A 20-year natural gas contract for 2.67 gigawatts rewrites the energy supply model for AI data centers, and tests how far Microsoft will go to keep its capacity promises.
The grid interconnection backlog that has stalled hundreds of announced data centers now has a workaround: co-locate the power plant on the same acreage as the server farm and never touch the public grid. Project Kilby is the most concrete execution of that model yet. Whether it closes, and on what timeline, is a different question than whether the model is correct.
The interconnection queue has turned the Permian Basin into an energy arbitrage opportunity for Microsoft and Chevron. Natural gas volumes in the Permian routinely exceed what regional pipelines can move, which forces operators to flare the excess and depresses local gas prices. Project Kilby converts that stranded supply problem into a competitive cost structure for a 2.67-gigawatt data center campus in Reeves County, Texas, near the city of Pecos.
The Chevron announcement, dated June 22, 2026, describes a 20-year power purchase agreement with Microsoft under which Energy Forge One, a Chevron subsidiary, will develop the co-located power facility alongside investment firm Engine No. 1 and turbine supplier GE Vernova. Additional generation capacity will come from Solar Turbines, a Caterpillar subsidiary. The plant will not connect to the Texas power grid. It will produce electricity exclusively for the Microsoft data center campus on the same site.
Three weeks ago, this analysis argued that the binding constraint on AI infrastructure is not money — it is interconnection queues, permitting, and turbine lead times. Project Kilby is one concrete answer: bypass the queue by owning the generation asset outright.
The model that replaces the grid connection
Behind-the-meter co-location has been building as a concept for two years, but most of the examples involved solar arrays and battery storage. Project Kilby is different in scale and fuel type. At 2.67 gigawatts, it sits in the range of a large coal or nuclear plant. The generation comes from gas turbines on the same site, fed by Chevron's existing Permian production, with no grid connection and no utility contract. Ratepayers on the Texas grid see none of this load.
Jeff Gustavson, Chevron's president of New Energies, framed the economics in the company's announcement: the ability to move power from the project's gas fields to the data center through a company-controlled supply chain, at competitive cost, is what makes the numbers work. The co-location design also provides the political cover that conventional grid-tied projects have struggled to maintain. Ratepayer anger over data center load additions to shared grids has driven moratorium legislation in over a dozen states. A campus that generates its own power sidesteps that objection entirely.
Engine No. 1 holds an option to acquire half the project and contribute matching capital. Total estimated cost, per Bloomberg reporting at the time of initial negotiations in April, runs to roughly $7 billion, though Chevron has not disclosed a final figure. That figure is vendor-supplied and unaudited. The company's own press release describes the project as targeting mid-teen returns and expects a final investment decision by the end of 2026.
Microsoft chose capacity over carbon profile
Microsoft has spent several years building a reputation for clean energy procurement. The Three Mile Island nuclear deal, framed as a zero-carbon power source for Pennsylvania data centers, was a credible expression of that. Project Kilby is something else. Natural gas combustion at 2.67 gigawatts over 20 years produces substantial carbon dioxide and criteria air pollutants regardless of how the project is framed. The Environmental Integrity Project, citing its own modeling, estimated more than 13 million tons of carbon dioxide as a potential output of the full build-out.
Noelle Walsh, Microsoft's president of cloud operations and innovation, described the project in the company's announcement as a response to AI demand that requires energy infrastructure that can scale quickly and reliably. That framing prioritizes reliability over carbon profile, which is a real shift from the positions Microsoft has taken publicly on sustainability. The company's capital expenditure budget for 2026 runs to $190 billion, up 61 percent compared to the prior year. At that scale of infrastructure commitment, reliability and speed of delivery carry more weight than fuel type.
The company that once anchored its infrastructure story on renewable procurement has now signed one of the largest dedicated natural gas supply contracts in the country. That is not a contradiction in terms. It is a prioritization made visible.
Enterprise buyers paying attention to cloud provider sustainability commitments should note what this deal reveals. The carbon accounting for workloads running on Microsoft Azure infrastructure connected to the Pecos campus will depend on how Microsoft structures its scope-2 emissions reporting for behind-the-meter gas generation. That accounting methodology has not been disclosed.
Chevron needed a business that does not track oil prices
Chevron's New Energies division has been searching for revenue streams that do not rise and fall with crude. A 20-year, fixed-counterparty power purchase agreement with one of the world's largest technology companies is close to the opposite of oil price exposure. It produces predictable cash flow independent of commodity cycles, which is what the investor presentation language around "resilient shareholder returns" is pointing toward.
Gustavson said publicly that others in Chevron's peer group are talking about similar moves and that Chevron is the one doing it now. That is a direct bid to establish a replicable model before other oil majors move. If Project Kilby closes on schedule and delivers first power in 2028, Chevron will have a proof of concept for co-located power supply that other majors with stranded Permian gas can reference.
The Final Investment Decision is still pending. Construction has not started. The development timeline runs five to seven years to full build-out. The deal that ran in the newspaper this week is an agreement to build something, not a confirmation that the thing exists.
The construction gap has not closed
Project Kilby sits inside the same infrastructure reality that applies to every other large AI capacity announcement: first power arrives in 2028, the site ramps through the early 2030s, and the Final Investment Decision happens later this year. A CIO looking at Microsoft Azure's capacity roadmap for the next 18 months is looking at capacity that does not yet exist in Reeves County.
Co-located natural gas generation at this scale still requires permitting, turbine procurement, and construction sequencing. GE Vernova large-frame gas turbines, which will supply the majority of Project Kilby's output, carry lead times that now run several years. The supply commitment is locked. The physical hardware still has to arrive and be installed.
The pattern here matches what has played out across the broader AI infrastructure buildout. Capital is visible and announces well. Steel and concrete on a specific site in a specific county, with a confirmed interconnection or a completed gas turbine installation, is the measure that matters for enterprise buyers.
Project Kilby is the most advanced execution of the co-located power model in the AI infrastructure space, solving the interconnection queue and ratepayer backlash problems at once. The timeline is the exposure. A provider whose capacity depends on a power plant with a pending Final Investment Decision and a 2028 delivery target cannot serve workloads that need capacity in 2026 or 2027. Understanding which specific facilities back the AI services you are buying is no longer optional due diligence.
Your AI workloads depend on your platform provider, whose capacity depends on power supply agreements, whose value depends on whether operational facilities exist when your business needs them. Project Kilby is a 2028 story at earliest. If your processing requirements land before that, the question to answer is what happens to your workloads if your provider cannot deliver because the energy supply was not ready. Ask for facility-level capacity status, confirmed power source, and operational date — not platform-level commitments. The gap between those two answers is where the risk lives.
