The A.I. Power Bill Is Arriving

Data-center expansion is no longer only a technology or infrastructure story. Across the largest regional electricity market in the United States, households and businesses are beginning to absorb the cost.

Category: Industry. Written by Jaime Garcia, Founder, SnowRock. Published May 19, 2026. 15 min read.

In short

• A single regional capacity auction is expected to add roughly $6.3 billion to electricity costs across 13 states, driven substantially by data-center demand.
• The core conflict is less about the fact that data centers use power and more about who pays for the supply, transmission and reliability they require. Today much of it is spread across ordinary ratepayers.
• For a mid-market operator, the lesson is efficiency: right-sized, measured AI is insulated from the cost curve that is now reshaping the grid.

The rapid construction of artificial-intelligence infrastructure is creating a new class of economic pressure: the cost of supplying enough electricity to support it.

An annual power-market auction conducted by PJM Interconnection, the largest regional grid operator in the United States, is expected to add approximately $6.3 billion to electricity costs across 13 states and the District of Columbia over the coming years. The increase is being driven substantially by data centers.

These facilities require immense and continuous supplies of electricity to operate servers, cooling equipment, networking systems and backup infrastructure. As artificial-intelligence companies and cloud providers build larger computing campuses, electricity demand is rising faster than new generation can be connected to the grid. The result is a widening imbalance between the amount of power the region will need and the amount suppliers can reliably provide during periods of peak demand, and that imbalance is now being priced into consumer electricity bills.

How data center electricity prices reach a household power bill

Electricity markets are complex, but the underlying mechanism is straightforward. PJM is responsible for maintaining the reliability of a vast regional grid serving approximately 67 million people. Its territory stretches from the Mid-Atlantic to parts of the Midwest and includes Northern Virginia, the largest concentration of data centers in the world.

Each year, PJM conducts a capacity auction. Power producers submit the prices at which they are willing to guarantee that electricity will be available during future periods of highest demand, and PJM selects enough capacity to meet projected needs and maintain a reserve margin for emergencies. The price established in the auction becomes part of the broader cost paid by utilities and, eventually, their customers. When projected demand rises while supply remains constrained, the auction price increases. That is what is occurring now.

Data-center developers are seeking enormous amounts of new electricity, frequently in areas where transmission systems and power generation were designed for much slower growth. New plants, batteries and transmission lines cannot be planned, approved and built at the same speed that a large technology company can announce a computing campus, so the market pays more to secure the available supply. Those costs do not remain confined to the companies creating the demand. They are distributed through the regional electricity system, affecting residential customers, small businesses, manufacturers and public institutions. A household that has never used an artificial-intelligence service may still pay more because one was trained or operated nearby.

The Socialization of Infrastructure Costs

The central policy conflict is not whether data centers consume electricity. That is unavoidable. The conflict concerns who should pay for the infrastructure required to support them. Under the existing market structure, a large portion of the cost is spread across all electricity customers, because capacity charges, transmission investments and reliability expenses are incorporated into utility rates throughout the region.

This creates an asymmetry. Technology companies capture the commercial value generated by data centers; they sell cloud services, develop artificial-intelligence models and accumulate computing assets whose value may reach hundreds of billions of dollars. The public absorbs part of the infrastructure burden. Residents may pay higher electricity rates. Utilities may need to reinforce local networks. Governments may finance roads, water systems and tax incentives. Existing customers may face a tighter power supply during extreme weather.

The arrangement resembles other periods of industrial expansion in which public infrastructure was built to support private economic growth. The difference is the scale and speed of the current demand. A hyperscale data center can require as much electricity as a city, and a cluster of them can reshape the economics of an entire regional grid. When several large projects seek connections simultaneously, the cost is no longer marginal; it can alter electricity prices for tens of millions of people. The $6.3 billion increase attributed to the latest auction is part of a larger pattern: since 2024, PJM capacity auctions have added an estimated $29 billion in costs across the region, with data-center demand identified as a major contributor.

