PJM’s AI Power Bill Leaves the Server Room

76% was the move. PJM’s average power price rose that much year over year in the first quarter, to $136.53 per megawatt-hour, amid data-center demand. The server owner bought the compute; the grid registered the premium.

Grid scarcity turns AI into a public-cost event

Hyperscalers buy the GPU, server rack, cooling system, and campus. Grid operators must supply the generation, transmission, connection capacity, and operating margin required to serve that concentrated load without weakening the surrounding system.

PJM makes the transfer visible. The operator of a 13-state grid expects demand to grow 4.8% annually over the next decade, driven by the AI boom. In parts of the same territory, electricity bills are projected to rise by more than 20% as data centers strain the grid.

These figures measure different things. One is a demand forecast. One is a customer-bill projection. The 76% figure is a wholesale-price move over another period. Collapsing them into a single causal estimate would be tidy and wrong.

Together, they trace the same path. A concentrated buyer adds demand faster than physical capacity adapts. Regional power prices and customer bills can move before the requested supply comes online. The buyer keeps the compute upside, while the grid retains the reliability obligation.

A hyperscaler paying for its own GPU does not settle who pays for the system around it. That gap is AI infrastructure becoming utility infrastructure: the regional grid must balance everyone attached to it.

Campuses displaced queries as the unit that matters

Over three years, the unit of concern widened from the query to the campus, then to the customer bill and the grid connection.

• 2023: AI electricity intensity remained uncertain, leaving efficiency as a material offset.
• 2024: Data-center construction ran into shortages of parts, property, power, and cooling equipment, whose lead times had increased fivefold.
• 2025: PJM bill projections moved the issue from corporate procurement into household and business budgets.
• 2026: PJM proposed requiring large data centers to self-supply or curtail to reduce outage risk.

The daily narrative still oscillates between “AI needs a small country’s electricity” and “models are getting more efficient.” Both can be true and miss the structure. Efficiency changes the amount of compute obtained per unit of power. It does not guarantee that generation, transmission, and interconnection capacity exist where a new campus wants them.

The question is no longer only how many kilowatt-hours a model uses. It is who has the right to draw firm power, at what location, and under which reliability conditions.

Local wires, not national supply, set the limit

A national electricity total is the wrong denominator. Data-center load is geographically lumpy. The grid must absorb it at a specific place, through specific infrastructure, while preserving service to existing customers.

Dominion Energy reported that data-center power demand in Virginia almost doubled during the second half of 2024, reaching 40.2 gigawatts in December. In metropolitan Atlanta, data-center capacity is projected to exceed 4,000 megawatts by 2028, roughly 30 times its 2012 level.

Those figures describe concentration, not merely growth. A country can possess enough electricity in aggregate while a particular region lacks the generation, wires, or firm connection needed for another large campus. National totals hide the point of failure.

Data-center developers can announce demand quickly, but utilities cannot build physical supply on the same schedule. That mismatch is capacity lag.

Announced demand is not realized demand, however. PJM cut its projected summer 2027 peak to about 160 gigawatts from about 164 gigawatts because some projects, including data centers, lacked firm service or construction commitments—a reminder that campus announcements are not meter readings.

The haircut identifies the correct base rate without erasing the bottleneck. Some proposed load will disappear, slip, or become more efficient. Firm projects still arrive through a system whose adaptation time is measured in years.

Large loads are becoming grid instruments

Historically, a commercial customer bought electricity and the system served it. PJM’s proposed bargain is different. A sufficiently large data center may have to bring on-site generation or accept curtailment when reliability is threatened. U.S. officials have also proposed requiring data centers to power down or switch to backup supply as blackout risk rises.

Those rules recast the data center from passive demand into a managed reliability participant.

ERCOT demonstrated the economic logic before the AI buildout reached its present scale. In August 2023, Riot Platforms received $31.7 million in energy credits for curtailing electricity use. The bitcoin it mined that month was worth about $8.9 million, giving Riot more than three times as much value for not consuming power as for mining.

Riot was a bitcoin miner, not an AI data center, and the evidence does not establish that AI workloads can curtail as readily. It establishes something narrower and more useful: interruptibility has economic value when electricity is scarce.

