AI Policy Wiki
Dashboard
⚲

Behind-the-Meter Power for Data Centers

medium confidence · updated 2026-09-17

On-site or co-located generation that supplies a data center directly, behind the utility revenue meter, bypassing or supplementing grid interconnection — the principal industry response to multi-year US connection queues for AI compute.

Behind-the-meter (BTM) power refers to electricity generation sited at or adjacent to a data center that is consumed directly by that facility, behind the utility revenue meter, either bypassing the public grid entirely or using it only as backup. It is distinguished from a conventional power purchase agreement or utility supply contract, in which power is sourced from the grid and metered accordingly (Source: datacenterhawk.com; orrick.com). The model has moved from a niche arrangement to a common development strategy as AI compute demand has outrun the pace at which utilities can grant grid connections.

Background: the interconnection constraint

The immediate driver is the length and uncertainty of grid interconnection queues. Enverus Intelligence Research reports that queue-to-commercial-operation timelines have grown roughly 60 percent since 2017 and now average over 2,100 days for projects with a first power year of 2025, meaning a new data center in a constrained market routinely takes five or more years to connect; by the same analysis, only about 10 percent of capacity sitting in interconnection queues is ultimately built (Source: enverus.com). Providers seeking utility power also face upfront load-study costs and per-megawatt capacity deposits with no guarantee of timing (Source: datacenterhawk.com).

The scale of demand behind the queue is large. The International Energy Agency projects $3.9 trillion of global data center investment between 2026 and 2030, against aging networks already strained by the broader energy transition (Source: aoshearman.com). In the European Union, the European Commission's Grids Package addresses connection delays by requiring grid operators to publish queue information and by accelerating permitting for grid upgrades (Source: aoshearman.com).

Adoption figures track the constraint. The Foley 2026 Data Center Survey found that 56 percent of developers were exploring co-located or on-site generation, the third most common power strategy after negotiating power purchase agreements and securing early grid interconnects (Source: datacenterhawk.com).

The pipeline is much larger than the operating base. The market-intelligence company Cleanview estimates roughly 97 gigawatts of behind-the-meter projects planned across hyperscalers including Meta, Microsoft and Oracle, against about two gigawatts online as of August 2026, most of it SpaceXAI's. Two of the largest disclosed projects are gas: a 7.65-gigawatt, 35-turbine plant in Pecos County, Texas developed by Pacifico Energy and known as GW Ranch, which Amazon confirmed on August 7, 2026 it is financing to supply a planned data-centre campus and which would be larger than any gas plant currently operating in the United States; and a 9.2-gigawatt facility planned in Ohio under a public-private partnership involving SoftBank. The Pecos County plant's permits would allow emissions of more than 30 million metric tonnes of greenhouse gases a year, more than any other single source in the country, though facilities rarely emit at their permitted ceiling (Source: finance.yahoo.com). Tesla and SpaceX confirmed on August 10, 2026 that the Terafab semiconductor plant in Grimes County, Texas will likewise run on newly built gas plants and "very large battery arrays" (Source: electrek.co).

Structures and fuels

In the basic arrangement the operator owns or contracts the generation asset directly and power moves from turbine or engine to facility load without touching the utility grid (Source: enverus.com). Natural gas is the predominant fuel, chosen for availability and speed of deployment; nuclear co-location and renewables paired with storage are the other principal routes (Source: enverus.com). The Talen Energy campus co-located with the Susquehanna nuclear plant in Pennsylvania is the most-cited nuclear example; the Federal Energy Regulatory Commission blocked its initial behind-the-meter structure in late 2024 and the arrangement was subsequently restructured. See Nuclear PPAs for AI and Federal Energy Regulatory Commission (FERC).

Hybrid structures sit between full BTM and standard interconnection. A "bring your own capacity" arrangement or flexible interconnection allows a facility to run on behind-the-meter resources during grid stress and draw from the utility otherwise; datacenterHawk reports research suggesting a 500 MW facility using this approach can reach full operation roughly three to five years earlier than through conventional interconnection (Source: datacenterhawk.com).

