on Aug. 6, up from 23 GW a year earlier. That figure covers Mitsubishi's fiscal first quarter, which runs March through June, so it lags roughly a month behind the calendar-quarter numbers above. CFO Hiroshi Nishio said orders booked during the quarter are scheduled for delivery between 2028 and 2030, and that the company is "being selective in the projects we contract."

Add the three headline numbers and you get 220 GW, which overstates the case, because none of them is counting the same thing. Of GE Vernova's 116 GW, only 53 GW is firm equipment backlog; the other 63 GW is slot reservations, paid options that haven't converted to orders yet. Siemens' 69 GW is firm backlog with no reservations mixed in. Mitsubishi's 35 GW covers large-frame turbines only, leaving out its aeroderivative and mid-size lines. None of the three would explain what the other two mean by "backlog" without a footnote.

What all three agree on, regardless of how each counts it, is direction and scale. Global orders in the second quarter hit a record 38 GW, up 71% year on year, with the U.S. accounting for half. Wood Mackenzie puts worldwide manufacturing capacity at 60 to 70 GW a year against roughly 110 GW of orders.

One Year of AI Demand Outstrips the Entire Industry's Output

That 36.3 GW figure from Goldman is the one that matters most, because it's larger than what the entire turbine industry can plausibly deliver in a single year. Everything GE Vernova builds annually, worldwide, for every customer on the planet, is about 20 GW. Only a fifth of the gigawatts under contract are earmarked for data centers in the first place. The rest goes to utilities replacing coal, Middle East power and desalination projects, industrial load and grid reliability work queued up long before anyone was talking about inference clusters.

Even in a fantasy where GE Vernova pointed its entire global output at American AI campuses, it wouldn't cover one year of the Goldman curve. And it couldn't, because those slots were sold years ago.

The turbines running data centers in 2027 were ordered in 2023 and 2024. Turbines ordered this month land somewhere around 2030 or 2031. Everything in between is either already spoken for, coming from something other than a new gas plant, or not happening.

This is also the part of the supply chain that scales worst. Hot-section castings, the blades and vanes that sit in a 1,000-degree gas path, come from a thin bench of specialized foundries, and the welders and machinists who assemble these machines were laid off a decade ago and never replaced. Even where a slot exists, there aren't enough qualified contractors to install the thing. Average lead times for a new combined-cycle plant have gone from three and a half years in 2023 to roughly five now, and seven for some heavy-duty frames.

PJM Paid Top Dollar and Still Came Up Short

PJM's capacity auction on July 14 is the clearest read available on what the gap costs.

The auction for the 2028/2029 delivery year cleared at $325 per megawatt-day, the FERC-approved cap, for the third consecutive year. It procured 138,318 MW of unforced capacity, which left the system 6,831 MW below PJM's own reliability requirement. New generation and uprates that cleared: 525 MW.

The largest wholesale power market in the country paid the maximum price its rules allow, came up nearly 7 GW short and attracted about half a gigawatt of new supply in return. PJM's own modeling showed that without the cap, the RTO would have cleared at $554.72 per megawatt-day and the ComEd zone at $776.69.

PJM has proposed a one-time Reliability Backstop Procurement targeting roughly 14.9 GW of new capacity, split between bilateral contracts starting this September and a pay-as-bid auction next March. Whether that fills a hole gas turbine makers can't touch inside three years is the open question.

Turbine Prices Have Nearly Doubled, and Gas Isn't the Reason

Prices have moved accordingly. BloombergNEF put the average combined-cycle project at $2,157 per kilowatt last year, up from under $1,500 in 2023. A GridLab review of actual project filings found plants due to finish in 2026 and 2027 reported $1,116 to $1,427 per kW, while projects targeting 2030 and 2031 routinely report $2,000 per kW or more. Wood Mackenzie expects turbine prices alone to reach $600 per kW by the end of 2027, a 195% jump from 2019.

The fuel itself has nothing to do with it. Henry Hub spot gas was trading around $2.79 per MMBtu in mid-August, and government trackers put the 2026 average near $3.31, both unremarkable by the last five years' standards. The bottleneck sits entirely in castings, welders and factory slots, not in the molecule being burned.

