Layer 1: The Reportable Facts

Reuters reported on June 16, 2026 that at least 57 off-grid U.S. power plants are proposed or under construction to serve individual data centers, with combined capacity of about 73,000 MW. The report said more than a dozen such projects had won approval in under a year and described a pattern of developers arguing that private, off-grid plants for single customers are exempt from parts of the normal review process. Reuters identified operating examples including xAI outside Memphis and a facility in Ashburn, Virginia serving Vantage Data Centers, and highlighted the Apollo Generating Station in Wood County, Ohio, built to serve Meta’s Bowling Green data center.

The Washington Post reported in February 2026 that data-center developers were building what it called a shadow power grid: private power plants, mostly fueled by natural gas, for large AI and cloud campuses across states including Texas, New Mexico, Pennsylvania, Wyoming, Utah, Ohio and Tennessee. The Post’s review, which relied in part on Cleanview research, described off-grid power as a response to utility delays, grid pushback and the need to bring AI capacity online faster than traditional interconnection timelines allow.

Cleanview’s June 2026 report identified 59 behind-the-meter data-center projects with roughly 90 GW of announced capacity, while also cautioning that most of that capacity remains planned rather than operating. Cleanview estimated that about 2 GW was online as of mid-2026, much of it tied to xAI’s Colossus sites near Memphis, and projected roughly 2.8 GW to 3.2 GW online by the end of 2026 if near-term projects progress. That distinction matters: the announced shadow grid is enormous, but the built shadow grid is still early.

Industry-side data points in the same direction. A Bloom Energy-sponsored 2026 Data Center Power Report, covered by Data Center Dynamics, said roughly one-third of hyperscalers and colocation providers expected to operate fully onsite-powered campuses by 2030. The same report said utility delivery timelines were running about 1.5 to 2 years longer than hyperscalers and colocation providers expected in key hubs, making onsite power a schedule tool rather than only a resilience feature.

Ohio is an early test case for the permitting side of the pattern. The Ohio Legislative Service Commission’s 2025 utilities digest says changes to Power Siting Board procedures required a completeness determination within 45 days, shortened hearing timing for certificate applications and required certificate decisions within 150 days after an application is deemed complete. Separate Ohio law also allows accelerated review for certain major utility facilities, a framework Reuters connected to faster approval pathways for AI-related power projects.

Layer 2: The System Read

Verified fact: AI data centers are creating demand that can exceed what utilities can deliver on the timeline hyperscalers want. System read: the industry’s response is vertical integration. The same companies and infrastructure partners that already secure chips, land and fiber are now trying to secure electrons through private generation. That changes electricity from a public utility service into a controllable input for AI production.

The strategic logic is straightforward. A gigawatt-scale AI campus is valuable only when it is energized. If GPUs sit in warehouses waiting for grid upgrades, the bottleneck is no longer silicon; it is permitting, turbines, gas interconnects and local approvals. Behind-the-meter power converts a regional constraint into a private project-management problem. It also gives the largest buyers an advantage smaller competitors cannot easily match.

The governance tradeoff is equally clear. Utilities are slow partly because grid planning, transmission buildout, environmental review and ratepayer protection are slow. Private onsite plants can move faster, but speed can come with opacity: shell entities, nondisclosure agreements, redacted filings, compressed hearing windows and residents learning about industrial energy assets after site work has begun. The infrastructure is private, but the air, roads, emergency services and land-use impacts are local.

This is why the AI power story is becoming an industrial-policy story. States that want data-center investment are rewriting siting and permitting pathways to attract campuses. Developers are choosing locations not only for tax incentives and fiber access, but for gas supply, permissive power-market rules and faster approvals. The map of AI capacity may increasingly follow the map of energy flexibility.

The central tension is that behind-the-meter does not mean consequence-free. If private plants compete for scarce turbines, gas pipeline capacity or skilled construction labor, they can affect the broader power system even when they are not formally drawing from the grid. If they run on gas, they also add air-pollution and greenhouse-gas questions that corporate renewable matching does not fully resolve at the local smokestack.

Layer 3: What To Watch Next

First, watch which announced projects become real steel. Cleanview’s distinction between announced, permitted, under-construction and operating capacity is the key filter. The market has many giant press releases; the signal is air permits, turbine deliveries, gas interconnection agreements, satellite-visible construction and signed compute tenants.

Second, watch state siting laws. Ohio, West Virginia, Texas and other power-hungry or gas-rich states are becoming laboratories for faster AI infrastructure approvals. The next phase will likely include legal challenges, local ballot fights, emergency-response questions and attempts to define when a private data-center power plant should be treated like a conventional public-facing energy facility.

Third, watch whether hyperscalers keep project sponsors at arm’s length. Reuters and The Washington Post both described cases where the ultimate tech customer was not always obvious from early paperwork. If public backlash grows, major AI buyers may be pressured to disclose more about counterparties, emissions, emergency plans and how behind-the-meter generation fits with their climate commitments.

Fourth, watch reliability. A data center wants near-continuous uptime, while gas turbines, fuel cells and mobile generators require maintenance and fuel logistics. If behind-the-meter campuses suffer outages, operators may seek backup grid service anyway. That would reopen the question of who pays for standby capacity and whether private power truly reduces public-grid burdens or merely changes their timing.

Fifth, watch the accounting. The next policy fight will not be only megawatts; it will be cost allocation. Utilities, regulators and consumer advocates will ask whether private AI power reduces ratepayer exposure or shifts indirect costs into grid planning, fuel markets, emergency services and environmental enforcement. The answer will shape whether behind-the-meter AI power becomes a durable model or a transitional scramble.

Conclusion

The move behind the fence is not a side story. It is the next layer of AI infrastructure competition. If hyperscalers can build private power faster than utilities can expand the grid, the winners of the AI race may be determined as much by air permits, gas turbines and county politics as by model architectures. The open question is whether the public gets a transparent bargain from that acceleration—or merely discovers the power plant after the data center has already arrived.