Grid operators can reduce the reliability risk from large loads if those customers genuinely curtail, shift to backup power, or supply themselves. PJM’s proposal tries to turn flexibility into an operating resource rather than merely impose a penalty.

Contracts alone cannot provide that resource. Grid operators need verified behavior during system stress, not promises with a diesel generator in the footnotes.

Private generation relocates the bottleneck

Developers facing grid-access waits of up to seven years have turned to aeroderivative turbines and diesel generators. Data-center demand has also contributed to a supply crunch for gas turbines. The procurement race is moving down the stack: from chips to servers, then land, turbines, fuel, commissioning capacity, and operating permits.

The same shift appears in Amazon’s energy-to-compute strategy. Its reported Texas plan pairs an off-grid AI campus with 7.65 gigawatts of gas generation, while a separate $2 billion facility in Gilroy required years of local negotiation. Control of power changes the path to operation.

It changes financing as well. An Ohio arrangement covered in AI data-center financing pairs a 20-year, 10-gigawatt OpenAI agreement with an Nvidia backstop on part of the completed facility’s value. Compute demand becomes financeable when power access, customer commitment, and residual asset value are bound together.

Behind-the-meter generation changes the interface with the grid rather than eliminating it. Data Center Alley illustrates the risk: the 30-square-mile region outside Washington contains more than 200 data centers and consumes roughly as much electricity as Boston. Unannounced disconnections can occur when facilities switch from grid power to local generators, and those sudden changes may contribute to cascading outages on the broader system.

A self-supplied campus must therefore coordinate with the system it is trying to avoid. The campus benefits from independence, but the grid needs synchronized behavior when thousands of megawatts connect or disconnect.

Ratepayer politics follow the bill

AI infrastructure once looked remote to most households. The server room was somewhere else, the capital spending belonged to Big Tech, and the consumer encountered the result as software. Electricity bills collapse that distance.

In one survey, 81% of economists said the U.S. AI buildout would add to inflation over the following year, including through electricity and software prices. Another found that 75% of U.S. adults had heard at least a little about data centers, while 38% viewed them as bad for home energy costs.

Public opinion is not a causal model, but it identifies the legitimacy constraint. Once voters associate data centers with higher bills, the allocation rule becomes inseparable from AI adoption.

The burden extends beyond prices. Data-center construction and operations have been linked to blackouts and water shortages in vulnerable communities globally. Nearly 60% of the 1,244 largest data centers measured as of June 2025 were outside the United States, so the physical burden is not confined to the markets capturing the largest corporate valuations.

Supporters have a substantive case. Data centers can support advanced technologies and domestic supply chains through demand rather than subsidies or tariffs. They can finance generation, equipment, construction, and technical capacity that might not otherwise be built.

But productivity and distribution remain separate columns. A project can expand national technological capacity while imposing local costs, and an economy can gain in aggregate while particular ratepayers fund the transition.

Power-market design is therefore becoming AI policy. Grid operators and regulators must decide who pays for incremental capacity, which loads receive firm service, who earns compensation for curtailment, and which communities absorb the environmental and reliability burden.

Reliable power carries the infrastructure premium

The early AI race rewarded access to compute. The next layer rewards control over the conditions that make compute operable: firm generation, deliverable transmission, connection rights, fuel, and credible flexibility.

A developer can wait for grid capacity, bring generation, accept curtailment, or fund enough incremental infrastructure to change the allocation argument. Each choice assigns the premium to a different balance sheet.

Grid operators gain leverage because they govern reliability. Utilities gain leverage because they control deployment speed in constrained territories. Turbine suppliers gain leverage because private generation requires scarce equipment. Large technology companies retain leverage through capital, but capital alone cannot compress every physical timeline.

Electricity will not replace chips as the only bottleneck. As companies finance away one constraint, cooling, land, transformers, turbines, or transmission can bind next. A buyer choosing a site must now price not just compute and land but years to firm power, curtailment obligations, backup generation, and turbine availability.

The 76% move does not forecast another 76%. It shows that the meter has changed jurisdiction: concentrated data-center demand meets a regional grid that must serve everyone at once. The server owner still buys the GPU; PJM’s $136.53 per megawatt-hour lands on another ledger.