Power, Not Chips, Is Now the AI Buildout's Binding Constraint

Microsoft walked away from anchoring a £2bn data center in Essex because the local grid said it could not deliver power until the early to mid-2030s. Elon Musk's xAI has eight converted airliner engines bolted to the ground in Mississippi to keep its newest supercomputer fed. OpenAI's first Stargate campus in Abilene, Texas, has been running on gas turbines since last year.

None of that is how data centers are supposed to work. It is what the buildout looks like once the constraint moves from silicon to electrons.

What Happened: The Grid Said No

According to TechCrunch's report on New York Times findings, a growing number of AI data center developers are making their own electricity with jet engines pulled from retired aircraft and rebuilt to burn natural gas. The units come from ProEnergy, which converts a GE engine commonly found on Boeing jets.

The appeal is lead time, not elegance. Rystad Energy analyst Audun Martinsen put the pitch plainly: with jet-powered units, "they can have a data center running in 24 months." Hooking up 100 megawatts to the grid in Northern Virginia, by contrast, takes roughly seven years according to the International Energy Agency, and a Mitsubishi Power executive told Bloomberg last week that most new capacity is not expected before 2031.

The roll call is already substantial. xAI installed eight converted jet-engine turbines in Mississippi for Colossus 2, per state permits and satellite imagery cited by the Times. Crusoe, building OpenAI's Abilene campus, ordered 13 ProEnergy units in April, after OpenAI had already deployed gas turbines there, then walked away in September from a $1.25 billion deal for 29 turbines from Boom Supersonic.

Then came the clearest demonstration that power beats everything else. Nscale's £2bn Loughton site, billed as the UK's largest AI data center, lost Microsoft as its anchor tenant after UK Power Networks said it could not supply enough electricity until the early to mid-2030s. Nscale says it remains committed and is exploring alternative power sources. Microsoft's exit is what an interconnection delay looks like once it reaches a balance sheet.

The trade-offs are real. Converted engines are less efficient and pollute more than turbines built for conventional gas plants, and it takes more of them to produce the same output, which makes converted jets a bridge rather than a power strategy.

Why It Happened: Queues, Not Shortages

The problem is not that America lacks generating capacity. It is that the interconnection queue has become a market of its own.

In Texas, ERCOT's queue held more than 1,800 projects representing over 474 gigawatts of connection requests as of August, more than five times the state's record peak demand, with roughly 90 percent of those requests coming from data centers. Governor Greg Abbott responded with a moratorium on new data center grid connections and an audit of every project in the queue. The pause explicitly did not apply to projects building their own on-site generation.

That exemption is the whole story. Research from CleanView cited by Ars Technica counted 59 data centers with roughly 90 gigawatts of planned capacity building power behind the meter, about a quarter of all planned US projects. xAI alone accounted for most of the roughly 2 gigawatts running off-grid as of mid-2026.

Hardware compounds it. Bloomberg reported that almost half the US data centers planned for 2026 were expected to slip or be canceled because developers cannot import enough transformers, switchgear and batteries, with lead times stretching to five years. Satellite analysis by the Financial Times and SynMax put nearly 40 percent of this year's projects behind schedule.

The Verge's reporting on a contested Utah project describes how the jam cascades: speculative interconnection requests jam the queue, wait times grow, builders go behind the meter, turbines get backordered for years, and a market appears for "powered land assembly" where sites with secured power sell for 20 times the raw land value. What buyers pay for is speed.

The financial mirror of that scarcity sits in Anthropic's updated S-1. Compute consumed 58 percent of revenue in the quarter ending September 30, up from 48 percent, on a $10 billion quarterly compute bill that included $5 billion under short-term commitments the filing says cost "significantly more than comparable compute procured under our multi-year commitments." Revenue per dollar of compute peaked above 2x and fell back to 1.72x. Scarcity now shows up in gross margin, not just queue positions.

Why It Matters: Three-Year Bets on Thirty-Year Assets

Converted jet engines are installed to buy time, but the infrastructure around them is financed on decade-long assumptions. That mismatch is the strategic risk.

First, capacity geography becomes a product decision. If a site can be live in 24 months with on-site gas but needs seven years for a grid connection, deployment follows the electrons rather than the users. Regions with permissive permitting and available gas become de facto cloud regions, and assumptions about latency, data residency and failover have to be rebuilt around power topology instead of network topology.

Second, the workaround is politically fragile. New York imposed a one-year moratorium on data centers of 50 megawatts or more, Maine moved toward pausing construction, and Texas halted grid connections to audit the queue. Burning gas in converted aircraft engines near populated areas is exactly the practice that attracts regulation once a community connects the noise and emissions to the facility next door, and experts quoted by the Times expect that shift within a year or two, shorter than the depreciation schedule of the hardware inside.

Third, inference inherits a power floor. Compute bought under short-term commitments at a premium does not get cheaper when models do; the premium gets passed through as a capacity charge and compresses margins for everyone reselling inference on top.

Builder Impact

Capacity contracts matter more than benchmarks. The differentiator between providers is no longer which model they host but where their power comes from and what happens to your workload when the fuel supply or permit is challenged. Ask for the site, generation source and interconnection status, not just GPU count.

Design for interruptible work. Behind-the-meter gas sites and grid-constrained regions will increasingly expose price and curtailment signals. Agent fleets that checkpoint, pause and resume are cheaper than workloads that must finish in real time, and scheduling batch jobs against power economics is becoming a real optimization lever.

Treat neocloud risk as a first-class dependency. The flexibility that makes neoclouds fast also makes them exposed: single-fuel dependencies, concentrated sites, permitting risk, short-term pricing. Multi-provider routing is a continuity strategy, not just a cost one.

Do not assume inference prices fall in a straight line. Demand for open-weight hosting is surging, but margins stay thin because compute is expensive to lease. Power scarcity limits how far unit economics can improve, so model self-hosting against real capacity costs, not list prices.

Watch the permitting calendar, not just model releases. The next constraint on what you can deploy will be decided in utility commissions, state legislatures and environmental reviews, and those dates matter more than most product announcements.

The buildout spent two years treating compute as the scarce input. It is now discovering that compute is downstream of a grid connection, and that the fastest available power source is a retired airliner engine burning gas in a field.