The artificial-intelligence economy is beginning to produce a visible utility bill.

A Grid Built for a Different Era

The American power system was not designed for sudden concentrations of industrial demand on this scale. For much of the past two decades, electricity consumption in many parts of the country grew slowly, as improvements in energy efficiency offset growth from population and technology, and utilities and grid operators planned around relatively stable demand. Artificial intelligence has disrupted those assumptions.

Training and operating advanced models requires dense clusters of high-performance processors. Those chips consume substantial amounts of electricity and generate heat that must be removed continuously, and companies are also building redundancy into their facilities, increasing the power and physical infrastructure required. Unlike many traditional industrial users, data centers operate around the clock. Their demand does not disappear overnight or during weekends, and they require highly reliable power because even short interruptions can disrupt services, damage equipment or interfere with model training. This creates a difficult profile for grid planners: large, concentrated, continuous loads that may arrive faster than supply.

PJM has acknowledged that electricity demand is growing faster than generation. The region needs more power plants, renewable-energy projects, batteries and transmission capacity, and it needs faster processes for connecting them. Yet adding supply is not simple. Natural-gas plants face permitting, pipeline and equipment constraints. Nuclear projects require long development periods and immense capital. Wind and solar projects may be built more quickly but require transmission, storage and complementary generation to maintain reliability. Large grid batteries remain expensive and cannot independently support extended periods of high demand. Transmission projects can take a decade or more to approve and construct, while data centers can move from proposal to operation far more quickly. The mismatch between digital development and physical infrastructure is becoming one of the defining constraints of the A.I. economy.

The Queue Problem

One of the sharpest criticisms of PJM concerns the time required to connect new sources of electricity to the grid. Thousands of proposed projects, including solar farms, wind developments, battery systems and conventional power plants, have sought permission to interconnect with the regional network, and many have waited years for technical studies and approval. Grid operators must determine whether a proposed facility can connect without destabilizing the system and what transmission upgrades will be necessary; the process is technically demanding, but the size of the queue has created significant delays.

As demand rises, those delays become increasingly costly. New generation that could reduce capacity prices or improve reliability remains stuck in the approval process, while data centers continue requesting access to the grid. This produces a distorted sequence in which the demand can arrive before the supply intended to serve it.

Governors and utility regulators have argued that PJM has been too slow to process new generation and storage resources, contending that the delays have contributed to higher prices and left the grid with a dangerously narrow reserve margin during periods of extreme weather. PJM has taken steps to reform its queue, but the underlying challenge remains: it must evaluate a large pipeline of projects while ensuring that new connections do not compromise reliability. The process must become faster without becoming careless, and the economic cost of failing to strike that balance is now measured in billions of dollars.

Reliability Is Becoming More Expensive

Electricity markets do not pay only for power that is consumed. They also pay for power that must be available if needed. Capacity auctions exist because the grid must survive the hottest summer afternoons, the coldest winter mornings and unexpected failures of major power plants or transmission lines, and suppliers are compensated for maintaining resources that can operate during those critical periods. When the reserve margin becomes thin, the value of dependable capacity rises.

PJM has repeatedly faced concern about how much excess power would be available during severe weather, and recent heat waves have pushed electricity consumption sharply higher across the region, leaving less room for equipment failures or forecasting errors. Data centers intensify that pressure because they add large amounts of relatively inflexible demand. A residential customer may reduce air-conditioning use during an emergency. A factory may agree to curtail production temporarily. Some data centers may be able to shift computational workloads or use backup generation, but many require continuous operation and exceptionally high reliability. The grid must therefore procure enough supply not only to serve ordinary demand but to support enormous facilities whose business models depend on uninterrupted power, and that reliability has a price. The more the system is stretched, the more consumers must pay generators to guarantee availability.