Reliability is the principal technical constraint. A&O Shearman notes that BTM improves speed-to-power and supply control but is difficult to reconcile with hyperscale uptime requirements, particularly where generation is intermittent: conventional short-duration battery energy storage may not satisfy the "five-nines" availability and 24/7 load-matching targets hyperscale operators set. The firm also observes that allowing data centers to run more flexible loads in exchange for faster grid connection is being explored but conflicts with operators' preference for maximum compute utilization (Source: aoshearman.com).

The turbine-component constraint

Behind-the-meter gas generation depends on a component supply chain narrower than the turbine market itself. Production of the vanes and blades inside large gas turbines — cast for service at 3,000 to 3,600 degrees Fahrenheit, some 800 degrees above the melting point of the alloys used — has been concentrated in four casting companies whose foundries are reported to be at capacity, one reason turbine makers including GE Vernova are sold out until 2030 (Source: theinformation.com). Elon Musk said in a February 2026 podcast that blades and vanes are "the limiting factor" in provisioning turbines, and that at current production rates "it's not clear to me that there's enough usable electricity to turn on all the AI chips that are being made."

SpaceX has taken steps toward casting the components itself. Job listings reported on August 29, 2026 advertised for a senior operations engineer "to serve as the extreme owner of the construction, buildout, and operational ramp-up of a new blades and vanes foundry" and for a programmer to mill parts in high-temperature superalloys at "SpaceX's new industrial gas turbine (IGT) blades and vanes manufacturing line in Bastrop," about 30 miles from Austin, Texas. Morgan Stanley analyst Adam Jonas said his own research indicated Musk is developing such a plant and that one facility would serve both data-center airfoils and Raptor turbopump castings, because they "draw on the same alloys, furnaces and trained workforce." Corey Trinetti, who publishes the newsletter Measured AI, found that SpaceX bought about 830 acres near its existing Starlink factory between March and June 2026 and that two new buildings appeared in July satellite imagery, but could locate no foundry permits or construction filings in Texas environmental or building records. SpaceX did not respond to requests for comment (Source: theinformation.com).

Musk has told SpaceX investors he could stand up 10 gigawatts or more of terrestrial data centers by the end of 2027; Jonas estimated 3 to 4 gigawatts of bridge power by then through purchase agreements. Musk bought the mobile-turbine supplier APR Energy personally from Fortress Investment Group, according to a Federal Trade Commission filing in May 2026; APR said in January it had about 1 gigawatt of turbines in its fleet. BloombergNEF has estimated that planned off-grid gas plants for data centers, if built, would raise US carbon dioxide emissions 20% (Source: theinformation.com). See AI Environmental Impact.

Regulatory treatment

Because a behind-the-meter facility contracts with a fuel supplier rather than a utility, its power costs fall outside the rate-setting jurisdiction of state utility regulators (Source: utilitydive.com). A separate and unsettled question is whether a data center with on-site generation must nonetheless be classified as a "network load," which would require it to contribute to transmission upgrade costs even if it rarely draws grid power (Source: orrick.com).

Federal policy has pushed in the direction of self-supply. Jeffrey Rissman and Eric Gimon of Energy Innovation identify the White House Ratepayer Protection Pledge of March 2026, alongside interconnection delay, as a reason data center projects increasingly plan their own generation (Source: utilitydive.com). The pledge's terms are not reproduced in that source. A National Republican Senatorial Committee memorandum of August 18, 2026 renders them as five commitments — that all new energy supply for data centers is new power with "no burden on existing customers"; that data centers "pay their own way on generation and infrastructure"; that there be "no higher costs on families or ratepayers"; that the investment strengthen the grid through upgrades and backup generation; and that communities gain jobs, workforce, and long-term growth (NRSC — Ohio Data Center Risk (memorandum, August 18, 2026)). That is a party committee's summary offered in a campaign document, not the pledge's own text, and the pledge is voluntary in any case; the first two commitments describe the outcome that the network-load and cost-allocation questions above are contesting rather than a mechanism for reaching it. The House-passed Ratepayer Protection Act, which the Energy and Commerce Committee described on introduction as codifying "the principles underpinning the Ratepayer Protection Pledge," does not mention the pledge in its text (Ratepayer Protection Act, H.R. 9340 — text as reported in the House (September 10, 2026); Source: energycommerce.house.gov). Emerald AI chief executive Varun Sivaram has argued that behind-the-meter campuses raise costs for other users by competing for the same turbines and transformers and deprive utilities of anchor customers (Source: fortune.com). Air-permitting has become a contested venue: xAI was sued over its Southaven, Mississippi site for operating gas turbines without a permit in alleged violation of the Clean Air Act (Source: utilitydive.com).

