Summary: BloombergNEF figures reported on April 27, 2026 show a new US combined-cycle gas plant became 66% more expensive to build over two years, rising from below $1,500/kW in 2023 to $2,157/kW in 2025. Data-center electricity demand is forecast to grow from about 40 GW to 106 GW by 2035. The article says power cost, turbine lead times, and community resistance are now first-order constraints on AI deployment.
New BloombergNEF (BNEF) figures, reported by TechCrunch on April 27, 2026, put a hard number on U.S. gas plant economics: a new combined-cycle gas turbine (CCGT) plant has become 66% more expensive to build over the past two years. Per-kilowatt construction cost moved from below $1,500/kW in 2023 to $2,157/kW in 2025, and projects now take 23% longer to complete. One force pushed all of this: electricity demand from AI data centers.
These figures turn the AI buildout from a chip story into a power story. Training and inference draw electricity faster than grids and plants can grow, so the availability and price of power now feed straight into what compute costs. Teams planning large-scale AI services after 2026 need no forecast — the next two procurement cycles will put this reality in front of them.
The 66%, Explained: From $1,500 to $2,157 per kW
BNEF measures what it costs to build each kilowatt of generating capacity. In 2023 a new CCGT sat below $1,500/kW; by 2025 the figure had climbed to $2,157/kW, a cumulative 66% rise. The schedule broke too: a new plant now takes 23% longer to finish. Against U.S. natural gas prices that have stayed low for years, the increase points almost entirely to scarce equipment and engineering capacity — not fuel.
In project terms, a 1 GW-class gas plant now runs about $660 million over a budget drawn up two years ago, and it waits in a longer line before ground breaks. AI loads need power at megawatt scale, delivered fast and predictably — three requirements that no longer arrive together.
The Tightest Choke Point: the Gas Turbine
At the root of the price spike sits the gas turbine. Wood Mackenzie’s April 1 report projected heavy-duty turbine prices at $600/kW by the end of 2027, up 195% from 2019, with turbines accounting for roughly 20–30% of a combined-cycle project’s total cost — a share that climbs higher for simple-cycle plants. The mismatch is stark: as of the end of 2025, global order books held about 110 GW against manufacturing capacity of just 60–70 GW, books are filled past 2027, and the longest lead times run six years.
Manufacturers are adding capacity, but not fast enough. GE Vernova has put more than $160 million into lifting large-frame turbine production from roughly 50 units a year to 70–80 by the end of 2026. Siemens Energy has moved key facilities to round-the-clock operation and is carrying out a $1 billion U.S. investment program. Mitsubishi Heavy Industries plans to have doubled its capacity by 2028. Where things truly bind is components: single-crystal blades and other hot-section parts come from only a few suppliers worldwide, and specialized workers are scarce.
One project conveys the scale of demand: SB Energy’s Portsmouth Powered Land scheme — 9.2 GW backed by $33 billion, announced in February 2026 — will take 24 to 30 heavy-duty turbines for its first phase alone. Wood Mackenzie also projects 63 GW of new U.S. gas capacity between 2026 and 2030, with the ordering peak landing in 2026.
On the Demand Side: 40 GW Heading to 106 GW
BNEF’s forecast for the data centers themselves is equally steep: consumption of about 40 GW today grows to 106 GW by 2035, a 2.7x expansion. The shape of facilities is changing too — only around 10% of data centers reach 50 MW or more today, yet within a decade the average facility will pass 100 MW. Wood Mackenzie projects data center electricity use up 96% between 2026 and 2031, the fastest-growing new load on the U.S. grid.
Why did cost and schedule break together? These numbers answer that. New load shows up in huge, lumpy blocks — a single hyperscaler campus can take an entire plant’s output — while the generation supply chain scales up on a rhythm of years and, at times, decades.
Gas or Renewables Plus Storage: The Industry’s Two Routes
Facing the shortfall, the tech giants split into two postures. Microsoft and Meta are accelerating gas plants of their own, placing generation right beside their data centers, and the Trump administration has publicly called on operators to “bring their own power” and lean less on the public grid. That route carries a political bill: once utilities fund gas units, the costs usually shift onto ordinary ratepayers, and the burden has already stirred a backlash — in local permitting fights, “data center” has turned into a loaded term.
Google bet elsewhere: renewables paired with long-duration storage. Its power playbook brings in Form Energy’s iron-air batteries — discharging for more than 100 hours, well past the 4-hour class of lithium — and wagers solar-plus-storage against gas. As solar and battery costs keep falling while turbine prices surge the opposite way, the crossover between the two routes arrives ahead of what most models assumed.
What Infrastructure and Product Teams Should Draw From It
Three concrete effects. One, acquiring power is now a first-order constraint on AI deployment: siting, latency, and launch-date plans must factor in turbine lead times and the interconnection queue alongside GPU availability. Two, costs pass through: owning plants, signing PPAs, or renting cloud GPUs all expose a project to rising power prices and community resistance, and where inference workloads sit decides your unit economics. Three, the contest between clean firm power with long-duration storage and gas will set the carbon and cost curve of new compute over the next five years. Reporting AI energy and emissions externally is moving from a marketing option to a compliance and community-relations requirement.