GE Vernova is collecting on the hardware side of it. First-half 2026 equipment orders were priced more than 20% above fourth-quarter 2025 orders. Second-quarter free cash flow hit $5.1 billion, helped by a $6.4 billion working capital benefit from customer down payments tied to slot reservations. Year-to-date free cash flow is running near $9.9 billion, already more than the company generated in all of 2025.

GE Vernova Is Getting Paid for Turbines It May Never Build

That down-payment structure is also the smartest thing the turbine makers have done this cycle, and it comes straight out of the last one.

All three lived through the last gas boom. Cheap shale in the 2000s pulled in a wave of merchant developers, the market saturated, orders were canceled, factories closed and skilled labor scattered. The industry spent the 2010s treating gas turbines as a cash cow rather than a growth business, and capacity stayed flat while electricity demand quietly started climbing again.

Nobody wants that hangover twice, so rather than bet the factory on hyperscaler forecasts, the OEMs built an instrument that makes the customer carry the risk. A slot reservation agreement is a paid option on future production. If the data center gets built, the customer converts it to a firm order. If it doesn't, the manufacturer keeps the deposit and resells the slot into a market that's short anyway.

GE Vernova's balance sheet is currently being financed, in part, by down payments on turbines that may never be needed. It's a good position to be in, and it explains why the capacity expansions are so measured relative to the size of the order book.

Exelon and Texas Are Quietly Walking Back Their Numbers

There's an honest counterargument here, and it keeps getting stronger.

Exelon cut its "high probability" data center load from roughly 18 GW to about 11 GW on July 30, a drop of nearly 40%. Its broader interconnection pipeline fell from about 43 GW to 25 GW in a single quarter. Only about 4 GW carries signed transmission service agreements backed by $1 billion in posted collateral. "We now weed out speculative projects," CFO Jeanne Jones said, "and it gives us proactive insight into what is real."

Texas went further. Gov. Greg Abbott ordered an audit of every data center in the ERCOT interconnection queue on Aug. 3 after requests swelled to roughly 474 GW, about 90% of it data centers and more than five times the state's record peak demand. ERCOT suspended its Batch Zero large-load process within hours and, on Aug. 10, formally asked the Public Utility Commission for a good-cause exception covering the missed classification deadline; the audit itself is expected to run several months, with scope and timing due at the PUCT's Aug. 20 open meeting. BloombergNEF estimates the pause puts nearly 50 GW at risk of delay, about a fifth of the national pipeline. New York had already halted new approvals in July.

Goldman itself assumes only 50% to 60% of the data center capacity scheduled over the next two years shows up on time.

Read the 116 GW order book one way and it's the leading edge of the biggest power buildout since the 1970s. Read it the other way and it's a queue stuffed with options on projects that were never financed, held by developers who'd rather forfeit a deposit than miss a slot.

This Bottleneck Gets Solved by Governors, Not Factories

It barely matters which reading is right, because the 2027 timeline holds either way.

If the demand is real, the turbines can't arrive fast enough and the shortfall gets covered by whatever can be delivered inside 24 to 36 months, mostly reciprocating engines. That's why Caterpillar and Wärtsilä are now selling prime power into AI campuses, and why their own order books are full into 2028 too. Fuel cells fill in around the edges. Existing plants run harder. Coal units that were supposed to retire stay open instead. And a meaningful share of the load that's currently on paper simply doesn't get built.

If the demand isn't real, the correction shows up as canceled slot reservations in 2028 and 2029, long after this year's capacity auctions have already repriced power for 67 million people in PJM.

Either way, the bottleneck doesn't clear on a factory floor. It clears in state capitals: governors in Austin and Albany, regulators writing separate rate classes for data centers, utilities demanding collateral before they'll even run an interconnection study. Almost nobody selling the AI buildout has priced in how much of that demand will have to shrink before the equipment can catch up.

By Michael Kern for Oilprice.com

More Top Reads From Oilprice.com

Oilprice Intelligence brings you the signals before they become front-page news. This is the same expert analysis read by veteran traders and political advisors. Get it free, twice a week, and you'll always know why the market is moving before everyone else.

You get the geopolitical intelligence, the hidden inventory data, and the market whispers that move billions - and we'll send you $389 in premium energy intelligence, on us, just for subscribing. Join 400,000+ readers today. Get access immediately by clicking here.