The gas turbine shortage behind AI data centers, explained: backlogs, lead times and who is waiting

Why can't AI companies just plug into more power? Often the answer is a machine: the large gas turbine, built by a handful of companies, with a delivery queue that now runs to the end of the decade. This explainer sets out who makes the turbines, how big the queue is, why the numbers you see are hard to compare, and what developers do about it.

The short version

Three manufacturers, GE Vernova, Siemens Energy and Mitsubishi Power, supply over 75% of large gas turbines, according to Turbomachinery Magazine, and their delivery slots reach late 2028 at the earliest, with many committed into 2030 or later. A developer who wants to power a new AI campus with gas either joins that queue, buys a smaller machine, or finds a used one.

The queue in numbers

OilPrice compiled the makers' latest disclosures:

• GE Vernova: 116 GW of gas power backlog and slot reservations at the end of Q2 2026, up from 100 GW three months earlier and 83 GW at the end of 2025. Only 53 GW is firm equipment backlog; the other 63 GW is slot reservations. The company expects at least 125 GW under contract by December. On its July 22 earnings call it was taking orders for 2031 delivery. Production runs at about 20 GW annualized, rising to 24 GW by 2028 and toward 30 GW by 2030.
• Siemens Energy: a 69 GW gas turbine backlog at June 30, 2026, after booking 15 GW and shipping 6 GW. Lead times are three years or more. Its chief executive, Christian Bruch, said the addressable market could reach 120 GW a year, roughly half American.
• Mitsubishi Heavy Industries: a 35 GW large frame backlog reported August 6, 2026, up from 23 GW a year earlier. Orders booked in the quarter deliver in 2028 to 2030, and CFO Hiroshi Nishio said the company is "being selective in the projects we contract."

The three headline figures add to about 220 GW, but OilPrice notes they count backlog differently, so the sum overstates the total. A "slot reservation" is a place in line, not a signed equipment order. That distinction matters for everything below.

Demand versus capacity

OilPrice cites Goldman Sachs figures from May 2026: US data center power demand rising from 31 GW in 2025 to 41 GW in 2026 and 66 GW in 2027. Capacity additions were 8.5 GW realized in 2025, with 13.6 GW scheduled for 2026 and 36.3 GW for 2027, and Goldman assumes only 50% to 60% of capacity scheduled over the next two years arrives on time. Global gas turbine orders hit a record 38 GW in Q2 2026, up 71% on the year, with the US about half. Wood Mackenzie puts worldwide manufacturing capacity at 60 to 70 GW a year against roughly 110 GW of orders.

Turbomachinery Magazine adds the demand side from Siemens Energy's Richard Reisig: about 25% of its gas turbine order book is now tied to data centers, "well into 2035." Other figures vary by source, and one outlet we reviewed gave a different share. We use the interview figure.

Why makers cannot just build faster

GE Vernova is investing $300 million in South Carolina and New York and targets 20 GW of annualized output by Q3 2026 and 24 GW by 2028, per Turbomachinery Magazine. Its Ihab Chaaban said the expansion "will not immediately shorten customer lead times" because it mainly serves existing backlog. Siemens Energy's Reisig says scaling is constrained by skilled labor, materials and component complexity. Baker Hughes's NovaLT line is effectively sold out through 2028.

Costs have risen too. BloombergNEF data cited by Turbomachinery Magazine puts combined cycle plant costs at about $2,157 per kilowatt in 2025, up from under $1,500 in 2023, and says time to bring a new plant online rose 23%.

What developers do instead

Three workarounds are in use. First, reserve early and hold deposits, which is the route that cost SoftBank about $3.7 billion when it cancelled orders it no longer needed, as reported by Semafor. Second, buy smaller or secondhand machines: Semafor says developers are increasingly seeking second hand equipment, and Turbomachinery Magazine reports Hallador Energy bought 460 MW of unfired Siemens turbines in June 2026. Third, convert other engines, as in our report on converted jet engines powering AI data centers. Anthropic's order of Caterpillar gas turbines for 1.7 gigawatts is another route, which we covered in our report on Anthropic's turbine purchase.

Turbomachinery Magazine says industry estimates suggest data centers could rely on about 35 GW of behind the meter power by 2030, mostly natural gas, and that natural gas is expected to supply roughly half of new data center power through 2030.

The grid side

Turbines are not the only bottleneck. In the PJM capacity auction on July 14, 2026, cited by OilPrice, capacity for 2028 and 2029 cleared at $325 per megawatt day, the FERC approved cap, for a third consecutive year. It procured 138,318 MW, 6,831 MW short of the reliability requirement, with only 525 MW of new generation and uprates clearing.

What a skeptic should ask

Gigawatt headlines mix reservations, orders and permits. Siemens Energy's chief executive told Semafor: "Some of the growth rates for data centers were unrealistic, it's just that nobody wanted to hear it." That cuts both ways: demand growth may be smaller than the loudest forecasts, and still larger than supply.

Our take

The turbine queue is the clearest physical limit on the AI buildout in the near term. It is why behind the meter plants, secondhand equipment and odd workarounds keep appearing in announcements. It is also why a project with no named delivery slot deserves caution. When you read that a campus will draw gigawatts of gas power, ask three questions: who holds the slot, which year does it deliver, and does the site have its permit? Only a yes to all three means a power plant rather than a press release. For the broader site by site picture, see our AI data center backlash tracker.