EPA’s Meter Draws the Law Amid a 97 GW AI Gas Rush

In 2025, Global Energy Monitor counted more than 97 GW of proposed US gas-fired power explicitly earmarked for data centers, up from 4 GW in 2024. On July 16, 2026, the EPA said a power plant serving only a data center may fall outside federal pollution programs that could govern the same plant if it served the public grid. The fuel and emissions can stay the same while the meter changes the rulebook.

A seven-year queue makes expensive power rational

The United States can have enough electricity for AI data centers and still leave a campus waiting seven years for power. A megawatt somewhere in the national supply balance is not a connection at a particular parcel, behind a particular substation, on a developer’s construction calendar. Developers are filling that gap with turbines, diesel tanks and air permits.

The original data-center design assumed that a utility delivered electricity and the campus consumed it. That division let developers concentrate on land, fiber, cooling and servers while grid operators handled generation and transmission. It worked while electrical service arrived within a tolerable construction schedule.

AI campuses broke that bargain. US data-center developers facing grid-access waits of up to seven years turned to aeroderivative turbines and diesel generators, even though on-site power can cost more and produce more emissions than ordinary grid supply. They were choosing the electron that could arrive before an expensive campus spent years idle.

Hyperscalers announced 46 GW of AI data-center capacity, enough at full utilization to consume as much energy as roughly 44.2 million US households. A national generation total cannot tell any one of those campuses when a substation, transmission path and firm-service agreement will be ready. The bottleneck has a street address: the site where planned computing equipment and deliverable power fail to meet on the same date.

Developers responded by moving generation inside the project boundary. A campus with its own power can proceed without waiting for every upstream grid project, but it also inherits the fuel contracts, emissions controls, operating permits and legal exposure that utilities once absorbed.

By replacing uncertain grid connections with dedicated generation, hyperscalers pulled environmental regulation into the AI infrastructure power stack. Permits that decide which generators can run, under what rules and after what public process now help allocate compute capacity.

Private resilience can become public instability

On-site generation turns a data center into an active grid participant. Its operator can remove a very large load from the grid in a single switching event, then restore that load according to the campus’s own operating needs. Grid operators must plan around both states.

Facilities in Northern Virginia have made the problem concrete. Data Center Alley contains more than 200 facilities and consumes roughly as much electricity as Boston. When campuses make “unannounced disconnections” by switching to local generators, those abrupt changes can contribute to cascading outages. The generator that protects one campus from a blackout can raise reliability risks beyond its fence.

PJM Interconnection responded by planning to require large data centers either to provide on-site generation or to curtail electricity use when necessary to prevent large-scale outages. Federal regulators also approved orders intended to process data-center power requests within 90 days while imposing new requirements on AI hyperscalers. Their rules treat campuses as controllable elements of the power system.

Developers installed private generators because the shared grid could not guarantee their schedules. PJM then had to govern those generators because a campus’s self-protection could threaten the grid it was trying to escape.

The meter determines which rulebook follows the turbine

The EPA’s July 16 letter identified the Clean Air Act’s Acid Rain Program among the federal programs that may not cover plants serving only data centers. The agency attached regulatory significance to the generator’s service relationship: whom the plant supplies can matter alongside what it burns and emits.

The EPA’s interpretation makes engineers’ one-line diagrams legally consequential. A turbine connected to the public grid occupies one regulatory position; a turbine dedicated to a single campus may occupy another. Engineers draw that distinction with a meter and switchgear. Lawyers now read the same drawing for the boundary of federal obligations.

The EPA then proposed eliminating the federal requirement that states publicize and solicit comments on air-pollution permit applications, including applications covering data centers and their power plants. The proposal would not itself approve a turbine, and states could retain their own procedures. It would remove a federal process that currently inserts public notice and comment into the project schedule.

Gas is being financed as schedule insurance

Global Energy Monitor’s pipeline measures financing intent before construction. It captured developers and capital providers preparing to use gas generation as schedule insurance.

Between 2024 and 2025, AI construction schedules raised the value of delivery control. A dispatchable plant beside a campus can offer timing that a remote generator, constrained transmission line or uncertain interconnection cannot. Developers began paying for that control even when it carried higher fuel, pollution and permitting costs.

Lenders must price those emissions through permit conditions, potential litigation and operating constraints. A durable air permit joins long-term power agreements and completion backstops among the documents that turn an announced campus into bankable capacity.

Federal policy offered another architecture for synchronizing power and compute. President Biden’s January 2025 order directed the Department of Energy and Department of Defense to lease land for gigawatt-scale AI data centers alongside clean-power facilities. Whether developers choose gas or co-located clean generation, they must align land, power, permits and a usable connection on one construction sequence.

A permit shortcut is only as durable as its enforcement

The EPA showed the limit of any claim that dedicated AI power sits beyond environmental law. In January 2026, the agency ruled that xAI acted illegally by using dozens of methane gas turbines to power its Colossus 1 and Colossus 2 facilities near Memphis. The ruling kept behind-the-meter generation within the agency’s enforcement reach.

State lawmakers have challenged data-center construction without broadly halting it. They introduced 12 moratorium bills in 2026; 11 stalled or were voted out, while a Maine bill still awaited a final vote. Communities can contest permits, pollution and siting even when legislatures decline to halt development statewide. Local procedures can govern a project even as federal policy accelerates construction.

When the Supreme Court reversed Chevron deference, it gave agencies less latitude when statutes leave room for interpretation. EPA climate actions became more vulnerable to narrow readings or judicial reversal. Courts can constrain enforcement and overturn permissive interpretations alike. A letter may guide project design, but it does not remove the courthouse from the construction schedule.

By 2026, an AI campus could gain or lose compute capacity before a chip arrived. The seven-year wait could shrink—or survive—on a one-line electrical diagram: a turbine, a meter and a pencil stroke separating “public grid” from “dedicated load.”