HWNEWS
From jet engines to orbital data centers: AI's distributed-power stress test
A 100 GW turbine order book and five year lead times have turned AI's energy gap into a procurement crisis.
A 100 GW turbine order book and five year lead times have turned AI's energy gap into a procurement crisis.
The constraint is structural. Large natural gas turbines now carry lead times of more than five years, and smaller units run 18 to 36 months. Average prices moved from roughly $2,000 per kilowatt to about $3,000 per kilowatt in six months, per EPRI data cited in Canary Media's 12 June 2026 survey of AI power strategies. Global gas turbine orders doubled in 2025 to 846 units and 100.3 gigawatts, and 70 percent of units sold were under 100 MW. The order book is no longer cyclical. It is a supply chain that has run out of slack.
That pressure is forcing AI companies to do what utilities and grid planners assumed only they would do: build the power plant themselves. Cleanview counts 59 planned U.S. data center facilities pursuing behind-the-meter generation — representing more than 25% of all planned U.S. data center capacity — as analyzed by Cleanview. The substitution is no longer optional. It is the only path to bringing a campus online inside a financing window.
The on-site generation tier is the spine of the story, because this is where real megawatts are being purchased. Crusoe has signed a $1.25 billion agreement for 29 jet-engine turbines from Boom Supersonic, repurposing modified supersonic engines as stationary generators. Wärtsilä, the Finnish marine-engine maker, has booked a 790 MW order for a Texas data center, with equipment scheduled for delivery in 2028 and full operational status expected in late 2029. Bloom Energy and Oracle have announced a fuel cell partnership of up to 2.8 GW, sized for Oracle's cloud buildout, per Bloom Energy's investor relations press release. Each of these is a procurement contract, not a press release, and each is sized to add firm capacity inside 36 months, assuming gas supply, air permits, and cooling water hold.
The second tier is conventional on-site gas generation, real but unfinished, wrapped in its own permitting and grid questions. xAI's Memphis site, running roughly 50 unpermitted gas generators on flatbed trailers, has become the regulatory test case for what "behind-the-meter" actually means when a generator is mobile, temporary, and never studied in an interconnection queue. The consequences are visible. Local moratoriums in Seattle and Flint, Michigan have paused new data center builds. Texas Governor Greg Abbott has recommended that new facilities cover their own interconnection and upgrade costs. Ohio, Delaware, and Pennsylvania are pushing bring-your-own-power mandates that put the generation obligation on the developer. None of these are anti-AI measures on their face. They are utilities and regulators catching up to a load that arrived faster than planning cycles.
The third tier is where skepticism earns its keep. Panthalassa has raised $140 million for wave-powered floating data centers, a category that has not yet proven the wave-energy half of the equation at scale. DeepGreen has proposed the Western Passage, an underwater AI data center near Eastport, Maine, drawing on cold deep water for cooling but raising new questions about cable landing, marine permitting, and the difference between a pilot and a fleet. Meta and Overview Energy have discussed space-based solar beamed to Earth via infrared laser, a concept that has lived in agency white papers for two decades. Elon Musk has floated colocating data centers in orbit, with the implicit assumption that launch cost and radiative cooling close on a curve the industry has not yet shown. None of these should be dismissed, and none should be confused with the second tier. They are announcements, branding, or research theater. Treating them as the same kind of thing as a Crusoe turbine order is exactly the error a flat catalog invites.
The harder question, the one grid planners and project finance teams are actually trying to answer, is which of these approaches adds firm, permitted, financeable megawatts to U.S. data center supply in the next three to five years. The answer, on present evidence, is the first tier, with caveats: jet engines and fuel cells can ship inside 36 months if gas and air permits cooperate. The second tier is constrained less by technology than by local politics and interconnection cost allocation. The third tier is buying optionality, not capacity.
What to watch next is the Bloom-Oracle 2.8 GW ramp, the Wärtsilä Texas commissioning date, and whether Crusoe's jet-engine turbines clear their first emissions test without modification. Those are the data points that will tell the market whether the on-site generation playbook is a bridge to a grid that catches up, or a permanent reordering of who builds power in the United States.