In Northern Virginia, connecting 100 megawatts from the grid takes roughly seven years. AI companies cannot wait. So they buy retired jet engines from passenger aircraft and retrofit them into natural-gas-powered generators—installing them directly beside their own data centers to generate electricity on-site. According to a November 3 report by The New York Times, this has become a legitimate business.
This is not an emergency workaround by some startup. Among those deploying this approach are xAI, OpenAI’s Stargate Phase I, and Crusoe—the firm building their campuses.
Specific deployments already underway
- xAI: Eight retrofitted gas turbines installed in Mississippi to power the Colossus 2 data center in Memphis. Equipment supplied by ProEnergy; engines modified from GE models commonly used on Boeing passenger jets.
- OpenAI: The first Stargate site in Abilene, Texas, scheduled for 2025, will deploy gas turbines.
- Crusoe: Crusoe, responsible for building the Abilene campus, placed an order in April for 13 retrofitted ProEnergy units.
- Boom Supersonic: Its core business is supersonic passenger aircraft, yet it also sells the power-generation version of the same engine to data centers. In September, Crusoe canceled a $1.25 billion order for 29 units but stated it would continue using gas turbines.
Why jet engines: the time equation
The core rationale boils down to one word: speed. As Audun Martinsen, analyst at Rystad Energy, put it: “It’s all about getting electricity as quickly as possible.” Retrofitted jet-engine unitscan bring a data center online in 24 months. By contrast, the International Energy Agency estimates that connecting 100 megawatts to the grid in Northern Virginia takes about seven years. Last week, a Mitsubishi Power executive told Bloomberg that most new power-generation capacity will not come online until after 2031.
Waiting is already causing real losses. Last month, Microsoft withdrew from Nscale’s Loughton, UK, data center project, declining to serve as anchor tenant—because the local utility indicated it could not supply sufficient power before the mid-2030s.
Policy direction is accelerating this shift: the Trump administration is encouraging data centers to solve their own power needs rather than queue up for grid connections.
The trade-offs: dirtier, costlier, and more numerous units
Retrofitted aircraft engines are not anyone’s idea of a proper power plant. Their shortcomings are clear:
- Lower efficiencyCombustion and heat recovery design is not optimized for long-term continuous power generation, resulting in lower electricity output per unit of fuel.
- Higher emissionsFor the same power output, pollutant emissions exceed those of standard natural gas power plants.
- More units requiredAchieving the same power output requires a greater number of turbine units, amplifying maintenance and noise issues.
This also sharpens environmental controversies. Our site previously reported on the xAI data center case involving 35 gas turbines alleged to be operating without permits—courts dismissed the motion to withdraw the lawsuit where the dispute was initially only about “whether a permit is required”; now, with aircraft engines and a larger number of units, evidence collection and regulatory oversight will become even more difficult. Multiple experts judge that efficiency and environmental impact will carry significantly greater weight within one to two years, making this solution more likely a short-term transitional measure.
How this affects your subscription bill
Though it appears to be energy-sector news, it actually sets the floor price for tokens. Electricity is one of the largest variable cost components in inference, and while building on-site generation may seem cost-saving, fuel, maintenance, unit depreciation, and future compliance costs will all ultimately be factored into API pricing and subscription tiers.
Several recent changes are, in fact, reflected on the same bill:Grok Bot reduced usage quotas,Claude’s $200 tier removed the 5-hour weekly limit but kept the overall weekly quota unchanged,OpenAI’s funding target cut to $200 billionAll three companies are doing the same thing: flattening demand via quotas and pricing until compute supply catches up.
The counterintuitive part is that on-site generation is not necessarily cheaper. Grid interconnection incurs a one-time waiting cost, while retrofitting turbine units entails ongoing fuel and maintenance expenses—plus carbon costs and risks of community litigation. Ultimately, energy efficiency determines long-term cost—which is why chip-side efficiency improvements are equally critical,NVIDIA’s Vera Rubin efficiency roadmap and space-based data center concepts (Google’s TPU satellite) all fundamentally address the same question: how to run more inference using less electricity.
Summary: A short-term fix, a long-term problem
Using retired passenger jet engines for power generation is the most pragmatic compromise when AI infrastructure outpaces electricity supply: online in 24 months, at the cost of efficiency, emissions, and regulatory risk. It will likely be replaced within two or three years by more standardized gas turbine units, energy storage, and grid expansion—but until then, whoever gets power connected first gets models running first. For ordinary users, this means inference prices won’t drop significantly in the near term—supply-side bottlenecks remain at the most fundamental level.
To see real cost differences across models, check Tool Comparison Engine the API price comparisons, or read GPT-6.1 Terra,Basic soldering, assembly, and firmware flashing skills. On the software side, users must follow documentation to configure Agent integration. Component creation requires no coding—components are generated from natural-language descriptions such detailed cost breakdowns.
Frequently Asked Questions (FAQ)
Why are data centers using aircraft engines for power generation?
Because it’s fast. Retrofitted aviation engine units can begin supplying power in as little as 24 months, whereas connecting 100 MW to the grid in Northern Virginia takes roughly seven years—and most new grid capacity won’t come online until after 2031. For AI companies racing to meet training and inference deadlines, time is more expensive than per-unit electricity cost.
Will this make electricity cheaper?
Not necessarily. Retrofit units operate at lower efficiency than standard gas-fired power plants; delivering the same output requires more units, higher fuel and maintenance costs, and added pressure to comply with emissions regulations. What they save is time—not generation cost.
Which companies are adopting this approach?
Publicly available information shows xAI installed eight units in Mississippi for Colossus 2; OpenAI deployed gas turbines for Phase I of its Stargate facility in Abilene, Texas; and Crusoe, the construction firm, ordered 13 retrofitted ProEnergy units. The equipment primarily comes from ProEnergy, which modifies GE aviation engines.
Will this affect AI subscription pricing?
Yes—but indirectly. Electricity is the largest variable component of inference cost; until supply bottlenecks ease, per-token pricing is unlikely to fall sharply, and vendors instead favor quota-based and tiered pricing to control costs—recent examples include Grok cutting quotas andClaude adjusting weekly limits, all following this logic.
What are the environmental concerns?
Retrofitted aviation engines are less efficient, emit more pollutants, and require more units than standard gas-fired power plants—making them more polluting overall. Some data centers have already faced lawsuits over gas turbine permitting issues, and increasing unit counts will concentrate regulatory scrutiny and community opposition.