AI Data Centre Cost Allocation: Who Pays for the Gap?

AI data centre cost allocation sounds like a dry tariff argument until a power plant is built for a customer that never turns up. The generator still needs paying. Someone receives the bill. On 9 October, the US Department of Energy backed a regulator's demand for PJM, the grid operator across a large part of the eastern United States, to explain how its proposed power procurement will charge the loads that caused the need. The argument is about more than new generation. It is about who carries the risk when an AI campus exists on a forecast but not on the ground.

The capacity shortfall is real. The customer list is less certain

PJM's proposed Reliability Backstop Procurement is a one-off attempt to secure new power capacity after its auction for the 2028/29 delivery year cleared 6,831 MW short of its reliability requirement. That number is an auction shortfall, not 6,831 MW of data centres waiting at a fence. The procurement target could fall if new supply is secured separately, including through bilateral deals with large electricity users. The proposed commitments could run for as long as 15 years.

The Federal Energy Regulatory Commission, or FERC, accepted the proposal for filing in September but suspended it for five months, subject to further proceedings. PJM did not start its planned September procurement. FERC invited PJM to file a revised plan sooner, which could avoid the full paper hearing. DOE's filing, served on 7 October and announced on 9 October, asks for that revision by 29 October. This is a request, not evidence that the auction has resumed or that the new allocation has been approved.

PJM needs supply. That much is plain. The awkward part is deciding whose demand caused the shortfall and whether those customers will actually connect. A power station cannot be ordered on the strength of a press release and then quietly handed to households if the project schedule slips. Conversely, refusing to procure until every campus is finished would leave the grid exposed. This is why a seemingly bureaucratic fight over forecast inputs can hold up real infrastructure.

The earlier power procurement bottleneck was whether enough electricity could be secured for AI infrastructure. PJM's dispute takes the next step: once capacity is procured, whose obligation is it?

A forecast can hide the load that caused the bill

Under PJM's original proposal, costs would first be assigned among zones using a measure of load growth, then passed to load-serving entities. FERC found a hole in the measure. Some big new loads may already have been folded into the baseline forecast, rather than recorded as later "Load Adjustments". A formula built around those adjustments could miss the very customers whose arrival made the procurement necessary. FERC did not rule that every data centre escapes payment. It said PJM had not shown that this allocation method fairly followed the causes and benefits of the costs.

Think of a site that was known early enough to be baked into the forecast. A neighbouring project is added later as a visible adjustment. Both may need capacity, yet a charge keyed to the adjustments risks catching the second and overlooking the first. The precise tariff treatment matters, but the commercial defect is easy to see. An accounting category can decide who pays for an asset that will be on the system for years.

FERC suggested that PJM consider updated forecasts, including the gap between actual load in the 2026/27 delivery year and the latest forecast for 2028/29, by zone. It also urged the most current load and resource forecasts for the initial procurement target. These were suggested routes for a revised filing, not an adopted final formula.

DOE wants PJM to identify material large-load projects at project level and track whether they arrive, shrink, stall or disappear. That would let the operator reconcile the baseline with subsequent adjustments and check for omissions or double counting. It is a practical demand. If the same megawatts appear twice, the market may buy too much. If a project drops out of the costing exercise, someone else can inherit its share. DOE's filing describes both risks; it does not claim PJM has already solved them.

The bill crosses a regulatory boundary

Even a better wholesale allocation cannot by itself guarantee that a particular data centre pays its whole share. FERC oversees PJM's wholesale procurement and the charges assigned through that system. States decide how those costs are recovered from retail customers. FERC says this division explicitly in its order. A wholesale charge can land on a load-serving entity; the state-level rate design decides how much of it reaches the new large load rather than other customers.

That distinction is easy to lose in the promise that "data centres will pay". A promise is not the same as an enforceable tariff, contract, deposit or rate class. Ask what happens if the customer signs up, the utility commits to infrastructure, and the customer delays by three years. Does the agreement hold the capacity reservation? Is there security if the customer exits? Who pays while the power is built and unused? These are questions for the final arrangements, not claims that a specific company has dodged a bill.

The stakes are not one-sided. Put every cost on a new site without a way to show it has brought its own qualifying supply and developers can pay twice. PJM's proposal allows confirmed new bilateral supply to reduce the central procurement target. FERC accepted that broad offset framework, while DOE urged a clearer route for large users to build, bring or buy capacity. Getting the offset right matters as much as getting the invoice right. Otherwise, the system can procure the same power twice, once in a private contract and again in a backstop.

Our earlier piece on the grid connection bottleneck asked whether a proposed site could actually get onto the wires. This case adds a second test: can the developer carry the cost of the capacity reserved in its name?

What the next filing needs to prove

The first test is the denominator. Which projects are in the baseline, which appear as adjustments, and which have dropped away? PJM needs a method that can be audited against a current project register without publishing commercially sensitive material wholesale. DOE's intervention calls for project-level identification and continuing status checks. FERC's order says a fixed allocation over a long commitment could be reasonable if it begins with accurate forecast information. The word "if" is doing work here.

The second test is risk after the contract is signed. A 15-year capacity commitment does not vanish when the load that justified it is late. The revised rules should make it clear who bears that exposure, what security is posted and when any obligation can change. FERC also asked for further examination of load-serving entity collateral and transmission-owner exit provisions. Those are separate issues in the proceeding, but they bear on whether the backstop can function when a counterparty fails.

Finally, watch the retail handoff. If the wholesale bill lands in the right zone but the retail bill is spread across existing customers, the cost-allocation headline has outrun the result. It is possible to build a tariff that looks fair at one layer and leaks at the next. The regulator, the state commission and the utility each need to show their part of the chain.

PJM has not bought this backstop capacity yet. Its original timetable was suspended; the revised filing is the next observable event. The useful question on 29 October will not be whether officials say the right customers should pay. It will be whether the tariff can find those customers when they move between forecasts, contracts and the physical grid.