Who Pays for the Next Gigawatt?

A data center pays its electricity bill. That fact alone does not settle whether its neighbors help finance the power system it needs. The bill for connecting a campus, the cost of keeping generation available, and the price of electricity delivered this evening are different obligations. A rule that assigns one of them to the developer can leave the others unresolved.

This is the distinction to watch as AI infrastructure moves from corporate spending announcements into utility contracts. Our earlier account of the race for grid access examined where companies can build. The next question is who carries the financial risk once a utility agrees to serve them.

Ohio offers a concrete example. In its July 9, 2025 decision announcement, the Public Utilities Commission of Ohio approved a settlement requiring a separate AEP Ohio data-center tariff. The commission identified the risk precisely: infrastructure built for customers whose eventual use might leave those investments underused. That is a problem of recovering committed costs. It is related to, but distinct from, a shortage that makes power more expensive for everyone.

Three bills behind one connection

The connection bill concerns the equipment and network work needed to serve a site. Its economic question is what happens after money has been spent if the customer arrives late, takes less capacity, or abandons the project. Contracts can place obligations on the customer before construction proceeds; whether other customers remain exposed depends on what those obligations actually cover.

The capacity bill pays for resources committed to be available when the system needs them. In PJM’s capacity market, electricity suppliers must secure resources for their customers’ demand and reserves. Capacity can come from owned resources, contracts, auctions or qualifying demand reduction. It is not a purchase of a particular quantity of electricity for immediate consumption.

The energy bill concerns electricity delivered over time. PJM’s explanation of its markets separates day-ahead and real-time energy trading from capacity procurement. A megawatt measures a rate of power use; a megawatt-hour measures energy. A campus drawing 100 MW continuously for one hour consumes 100 MWh. Reserving a 100 MW connection says nothing by itself about how many hours it will operate.

These distinctions matter for claims about subsidies. An unpaid share of an upgrade and a higher market-clearing price are different mechanisms. Both can affect existing customers, but they need different evidence and different remedies.

What Ohio’s minimum charge does

AEP Ohio’s tariff summary describes minimum billing demand, including a floor based on contracted capacity and a separate test tied to prior demand. The contract term includes a ramp period of up to four years followed by eight years. The structure gives the utility a payment commitment that extends beyond a customer’s immediate electricity consumption.

For an isolated 100 MW contract, the published capacity-floor formula above 75 MW gives 57.5 MW plus the 25 MW increment: 82.5 MW, below the 85 MW cap. At 200 MW, the uncapped result is 182.5 MW; the cap reduces it to 170 MW. These are our calculations of the contract-capacity component only, assuming full ramp-up. Actual demand and the prior-demand test can produce a higher billing basis. Affiliated loads and individual contract provisions also matter.

Neither result is an energy purchase in MWh or a complete dollar bill. The shorthand “85% of electricity” hides both the units and the sliding scale.

The protection is easiest to understand through a hypothetical. A developer reserves a large connection; the utility commits to construction; the developer then needs much less power than expected. Without a sufficiently enforceable commitment, the revenue expected to support the investment can disappear while the asset remains. A minimum charge can reduce that mismatch. It cannot establish that the original investment was appropriately sized, that every shared cost has been assigned correctly, or that the customer will always be able to pay.

That last point puts credit quality beside engineering. A long payment promise from an entity with limited assets is not equivalent to cash already collected. When reviewing a project, the useful questions are which entity owes the money, what security supports its promise, and what survives cancellation. The size of the announced campus is not an answer to any of them.

The regional bill is a separate decision

PJM’s July 31, 2026 reliability-backstop proposal addresses a different gap: securing additional supply. It proposes resource commitments lasting up to 15 years. Crucially, PJM says allocating retail costs among electricity suppliers and end users requires state participation. Buying the resource and assigning its cost to a particular data center are separate steps.

On August 13, PJM filed another proposed framework for large customers and an Interim Resource Adequacy Service. The proposal would use new supply commitments, a large-load registry and emergency demand reductions to manage the reliability risk. Retail allocation and implementation would still involve states and local utilities. These are descriptions of PJM’s filings, not a claim that their proposed provisions are already effective rules for every customer.

Our reading is that a credible protection package has to connect these levels. A developer’s contribution to a substation does not demonstrate that it has also funded enough dependable generation. Conversely, paying for generation does not prove that local network costs have been recovered. A promise to bring power needs a location, a delivery date, an enforceable obligation and a treatment for delays.

There is also a timing problem. Consider a hypothetical campus that starts consuming before the additional generator intended to serve it becomes available. Even if both projects eventually operate as planned, other buyers can face tighter supply in the interval. The size and duration of that effect require a market analysis; they cannot be inferred from the campus’s construction budget. Nor can a household bill increase be attributed entirely to AI without separating fuel costs, weather, retirements, network spending and the household’s own consumption.

What would demonstrate that the protection works?

For a particular project, we would look for four pieces of evidence:

An allocation of construction costs: which dedicated and shared upgrades the customer must fund, and which costs remain with other users.
Protection against a smaller or cancelled project: payment floors, security and enforceable exit obligations matched to the investment at risk.
A supply commitment with a credible timetable: what additional dependable capacity arrives, when it arrives, and who covers a delay.
The final retail treatment: how an approved tariff or rate order assigns the remaining costs, followed by evidence from actual billing and collections.

This is our framework for evaluating a project, not a claim that every jurisdiction uses the same rules. It also leaves room for a favorable outcome: a well-secured new customer can support investment and contribute to shared costs. Establishing that benefit requires the same discipline as establishing a subsidy—compare the system with the project against a credible case without it, and identify who receives the gains and bears the risks.

The AI infrastructure dossier follows spending, financing and the eventual economics of the build-out. Grid obligations belong in that analysis even when they are absent from a headline about buying servers. The strongest evidence that households are protected will be contracts and rate decisions that remain effective when the data center’s business plan changes.

Sources and scope

This analysis was prepared on September 23, 2026. The numerical examples apply the contract-capacity component in AEP Ohio’s published summary; they do not reproduce a complete customer invoice. The Ohio example concerns AEP Ohio’s service territory. PJM’s filings are discussed as proposals rather than assumed approvals. No estimate of a household bill increase is made.