Microsoft and Chevron sign a 20-year, 2-gigawatt power deal for Pecos

Microsoft is building a 2-gigawatt AI datacenter in Pecos, Texas, and it is paying Chevron directly for the power. On June 22, 2026, Microsoft called the new campus "one of the largest single capacity additions in our history." The same morning, Chevron announced a 20-year agreement to deliver 2.67 gigawatts of dedicated, behind-the-meter generation to the same site through a project called Kilby.

The size of the commitment is the headline. Two gigawatts is roughly the output of two large nuclear reactors, or the average demand of more than 1.5 million U.S. homes. Microsoft's president of cloud operations, Noelle Walsh, said the project is "grounded in a simple principle: we build where our customers need us, and we build for the long term." Microsoft said the campus will support more than 6,000 construction jobs at peak build-out and create hundreds of permanent operational roles in Reeves County. Chevron's release adds that Kilby alone is expected to generate more than $10 billion in state and local tax revenue and support almost 2,000 jobs in the region.

The financial structure is the part that will get the most attention. Under the Chevron deal, Energy Forge One, a wholly owned Chevron subsidiary, will build and operate a co-located natural gas power plant next to the datacenter. The generation is dedicated to Microsoft under a 20-year power purchase agreement, which insulates the campus from the kind of grid interconnection delays that have stalled other large AI buildouts. Chevron and Engine No. 1, the activist fund that helped push Exxon into a new strategic review, have been collaborating on the development. Kilby is expected to deliver about 2.67 GW of capacity, with a majority coming from GE Vernova turbines and additional capacity from Solar Turbines, a Caterpillar subsidiary. First power is targeted for 2028, and Chevron's final investment decision is expected by the end of 2026.

The Permian Basin is the obvious reason the deal is happening where it is happening. The basin is the largest oil-producing field in the United States and the largest single source of associated natural gas, much of which has historically been flared or sold at deep discounts because pipeline capacity out of the region has not kept up. Tying a new power plant to that stranded gas supply is one of the few ways a hyperscaler can lock in multi-gigawatt capacity on a timeline shorter than the five-to-seven-year window for a new combined-cycle gas plant plus interconnect on the public grid. Microsoft framed the choice plainly in its announcement: "By pairing new datacenter infrastructure with dedicated energy supply located onsite, we can bring capacity online at the pace our customers require while maintaining operational reliability."

Why Microsoft is paying for its own power

The power-purchase structure is the strategic pivot in this announcement. Until this year, the major hyperscalers have generally leaned on regulated utilities and independent power producers to supply their new AI campuses. Microsoft itself has been signing large renewable energy contracts in Texas, including 4.7 GW of contracted renewable electricity in the state, and has been building out nuclear agreements with Constellation and others. Behind-the-meter fossil generation is a different bet. It trades the carbon profile of natural gas for a much faster path to dispatchable, large-scale power, and it puts the financing risk on the buyer rather than on the regional grid operator or the utility.

Chevron's framing makes the same point from the other side of the table. "AI is reshaping the global economy, and abundant, affordable, reliable energy is essential to fueling that transformation," said Jeff Gustavson, Chevron's president of new energies. "Chevron is uniquely positioned to deliver power to customers with certainty, speed and at a competitive cost, leveraging Permian natural gas and our proven execution capabilities." The economic logic for Chevron is also significant. The press release said Kilby is targeting "mid-teen returns" and is expected to generate diversified cash flow that is independent of oil and gas price cycles. For an integrated oil major whose stock has traded in a tight range for most of the last two years, a long-dated power contract with a hyperscaler is the kind of cash flow profile investors tend to reward.

This is also the first time a U.S. oil and gas major has tied a single power plant of this size to a single datacenter customer. Smaller co-located gas plants have been built for individual campuses before, but Kilby, at 2.67 GW, is the largest dedicated, behind-the-meter gas generation deal between a U.S. energy company and a hyperscaler that has been publicly disclosed. The closest comparable deal is the broader 20-year gas supply agreement that a different hyperscaler signed with a different oil and gas producer for fuel, not for power. The Kilby structure makes Microsoft a power off-taker, not a fuel buyer, which is a different and more binding commercial relationship.

Pecos sits in the middle of a stretch of West Texas that has become one of the most contested pieces of energy real estate in the country. ERCOT, the operator of the Texas grid, has been working through an unprecedented queue of large industrial loads seeking interconnection, and the pace has been the single biggest constraint on AI buildouts in the state. A 2 GW behind-the-meter load does not appear on that queue, because the campus is not interconnecting to the public grid at the start. The Microsoft press release explicitly says the plant will serve the campus "directly and independently of the public grid, so this demand does not take from the current grid."

That detail matters for two reasons. The first is local. Reeves County and the surrounding region have seen a sharp jump in large industrial load applications over the last 12 months, and adding 2 GW of new demand to the public grid would have required new transmission, new generation, and a multi-year approval process. The co-located plant sidesteps the entire transmission build. The second is at the state level. ERCOT is in the middle of a rule-making cycle on how to allocate grid costs to large new loads, and one of the open questions is whether hyperscalers should pay a premium for new transmission built to serve them. By going behind the meter, Microsoft effectively removes itself from that debate for this campus. It is the cleanest demonstration to date that hyperscalers can choose not to participate in the contested grid-interconnection process at all if they are willing to pay for their own generation.