The Governance Problem

PJM occupies an unusual position in the American political system. It makes decisions with major consequences for electricity prices across multiple states, yet it is not directly controlled by the governors or utility commissions serving those states; it is regulated primarily by the Federal Energy Regulatory Commission. This regional structure exists because electricity moves across state boundaries and must be coordinated across a large interconnected network, and no individual state can manage the entire system independently. But the arrangement creates an accountability gap.

A governor may face public anger over rising utility bills while having limited authority over the auction structure or interconnection process that helped create those costs. State regulators can oversee local utilities, but they cannot easily direct the regional market. This tension has become increasingly visible. Pennsylvania Governor Josh Shapiro sued PJM in late 2024 over capacity prices, and the dispute ended in a settlement that imposed a cap on the auction price, preventing even larger costs from flowing to consumers.

The intervention showed that political pressure can alter market outcomes. It also demonstrated how unstable the current system has become. When elected leaders must sue a grid operator to limit electricity costs, the institutional model is no longer functioning as a purely technical arrangement. It has become a significant political contest over affordability, industrial development and control of regional infrastructure.

The Backlash Against Data Centers

Public opposition to data-center construction is expanding. Communities have raised concerns about electricity prices, land use, water consumption, noise, diesel backup generators, transmission lines and the limited number of permanent jobs created by highly automated facilities. The debate is especially intense because the economic benefits are uneven. A local jurisdiction may receive construction activity, property-tax revenue or infrastructure investment; a technology company may gain access to a strategically valuable site; the state may strengthen its position in the digital economy. But residents may bear higher utility costs, environmental burdens or changes to the character of their communities.

New York’s decision to impose a temporary statewide moratorium on new data-center construction reflects the growing political resistance; the pause is intended to provide time to assess energy and environmental consequences before additional projects proceed. Moratoriums may slow development, but they do not resolve the underlying demand. Artificial-intelligence companies will continue seeking access to power, and if one state becomes restrictive, developers may move projects elsewhere, shifting the pressure rather than reducing it. A durable policy must address pricing and infrastructure directly. The issue is not whether data centers should exist. It is whether they should be permitted to consume scarce grid capacity without bearing the full cost of the supply, transmission and reliability measures they require.

Data Centers May Need a Different Rate Structure

Traditional utility pricing was developed for a system dominated by households, commercial buildings and industrial facilities whose demand grew gradually. Hyperscale data centers do not fit comfortably within that structure; their size, speed and concentration make them closer to infrastructure projects than ordinary customers. Regulators may need to establish special rate classes or interconnection agreements for exceptionally large loads.

These requirements would not eliminate all shared costs. Electricity networks are interconnected, and major investments often benefit multiple users. But they would reduce the extent to which ordinary customers subsidize speculative or highly concentrated industrial demand. The principle should be clear. A company requesting the power consumption of a city should not be treated like another commercial account.

The Risk of Overbuilding

The current expansion also creates the possibility of stranded infrastructure. Utilities and grid operators are planning around forecasts of extraordinary data-center growth, forecasts that assume demand for artificial intelligence will keep rising, that companies will build the facilities they have announced, and that computing efficiency will not reduce electricity use enough to change the outlook. Each assumption may prove correct. None is guaranteed.

Technology companies frequently reserve access to multiple sites before deciding which projects to complete. Some proposals may never move beyond the planning stage; others may be delayed by financing, chip availability, local opposition or changes in model architecture. Artificial-intelligence systems may also become substantially more efficient, as new chips, cooling systems or training methods reduce the electricity required for a given amount of computing. If utilities build power plants and transmission lines based on demand that does not materialize, customers could remain responsible for the cost.

The opposite risk is equally severe. If infrastructure investment is too cautious and demand does materialize, electricity prices will remain high, reliability will weaken and economically valuable projects may be unable to connect. Grid planners must therefore make multibillion-dollar decisions under unusual uncertainty. The data-center boom may be durable. It may also contain elements of speculative overcommitment, and a responsible system should place a meaningful portion of that forecasting risk on the companies creating it.