Debate over ratepayer and emissions effects

Whether behind-the-meter generation insulates other customers from data center demand is disputed. Rissman and Gimon argue in an opinion piece for Utility Dive that it does the opposite: because natural gas is a market-traded commodity, on-site gas plants compete with other gas buyers and raise gas prices, and because gas supplies 43 percent of US electricity and sets the marginal price in most hours, higher gas prices feed back into electricity bills. They cite a Bloomberg New Energy Finance analysis finding 100 GW of on-site gas capacity planned to serve US data centers, equivalent to 18 percent of existing US natural gas generating capacity, and note that grid-connected data centers can at least be assigned cost-covering rates by regulators while behind-the-meter projects cannot (Source: utilitydive.com).

The same authors argue that efficiency compounds the effect: the most efficient combined-cycle turbines are back-ordered five to seven years, so developers rely on less efficient equipment that burns more gas per unit of output and emits more. They propose that policymakers require data centers to bring their own clean electricity rather than simply their own electricity, using siting and zoning rules, air-quality regulation, permit-by-rule for renewables, pre-authorized generation zones such as Texas's Competitive Renewable Energy Zones and Nevada's Solar Energy Zones, and harmonized local permitting on the model of Michigan's HB 5120 and Illinois's HB 4412. As examples that this is achievable they cite Google's planned Minnesota data center, to be supplied by 1.6 GW of wind and solar with 300 MW of battery storage, and Amazon's purchase of 1.2 GW of solar capacity and an equal amount of battery storage for its Oregon facilities (Source: utilitydive.com). This is an advocacy position from a clean-energy policy organization and is presented as such.

The two examples the authors open with illustrate the scale at issue: a facility in Richland Parish, Louisiana drawing 2.2 GW, roughly twice the peak summer demand of New Orleans, and a project near Cheyenne, Wyoming whose first phase requires 1.8 GW and which is designed to scale to 10 GW, comparable to New York City's peak draw (Source: utilitydive.com).

Relation to siting and local opposition

Behind-the-meter generation shifts, rather than removes, the local politics of data center construction, because the generating plant is itself sited in the host community. Rissman and Gimon report that most US voters now oppose data center development on grounds including home energy costs, environmental effects, and quality-of-life impacts such as turbine noise and soot (Source: utilitydive.com). The local-permitting dimension is treated at Data Center Siting / AI Power Politics; aggregate power demand at AI Data Centers and Energy and Electric Power Sector; emissions at AI Environmental Impact.

Relationships

Sources

  • Enverus / Francesca Costello, "Why data centers are looking to natural gas for behind-the-meter power" (May 28, 2026): enverus.com
  • A&O Shearman, "Powering data centers: the rise and challenges of the 'behind the meter' model" (July 16, 2026): aoshearman.com
  • Jeffrey Rissman and Eric Gimon (Energy Innovation), "Behind-the-meter data center gas plants will raise US energy bills," Utility Dive opinion (June 8, 2026): utilitydive.com
  • datacenterHawk, "Behind-the-Meter Power Solutions: The Data Center Industry's New Reality": datacenterhawk.com
  • Orrick, "Powering Data Centers | Megawatts to Megabytes" (November 2025): orrick.com