Water, emissions, and the broader buildout

The environmental design choices in the Pecos announcement are notable because they try to head off two criticisms that have followed large AI datacenter builds in the desert: water use and local air emissions. The Microsoft post commits to closed-loop cooling, which only needs an initial water charge at the start of operations and does not consume additional water during steady-state operation. Microsoft says the total lifecycle water use of the campus is "only a fraction of that consumed annually by a typical fast-food restaurant." On emissions, the campus will be served by a plant equipped with Selective Catalytic Reduction systems to control nitrogen oxide output, and Microsoft says it will "continue to drive additional improvements in environmental performance in line with our corporate commitments."

The water design choice in particular is a shift from the cooling towers that are common on older hyperscaler campuses in Arizona, Nevada, and west Texas, where evaporation losses are the single largest component of municipal water demand. Closed-loop systems, by contrast, are sealed and use air-cooled heat exchangers or similar designs that recycle the same water for years. For Reeves County, where groundwater levels have been a politically sensitive topic for more than a decade, the choice is also a way for Microsoft to avoid the kind of pushback that has followed other large water-using builds in west Texas. Chevron's release adds that the Kilby power plant will use "non-potable, brackish groundwater" rather than freshwater for its own operations, and that the partners are exploring the reuse of produced water from oil and gas operations. None of these choices eliminate the project's environmental footprint, but they do align the build with the political constraints that have shaped data center siting decisions in the region.

Pecos is the largest single piece of a much larger buildout that Microsoft has been signaling since the start of 2026. In April, Microsoft announced a $10 billion investment in Japan for AI infrastructure through 2029, and a separate A$25 billion commitment to AI capacity in Australia, both of which will run into the same grid constraints that have defined the U.S. buildout. In May, the company signed 4.7 GW of renewable contracts in Texas, which is the largest single-state renewable commitment Microsoft has made. The Pecos campus adds roughly 2 GW of dispatchable gas-powered capacity on top of the renewables, which suggests Microsoft is increasingly thinking of its AI infrastructure as a portfolio of generation technologies rather than a single procurement.

That portfolio framing is the most important signal in the announcement. The peer group is moving in the same direction. Anthropic and Amazon announced a 5 GW AI compute agreement in April, with much of the new capacity tied to a multi-year power structure. Oracle signed a large new power deal the same month for an unspecified data center buildout, and KKR launched a $10 billion AI datacenter joint venture with Nvidia and Vistra in late May. Each of these deals uses a different financing structure and a different generation mix, but they share a common feature: the buyer is taking direct responsibility for the power that runs the campus, either through a long-dated power purchase agreement, a behind-the-meter plant, or a hybrid of both. For a deeper look at how those financing structures differ, see our AI infrastructure resource page.

The second-order question is whether Kilby is replicable. The deal works because Chevron had stranded gas, Engine No. 1 had a structural mandate to push Chevron into new businesses, Microsoft had demand that could not wait for grid interconnection, and Reeves County had the right combination of land, water rights, and political support. None of those four conditions is easy to copy in another basin, and the largest gas-producing regions in the U.S. outside the Permian either lack the pipeline network, the political alignment, or the hyperscaler demand to support a similar deal at scale. That is why Microsoft's other large AI buildouts are still anchored on utility-scale renewables and nuclear, not on more behind-the-meter gas.

Three signals that will decide what comes next

Three indicators in the next 90 days will tell us whether the Pecos build is the start of a pattern or a one-off. The first is whether Chevron hits its final investment decision on Kilby by the end of 2026 as planned. A miss on that timeline would imply either permitting friction in Reeves County, turbine supply problems at GE Vernova, or a softer Microsoft demand picture than the announcement suggests. The second is whether Microsoft signs a second behind-the-meter gas deal with a different partner in another basin. A second deal would mean Microsoft has decided the model is the new template, not a one-off. A failure to announce a follow-up would mean the model is being held in reserve for sites where the grid is genuinely unworkable. The third is whether the Permian gas pricing that underpins the economics of Kilby stays in the range Chevron is assuming. The press release does not disclose the gas price, but any sustained move above the mid-single-digit dollar range per thousand cubic feet would compress the mid-teen return target and could slow the next phase of expansion.

For background on how the power and datacenter story has shaped Microsoft's broader enterprise AI strategy, see our recent coverage of the KKR-Nvidia-Vistra AI datacenter joint venture. The original Microsoft announcement, including Walsh's full statement and the underlying community commitments, is on Microsoft's official blog.

The broader lesson is that the AI infrastructure cycle is no longer just a story about GPUs and model launches. It is increasingly a story about how hyperscalers finance, site, and power the next 10 to 20 gigawatts of training and inference capacity. The Microsoft-Chevron deal at Pecos is the cleanest example so far of what that financing structure looks like when the hyperscaler and the energy company are willing to sign a 20-year contract and share the same site. The next question is whether the same template can be built quickly enough to keep up with the demand curve, or whether the buildout will end up gated by the same turbine, permitting, and gas-supply constraints that have shaped every other large energy project of the last decade.