Behind the meter generation, the practice of producing power on site and bypassing the local utility, is the fast growing workaround the biggest AI and cloud operators are using to escape multi year grid queues and reshape who pays for the wires.
On June 24, 2026, FuelCell Energy and Fit Energy announced up to 380 megawatts of modular fuel cell power for AI data centers, with an initial 30 MW equipment deposit already paid. 380 MW is roughly the continuous output of a mid-sized gas plant. One week earlier, Chevron and Microsoft had signed a 20-year power supply agreement for a West Texas data center campus, a project that Construction Review valued at roughly $7 billion. Neither deal passes through the local utility meter. Neither waits for an interconnection queue. The largest power customers in the country have stopped waiting for the grid to catch up. They are building around it.
The phrase that captures this shift is "behind-the-meter." It is industry shorthand for power that is generated on the data center operator's side of the property line, flowing straight from generators into server racks without ever passing through the local utility meter or the public grid. The Construction Review framing of "private utility empires" is the visual: a hyperscaler campus that looks, from the road, like a small city's worth of generation and load sitting behind a single fence. The bigger question is why the biggest buyers of electricity in the world are suddenly opting out of the system that was built for them.
The answer is speed, not ideology. The regulated grid in the United States was designed for a world where the largest single load was a factory shift change or a steel mill ramping up. The processes that connect a new megawatt of demand to the transmission system, from utility integrated resource plans to state public utility commission dockets, regional transmission organization studies, and multi-year interconnection queues, were calibrated for that scale and that pace. The 2 GW Pecos, Texas campus that Microsoft and Chevron are developing, or the 380 MW block that FuelCell Energy and Fit Energy are staging for delivery later in 2026, do not fit on that schedule. Hyperscaler buildout timelines are measured in months, and the construction crews show up before the dockets close.
That gap is not theoretical. The U.S. Energy Information Administration's 2026 analytical work on data center load growth and firm clean power options names 24/7 firm power as a distinct planning problem, and treats on-site nuclear and fuel cell generation as candidate solutions precisely because grid-side delivery cannot be assumed on operator timelines. The economics work in the same direction. Once a hyperscaler has decided that a campus will be live on a specific date, the cost of waiting for an interconnect is no longer a utility tariff. It is a deferred revenue line, a missed model training window, a contractual penalty. The cheapest kilowatt is the one that arrives on schedule.
The result is a quiet but consequential rewiring of who pays for the grid. The regulated utility model rests on a cost-sharing arrangement. Large industrial and commercial customers fund a disproportionate share of transmission and distribution maintenance, and residential and small commercial rates are set on top of that base. When a 2 GW customer defects to on-site generation, the maintenance burden it would have carried does not disappear. It is spread across the remaining ratepayers. That is the structural critique embedded in the "private utility empire" framing, and it is the part of the story that the headline scale of the deals can obscure.
The pivot is also contested on other grounds. On-site gas-fired or fuel cell generation raises direct questions about methane leakage, water use in arid siting regions, and air permitting at a scale that the local regulator was not staffed to review. The West Texas siting choices that make sense for cheap fuel and low population density also concentrate those environmental costs in counties with limited review capacity. The fuel cell path that FuelCell and Fit Energy are pursuing is cleaner at the point of generation, but the same load-departure problem applies. A clean megawatt behind the meter still leaves a fossil megawatt somewhere else paying the same grid bill.
Nuclear points to a different resolution. Amazon's acquisition of a nuclear-powered data center from Talen, followed by the expanded supply agreement that grew the Talen-Amazon partnership, anchors nuclear as the firm clean power option that data center operators are actively buying. In that case the on-site generation is a steam-side connection to an existing reactor, not a parallel grid. It keeps the megawatt hours inside a regulated rate structure rather than defecting from one. Whether regulators and grid operators treat nuclear retrofits as grid assets or as private bypasses will shape how fast that model scales next.
The question to watch is not whether the behind-the-meter trend continues. The FuelCell equipment deposit, the 20-year Chevron-Microsoft agreement, and the Amazon-Talen expansion are already on the books, and the broader pattern of hyperscalers cutting direct power supply deals is now the default rather than the exception. The question is whether state public utility commissions, regional grid operators, and the federal regulators who set transmission policy treat that defection as a problem to be priced or a precedent to be accommodated. The first major commission ruling that forces a hyperscaler to either pay its allocated share of stranded transmission costs or surrender the on-site option will do more to clarify the future of AI power than any new gigawatt announcement.