17JUN

Chevron builds Microsoft a gas plant

4 min read

10:13UTC

Chevron signed a 20-year deal with Microsoft on 22 June to build Project Kilby, a 2.67 GW gas plant in West Texas that feeds one data centre and skips the grid entirely. At roughly $7bn it is the largest dedicated-power deal struck for a data centre this year.

← Back to Virginia taxes the backup, not the draw Jump to analysis ↓

IndustryDeveloping

Key takeaway

The cheapest route to AI power in 2026 is to build a private plant, even at $7bn.

Chevron signed a 20-year power purchase agreement (PPA), a long-term contract to buy electricity at a fixed price, with Microsoft on 22 June to build a 2.67 GW (gigawatt) gas plant beside a West Texas data centre. The roughly $7bn project, branded Project Kilby and run through Chevron's Energy Forge One unit, feeds the data centre directly, behind the meter, with no grid connection, no regional operator, and no curtailment order in the way ¹. At 2.67 GW it is the largest dedicated-generation deal struck for a data centre this year, ahead of xAI's 1.2 GW Colossus turbines and Meta's 365 MW Wyoming solar build ².

Operators are pouring capital into their own generation because every route to grid power slowed at once, and the cheaper escapes are closing as fast as they open. Washington has twice switched off the behind-the-meter route during heat events, when the Department of Energy ordered data centres running their own generation to curtail . Kilby answers that risk by owning the plant outright, but it inherits a different queue: its GE Vernova turbines come from a backlog already booked into 2029 , so first power is not promised before 2028 and the final investment decision (FID) waits until the end of this year.

A 20-year contract on a single-customer plant resembles a project-financed power station more than a corporate clean-energy deal, with the data centre as its sole creditworthy anchor. That structure is why a supermajor will commit roughly $7bn against one offtaker. Energy-intensive industry has done this before: aluminium smelters built their own captive generation through the 20th century whenever grid supply was too slow or too thin. AI compute is repeating the pattern rather than inventing one.

Deep Analysis

In plain English

Microsoft, Google, and Amazon each operate large computing facilities called data centres, which run the servers behind cloud services, email, and online shopping. Those facilities consume vast amounts of electricity, and the normal path is to connect to the public electricity grid, the same network that powers homes and offices. But in the US, getting a new grid connection approved now takes several years, and the federal government has twice this year ordered data centres to switch off their own backup power generators during heatwaves to protect the public grid. So Microsoft and Chevron, the oil company, have agreed to build a private gas-fired power station right next to Microsoft's data centre in West Texas. The electricity goes directly from the gas plant to the data centre building, never touching the public grid. This arrangement is called behind-the-meter generation, because the power plant sits on the customer's side of the electricity meter. At 2.67 GW (gigawatts), Project Kilby would generate roughly as much electricity as a medium-sized city uses, all going to one facility. It costs about $7bn to build and will take until 2028 to produce its first power.

Deep Analysis

Root Causes

Three structural failures make the $7bn co-located plant rational rather than extraordinary.

First, FERC's interconnection queue was built for utility planning cycles of 5 to 10 years, not 18-month hyperscaler deployment calendars. Docket RM26-4-000 has been open since October 2025 and as of 18 June still produced no binding standard, meaning Microsoft's West Texas capacity need cannot be met through the regulated queue in any plausible timeframe.

Second, DOE's Section 202(c) emergency curtailment authority, exercised twice in 2026, has paradoxically accelerated BTM investment. A grid-served data centre now carries curtailment exposure the operator cannot hedge; an operator-owned generator removes that exposure entirely.

Third, GE Vernova's $163bn backlog means turbine access is the binding constraint on BTM builds. Chevron's oil-and-gas procurement relationships give Microsoft access to GE Vernova delivery slots that a hyperscaler placing an open-market order in 2026 would wait three or more years to receive. The Permian Basin location also provides direct pipeline access to Chevron's own gas production, removing fuel-supply risk from the equation.

What could happen next?

Precedent
The 2.67 GW deal sets the scale reference for oil-major BTM power-services contracts; BP, Shell, and ExxonMobil now face an explicit build-or-miss-market decision on dedicated hyperscaler generation.
Short term · Assessed
Risk
If courts rule in the ID:3834 rehearing that DOE curtailment authority covers operator-owned BTM generation as well as backup generators, Project Kilby's core rationale collapses before first power in 2028.
Medium term · Suggested
Consequence
Each GW of BTM generation built behind the meter removes load from the public grid, spreading fixed grid costs across fewer customers and raising residential electricity rates in the host region.
Medium term · Assessed

Source Landscape

This story draws on mixed-leaning sources from United States

United States

Primary parallel: The Public Utility Regulatory Policies Act of 1978 (PURPA) opened the same structural escape route for 1980s US industrial users.

