FERC's Data Center Docket Will Decide Who Pays for AI's Power

AI's electricity problem has reached the regulator's desk.

The question is no longer only whether data centers can find enough power. It is whether they can get fast grid access without shifting too much cost and reliability risk onto everyone else connected to the same power system.

That is why the FERC data center docket matters. Docket No. RM26-4-000 is not a software story, and it is not a simple energy-demand story. It is where the AI buildout becomes a tariff, interconnection and ratepayer problem.

Why the FERC data center docket matters now

FERC opened the large-load interconnection proceeding after the U.S. Department of Energy pushed the Commission to consider reforms for connecting very large electricity users to the transmission system.

The category is broader than AI. It can include manufacturing, industrial facilities and other major loads. But the political force behind the docket is obvious: AI data centers are arriving as sudden, concentrated, power-hungry customers.

FERC said on April 16, 2026 that it would act by June 2026 on the proceeding. The Commission described rapid demand growth from data centers and other large customers as a shift reshaping the transmission landscape, and said staff had reviewed more than 3,500 pages of public comments.

That timing matters. AI companies are trying to build compute capacity on venture, cloud and geopolitical timelines. Utilities and transmission systems move through planning cycles, equipment lead times, regulatory approvals and cost-allocation rules. The mismatch is becoming one of the most important constraints in compute infrastructure.

The docket asks a hard question: can the power system connect large loads faster while still being reliable, fair and legally durable?

The cost question under the AI power boom

The core issue is who pays.

FERC's docket explicitly asks whether large loads and co-located facilities should pay the full cost of grid upgrades needed for interconnection. It also asks whether those costs should later be credited back, and over what period.

That is not a technical footnote. It is the public bargain underneath AI infrastructure.

If a data-center campus needs transmission upgrades, substations, studies, backup service or additional capacity, the money has to come from somewhere. It can be charged directly to the project. It can be spread through broader utility rates. It can be handled through market rules. It can be blurred through incentives, delayed upgrades or future customer bills.

Each choice changes the economics of AI.

Direct assignment makes private compute projects carry more of the cost they trigger. Broad socialization can make projects easier to build, but it risks asking households, small businesses and unrelated industrial users to finance infrastructure whose benefits mainly flow to hyperscalers and AI platforms.

That is why this docket belongs in the AI story. The AI race is not only about chips, models and training clusters. It is also about which costs stay inside the data-center business model and which costs leak into the shared grid.

Co-location is not a free escape from the grid

Data-center developers have a natural answer to grid congestion: bring power closer to the load.

That can mean co-locating a large load with an existing or planned generator. In theory, a data center gets electricity without waiting for every normal grid upgrade. In practice, the arrangement still raises grid questions.

FERC's December 18, 2025 PJM order shows why. The Commission directed PJM to create clearer rules for AI-driven data centers and other large loads co-located with generating facilities. FERC said PJM's tariff lacked clear and consistent rates, terms and conditions for those arrangements.

The reason is simple. A co-located data center may still rely on the grid for backup, balancing, ancillary services, emergency support or market participation. It may affect capacity rights when a generator serves an on-site customer instead of injecting power into the regional system. It may create reliability effects that are not visible if everyone pretends the load is separate from the network.

PJM's own workshop material makes the tradeoff plain. Large data-center loads want time to market. PJM wants resource adequacy, reliable operations and cost allocation that reflects how the transmission system is actually used.

Those goals can conflict.

A data center can look private when it is buying power. It can look public when it needs backup from the grid.

Ratepayers are becoming part of the AI stack

The ratepayer is now inside the AI infrastructure story.

Vastkind has already covered why AI's grid bottleneck is transformers and why utilities are becoming part of the compute stack. FERC's docket adds the governance layer: who absorbs the bill when those bottlenecks have to be solved quickly.

Monitoring Analytics, the Independent Market Monitor for PJM, has argued that data-center load growth is a major driver of recent and expected PJM capacity-market conditions. Its 2025 complaint against PJM asked FERC to clarify that large data-center loads should be added only when they can be served reliably under transmission and capacity adequacy.

That does not prove every electricity-price increase is caused by AI. The grid has many pressures: retirements, fuel prices, transmission constraints, weather, policy, aging infrastructure and normal load growth.

But it does show the institutional pressure. Large AI loads are no longer invisible customers. They are large enough to change planning forecasts, capacity markets, interconnection queues and public arguments about affordability.

Once that happens, the data center stops being just a private facility. It becomes a claim on a shared system.

What remains unsettled

The boundary is still moving.

As of June 15, 2026, the final shape of FERC's June action was not yet public. The docket could move toward standardized study rules, faster processes for flexible loads, clearer cost responsibility, transitional handling for projects already in the queue, or some combination of those.

There is also a jurisdiction problem. Electricity regulation is split between federal and state authority. FERC has authority over interstate transmission and wholesale markets. States still control large parts of retail service, utility siting, resource planning and local rate design. NARUC's comments in the docket warned that federal action could run into state-jurisdiction concerns if the rules are not carefully drawn.

That matters because AI infrastructure is not built in one legal layer.

A hyperscaler may negotiate with a utility, a generator, a state economic-development office, a regional grid operator, a transmission owner and a local government. Each actor controls a different piece of the path from power plant to rack.

FERC can clarify part of that path. It cannot make the entire political economy of data-center power disappear.

Why This Matters

The FERC data center docket matters because it turns AI power into a public allocation problem.

If a data center pays for the upgrades it triggers, the economics of compute become more honest. If costs spread through the system, AI capacity can appear cheaper than it really is because part of the bill moves into public rates, capacity markets or future grid planning.

That is the shift most AI coverage misses.

The central bottleneck is not only megawatts. It is permission to claim megawatts on terms the rest of the system can live with.

Utilities have to keep the lights on. Regulators have to prevent unfair cost shifting. Grid operators have to manage resource adequacy. Local communities have to absorb land, water, noise, transmission corridors and reliability risk. AI companies want capacity fast enough to match model roadmaps and cloud demand.

FERC's docket is where those timelines collide.

What to watch next

Watch whether FERC makes cost causation explicit. The stronger the rule, the harder it becomes for large compute loads to hide grid costs inside general customer bills.

Watch whether flexible or curtailable data-center service becomes a serious path. If AI workloads can accept limits during stressed grid conditions, some projects may connect faster. If they demand firm service around the clock, they should expect a different cost profile.

Watch PJM first. It is the largest U.S. grid region, serves many of the data-center-heavy areas in the Mid-Atlantic, and has become the clearest early battlefield for this question.

And watch whether AI companies start talking differently about power. A serious compute builder will not only announce megawatts. It will explain interconnection status, upgrade responsibility, curtailment rules, backup emissions and ratepayer exposure.

That is what adulthood looks like for AI infrastructure.

The useful next question is not whether AI will use more electricity. It will.

The question is whether the bill follows the compute.

Read next: start with AI's Grid Bottleneck Is Transformers for the physical constraint, then read AI Data Center Power Demand Is Turning Utilities Into Compute Infrastructure for the utility-strategy layer.