Under a project dubbed Kilby, on site gas turbines will deliver dispatchable power next to Microsoft's data center, sidestepping the transmission queues that now stall most AI buildouts.
Chevron is not selling oil to Microsoft. It is selling the thing the Texas grid cannot deliver fast enough to house AI compute: dispatchable power without a queue. That is the structural change hiding inside the 20-year power purchase agreement Chevron and Microsoft announced this month for a project the companies have dubbed Kilby in West Texas.
The headline numbers frame the deal as another oil-and-Big-Tech handshake. They are bigger than that. The agreement covers about 2.67 gigawatts of on-site gas-fired generation, phased in modularly, with first power targeted for 2028. The site is "co-located," meaning the turbines sit next to Microsoft's data center rather than feeding into the public transmission grid; the power goes directly to the servers behind the meter. Chevron, acting through a special-purpose vehicle called Energy Forge One, has partnered with Engine No. 1 to develop the project. Both companies' press materials frame it as one of the largest co-located gas-and-data-center sites in the United States.
What the framing leaves out is the reason Microsoft needed a 20-year offtake in the first place. "Offtake" is the energy-industry term for a buyer's contract to purchase a project's output; in this case, Microsoft is signing up to buy Kilby's power for two decades. The Texas grid, run by ERCOT, has become one of the tightest chokepoints in the U.S. AI buildout. Transmission lines take years to permit, and new interconnections routinely sit in queues measured in gigawatts. A hyperscaler that wants multi-hundred-megawatt capacity to train and run AI models does not have that time. Co-located gas generation, even at a higher marginal emissions rate, lets a buyer sidestep the queue entirely and put a megawatt behind a server the same year the turbine is commissioned. Chevron is monetizing the precise commodity Microsoft cannot otherwise secure: a long-dated, queue-bypass contract, with the Permian gas supply sitting in the same basin as the load.
The trade-off is real and worth naming. When a hyperscaler builds behind the meter, the costs and externalities of generation do not vanish. They sit with the offtaker, the host community, and ultimately the climate math of the AI buildout. Chevron has cited roughly $10 billion in projected state and local tax revenue and around 2,000 jobs tied to the project; those numbers come from Chevron's own disclosures and SEC filings, and no independent Texas economic-impact study has been cited. West Texas counties taking on the permitting, air-quality, and water burdens of a multi-gigawatt gas plant are not the same constituency as Microsoft shareholders, and the deal does not directly compensate them for the queue-bypass.
A useful counterframe, laid out by independent analyst Jeremy Lanin, reads the deal the other way: Microsoft owns and operates the compute asset and the customer relationship. Chevron is selling power, interconnection rights, and long-dated gas monetization. The two companies are splitting the AI-power stack between them, with the data center as Microsoft's side of the ledger and the 20-year offtake as Chevron's. If that read holds, the "oil company wins AI deal" framing is a category error. The deal is closer to a long-term utility-style contract than an oil sale.
Whether more hyperscalers follow depends on a handful of concrete signals over the next 18 to 24 months. ERCOT's interconnection queue and any move by the Public Utility Commission of Texas to curb behind-the-meter arrangements would change the math. EPA action on permitting for new gas turbines and any climate rule targeting data-center emissions would compress the cost gap between co-located gas and grid-supplied renewables-plus-storage. The first Kilby phase coming online in 2028 will be the test of whether phased modular gas actually delivers on the queue-bypass promise. By the end of this decade, the more interesting question is not whether AI is energy-hungry, but whether the U.S. AI buildout can still credibly run on a public grid that was not built for it.