Steel mills, chemical plants, and paper mills built co-located cogeneration plants under PURPA's avoided-cost framework, bypassing utility monopoly pricing in exactly the logic Microsoft now applies: own the generator, skip the grid queue, achieve cost certainty. FERC Order 888 in 1996 then forced those self-generators into open-access interconnection anyway, partially reversing the bypass.

Counter-parallel: US aluminum smelters in the Pacific Northwest built captive hydropower contracts with the Bonneville Power Administration in the 1990s on 20-year terms, reasoning they had locked in cheap baseload.

Bonneville's capacity constraints in the 2000s made those contracts a liability rather than an asset, because the physical infrastructure could not deliver what the contract promised. Project Kilby is different in that Chevron owns the generator itself, removing the counterparty's physical delivery risk.

At $7bn capex, Project Kilby represents approximately 10% of Chevron's 2025 full-year capital budget, a material commitment for what is formally a power-services business Chevron does not currently operate at scale. Chevron's entry into dedicated hyperscaler power services opens a new revenue vertical for integrated oil majors; BP, Shell, and ExxonMobil face the same question of whether to follow.

For Microsoft, the economics turn on avoided grid cost rather than absolute power cost. ERCOT's wholesale market has delivered spot prices above $9,000 per MWh during stress events (February 2021) and carries curtailment risk under federal emergency orders.

A 20-year fixed PPA at Permian Basin gas economics provides cost certainty that ERCOT spot pricing cannot. The 20-year term at 2.67 GW implies roughly 400 TWh of total generation at current load factors, making this a $20-30bn all-in energy commitment when fuel costs are included.

Consensus view: SemiAnalysis analyst Dylan Patel has argued publicly that behind-the-meter (BTM) gas generation is the only rational response to an interconnection queue that now takes three or more years to clear. At 2.67 GW, Project Kilby is a capital commitment that reflects that calculation at its most explicit: Microsoft and Chevron are building the utility, not waiting for one.

Counter-view: The Institute for Energy Economics and Financial Analysis (IEEFA) has consistently argued that long-term gas PPAs signed under grid-queue pressure lock in fossil-fuel infrastructure for 20 to 30 years, precisely when grid decarbonisation targets require the opposite trajectory. IEEFA's position is that queue-clearing reforms, not captive gas, are the correct response.

Key tension: Whether the DOE's two Section 202(c) curtailment orders in 2026 extend legally to operator-owned BTM generation or only to backup generators connected to the grid, a distinction that determines whether Project Kilby's escape from curtailment exposure is durable or temporary.

First Reported In

Update #8 · Data centres build their own power plants

Chevron· 28 Jun 2026

Read original →

Causes and effects

Caused by

DOE lets PJM switch off data centres

Operators build BTM generation like Project Kilby specifically to bypass the grid queue that DOE has twice ordered curtailed under 202(c)

Occurred 18 May 2026

Read story →

GE Vernova prices the transformer bottleneck

Project Kilby draws on GE Vernova turbines from a backlog already booked into 2029, the same constraint GE Vernova disclosed in April 2026

Occurred 22 Apr 2026

Read story →

This Event

Chevron builds Microsoft a gas plant

When an oil major commits $7bn to a power station that serves one data centre, the grid has stopped being a usable route for compute that needs power before 2028.

Different Perspectives

Government of Odisha

Odisha signed a $1.5bn agreement with HCLTech and Sarvam AI on 24 July for a sovereign AI data centre in Bhubaneswar, targeting 2028 operations. While regulators elsewhere price developers out of the queue, Odisha is bidding state financial assistance to bring capacity in, betting the AI infrastructure boom is worth courting rather than rationing.

Switch Datacenters and Uithoorn

The developer, the municipality, Liander and dozens of greenhouse growers signed a waste-heat agreement letting a 60 MVA facility proceed under a provincial rule that bans sites above 5 MVA. The public-interest exception argues the site relieves rather than adds to grid congestion, working around a standing ban from inside rather than waiting for it to change.

Iskandar Puteri residents

Residents protested in February over falling water pressure they attribute to ZDATA's 300 MW construction; ZDATA disputes causation and has not yet resolved the dispute with its wastewater-only pivot. The complaint sits unadjudicated while the operator advertises compliance to state authorities the residents cannot address directly.

Microsoft

DC Byte reports Microsoft looking to the Nordics for new capacity while Dublin's regional share falls to 15 per cent, though this is a market-intelligence inference rather than a company statement. Microsoft keeps its existing Dublin cloud region, which is costly to relocate, while directing new investment where power is cheaper and the queue is shorter.

ISO New England

ISO-NE proposed excluding new large loads from its Installed Capacity Requirement entirely, forcing data centres to bring their own generation or accept curtailment, even though New England's demand has been flat or falling for a decade. The reform is pre-emptive, written before the region has felt the strain PJM and MISO already report.

Ofgem

Ofgem opened its Data Centre Commitment Fee consultation on 29 July, arguing that free options on grid capacity let developers bank places on ten sites and abandon nine, crowding out projects that would build. It expects the deposit to filter the queue before the rule even takes final form.