A developer can buy land, order servers and announce a major AI campus years before the electricity grid can supply it. Across the eastern United States, that mismatch is turning proposed data-center capacity into a question about transmission, connection rules and reliability, rather than just server procurement.

On November 17, 2026, Reuters reported that developers want greater clarity on large-load rules in the PJM region. PJM coordinates wholesale power markets and bulk-grid operations across parts of 13 US states and the District of Columbia. Demand for computing remains strong. What is less certain is when every announced project can actually draw the power it expects.

Twenty gigawatts of projects are not twenty gigawatts of delivered power

Reuters cites Wood Mackenzie estimates that around 20 gigawatts of new data-center projects entered active development in PJM during the first half of 2026. That is development-pipeline capacity, not a measurement of electricity being consumed and not a guarantee that those projects will all connect.

This difference is often lost in discussions of AI power demand. Megawatts (MW) measure instantaneous power. Megawatt-hours (MWh) measure energy delivered over time. A hypothetical facility drawing a steady 100 MW for 24 hours consumes 2,400 MWh. That arithmetic does not describe the utilization of any particular announced campus.

Some proposals will be delayed, redesigned or canceled. Others may expand. Treating every application for a connection as a completed project exaggerates near-term supply requirements, while dismissing the entire queue as speculative would be equally irresponsible for planning.

There is also a distinction between the servers' IT load and the facility's total electrical demand. Cooling, power conversion, pumps and supporting systems add consumption. Workload utilization changes the actual draw over time. An announced server count or GPU budget cannot be converted directly into a precise grid-demand forecast without those assumptions.

A region can have power but lack capacity at the right location

Electricity has to reach a facility through physical equipment: generation, high-voltage transmission, transformers and local connections. Available generation somewhere in PJM does not mean a particular substation can supply another major campus on the date its developer prefers.

PJM's load-forecasting explanation says forecasts inform both day-to-day operations and long-term decisions about generating resources and network upgrades. Fast-growing forecast demand puts pressure on planning schedules because building transmission and substations is not as quick as ordering computing equipment.

Who pays for the upgrades is contentious. Developers want predictable costs and connection dates. Regulators and other electricity consumers need to avoid paying for infrastructure built around projects that never materialize. A straightforward 'build more power plants' response does not resolve the siting, network and cost-allocation questions.

A prospective tenant can therefore find that one data-center location has available electrical capacity while a seemingly similar location has a multi-year connection problem. Rack density is not a complete description of deliverable cloud capacity.

Four gigawatts vanished from grid demand in one event

PJM's concerns extend beyond steadily rising consumption. Large data centers can change their electrical load suddenly.

In a September 10 account of a July 22 incident, PJM said nearly 4,000 MW of data-center load unexpectedly disconnected in northern Virginia. The facilities transferred to their own backup generation. Grid operators then had to manage the resulting imbalance between demand and supply and associated voltage and frequency effects.

The event did not create unresolvable systemwide reliability impacts, according to PJM. It also did not mean that 4,000 MW of computing equipment had been destroyed. The concern was an abrupt withdrawal of a large load from the grid.

PJM consequently proposed requirements addressing ride-through, meaning that certain large facilities should remain connected through normally cleared transmission disturbances instead of disconnecting together. Such behavior requires careful engineering: backup power is valuable for an individual facility, but simultaneous switching across many sites can create problems for the wider electricity system.

This reveals an unusual trade-off. A data center's locally optimal response to electrical disturbance may be destabilizing when many sites take the same action at regional scale.

The temptation to generate power on site

Developers sometimes consider on-site generation or alternative power arrangements to control scheduling and reliability. Those options can reduce dependence on particular grid upgrades but do not make fuel, emissions, maintenance or interconnection constraints vanish.

A backup generator designed for occasional outages is not the same economic or engineering proposition as a primary power plant. Battery systems also have finite stored energy: a high short-duration power rating does not imply the ability to sustain a campus for days. Fuel delivery can introduce another shared infrastructure dependency.

Nor is a local generating plant automatically isolated from regional grid rules. Whether it runs connected to the system, exports electricity or serves a dedicated load changes the relevant technical and regulatory questions.

Our earlier analysis of AI rack power and cooling examines what happens inside the facility. The PJM debate concerns the equally important boundary outside it.

What cloud and hosting buyers should infer

The effect is not limited to the largest AI training labs. New electrical infrastructure, transformers and transmission capacity must compete for equipment, construction workers and physical corridors. Smaller cloud and colocation providers operating in the same region may face different costs or delivery dates as a result.

Regional diversification can reduce exposure to a local electricity problem, but it does not guarantee resilience. Two zones may share a grid, transmission corridors or other supporting services; see our discussion of correlated failures across cloud regions and availability zones.

The most useful way to examine a new deployment is to compare three independent dates: hardware availability, confirmed electrical-service readiness and full facility commissioning. The slowest determines when capacity is commercially usable. For major AI campuses, a public announcement can precede that date by years.

PJM's debate does not prove that AI infrastructure growth must stop. It shows that compute capacity is no longer just a quantity of chips. It is the amount of computing that a particular site can power, cool, connect and operate under real electrical constraints.