The Economic Development Trade-Off

States are competing aggressively for data centers because the facilities signal participation in the artificial-intelligence economy. Governments offer tax abatements, accelerated permitting, infrastructure support and favorable electricity arrangements to attract them, and officials argue that these investments create construction work, expand the tax base and attract related technology companies. The benefits can be real. But the economic-development calculation must include the opportunity cost of electricity.

Power allocated to a data center cannot simultaneously support a factory, housing development, hospital or other commercial expansion unless additional supply is built. In regions with constrained generation or transmission, electricity itself becomes a scarce development resource. A data center may produce fewer permanent jobs per megawatt than almost any major industrial use. That does not make it economically undesirable; the facility may generate substantial tax revenue, strengthen digital infrastructure and support a strategically important industry. It does mean that officials should evaluate projects based on more than headline investment figures.

Without those answers, a celebrated technology investment can become a transfer of cost from a global corporation to local ratepayers.

A.I. Is Becoming an Energy Industry

Artificial intelligence is usually described as a software revolution. That description is incomplete. The industry increasingly depends on physical systems: semiconductors, data centers, cooling equipment, electric generation, transmission networks, water supplies and land. Its growth is constrained not only by the quality of algorithms but by the ability to deliver power.

This changes the strategic landscape. Technology companies are signing long-term electricity agreements, investing in nuclear power, supporting natural-gas generation and exploring proprietary energy projects. Utilities are revising demand forecasts. States are reconsidering how large computing facilities should be regulated. The distinction between the technology sector and the energy sector is beginning to blur. The most advanced A.I. system in the world has little economic value if it cannot be powered at scale, and electricity is becoming one of the critical inputs of machine intelligence, alongside data, chips and technical talent. The cost of artificial intelligence is no longer confined to venture investors, technology companies or corporate customers. It is appearing on ordinary utility bills.

The New Standard for Data-Center Development

The growth of artificial intelligence will require substantial expansion of the electrical system, and that investment can strengthen the grid, create new generation and modernize infrastructure the country already needs. But the development model must be economically credible. Large power users should provide firm commitments before utilities construct dedicated infrastructure. They should pay for the costs uniquely attributable to their projects. Grid operators should accelerate viable generation and storage projects. States should coordinate rather than compete through subsidies that conceal the true cost of development.

The public should also receive greater transparency. Communities deserve to know how much electricity a proposed facility will require, how its connection will affect rates, what infrastructure must be built and which party will bear the financial risk. Without those reforms, the expansion of artificial intelligence could produce a regressive outcome, in which some of the richest companies in history gain access to increasingly scarce electricity while millions of households and smaller businesses pay more to preserve the reliability of the system serving them.

The $6.3 billion increase projected from the latest PJM auction should be understood as more than a temporary market fluctuation. It is an early invoice for the physical economy behind artificial intelligence. The question now is whether future invoices will continue to be distributed across the public, or directed more precisely toward the companies generating the demand.

The SnowRock read

We are an AI firm publishing a caution about the cost of AI, which deserves a word of explanation. The version of this technology that shows up in these headlines, trillion-dollar campuses and grid-bending demand, is not the version most of our clients need. A mid-market company does not require a hyperscale data center or a frontier training run. It requires a few well-chosen models doing specific, measured work, and the energy footprint of that is closer to a rounding error than a regional crisis.

That distinction is quietly becoming a competitive one. As power, and therefore compute, grows scarcer and more expensive, the discipline we have always argued for pays off twice: right-size the model, run it only where it changes the answer, and you are largely insulated from the cost curve this analysis describes. The companies that treated AI as a reason to consume without measuring will feel the invoice. The ones that treated it as an engineering decision, with a bar and a budget, will barely notice. Efficiency was always the honest strategy. The power bill is simply making it the obvious one.