A single large AI data center consumes roughly as much electricity as a medium-sized city. There are hundreds of them. There will be thousands. And almost none of the people who live in the communities around them, who pay the utility bills that help finance their power supply, or who fund the public infrastructure that serves them, had any say in the decision to build them.
This is not a technology story. It is an economics story about who pays for things and who decides they don't have to.
Key Points
- AI data centers will consume nearly 1,000 TWh of electricity globally by 2030, roughly 3% of total world power demand.
- Data center emissions are projected to double, potentially accounting for 1% of all global greenhouse gas output.
- US states have granted billions in tax subsidies to data center operators, costs that fall on ordinary taxpayers.
- Electricity bills for residential consumers are already rising in regions with heavy data center concentration, including Virginia and Texas.
- Most people bearing these costs did not choose to use AI and had no input into the infrastructure decisions driving the costs.
The Scale Nobody Is Advertising
AI-specific energy consumption is currently measured in tens of terawatt-hours per year. By 2030, according to research by management consultancy Arthur D. Little, it is projected to approach one thousand — a tenfold increase in under a decade, representing roughly three percent of total global power demand. A single sector, largely concentrated in a handful of companies, drawing one in thirty watts of electricity generated anywhere on Earth.
Data center emissions are expected to double over the same period, potentially accounting for a full percent of all global greenhouse gas output. In Europe alone, the energy demand from AI infrastructure is projected to hit two hundred terawatt-hours annually. The infrastructure required to meet this demand does not exist yet. Building it requires new power generation, new transmission lines, new cooling systems drawing millions of gallons of water daily. In some regions, the queue for a new grid connection is now seven years long.
None of this complexity appears on the pricing of a ChatGPT subscription or a Gemini API call. As Arthur D. Little's Albert Meige put it: "AI feels cheap today because its real economic and environmental costs are essentially hidden."
How an Electricity Bill Absorbs a Data Center
The mechanism is not complicated. When a hyperscale data center connects to a regional grid, the utility company must upgrade substations, transmission lines, and backup generation capacity to accommodate the new load. These infrastructure upgrades are financed through utility rate structures. The cost is then distributed across the entire customer base: residential households, small businesses, schools, hospitals, through a line item on their monthly bills.
The data center negotiates its own rate, often at a substantial discount secured through long-term contracts and political leverage. The infrastructure built to serve it is co-financed by everyone else. It is not the data center that pays for the grid upgrade that made it possible. It is the retired teacher two towns over who does not own a smartphone.
This is not a theoretical scenario. Across major US grid regions, wholesale capacity prices have risen sharply in recent years, with data centers cited as the primary driver. Residents in affected areas are already absorbing double-digit monthly bill increases, with further increases projected through the coming years. The pattern is consistent: infrastructure costs generated by commercial AI buildout are distributed across utility customers who had no seat at the negotiating table.
Ireland Is a Warning
In Ireland, data centers now consume more than twenty percent of the country's total national electricity. The concentration is so severe that Dublin has imposed connection moratoriums: new data centers cannot access the grid in certain areas because the infrastructure physically cannot absorb more demand without jeopardizing supply to existing customers.
This is what the end of the runway looks like. Ireland has a functioning grid, advanced infrastructure, and significant political will to attract foreign investment. It has still reached a point where the demands of a single sector, which employs a small fraction of the population and exports most of its economic value to parent companies abroad, have begun to structurally compromise the energy security of ordinary residents and businesses.
The Irish situation is not unique. It is early. Other regions are on the same trajectory with different lead times. What makes Ireland instructive is that the outcome was foreseeable, was foreseen by regulators, and proceeded anyway because the short-term investment incentives overwhelmed the longer-term capacity planning.
Water: The Bill You Don't Get
Energy is measurable and priced. Water is harder to track, which may be why it receives less attention. A large AI data center consumes on average five million gallons of water per day for cooling, equivalent to the daily use of a town of tens of thousands of people, concentrated in a single facility that typically employs a few hundred workers.
In Texas, where data center development has accelerated sharply, the projected water demand from the sector is expected to reach hundreds of billions of gallons annually by 2030, drawn not from a single source but from aquifers, rivers, and municipal systems distributed across a state already experiencing chronic drought stress.
In Chile, which has attracted significant data center investment, fifteen consecutive years of drought have depleted aquifer reserves. The water used to cool servers in Santiago is water that will not be available to agricultural communities downstream. The cost does not appear on anyone's invoice. It appears in harvest yields and groundwater depth and the price of food.
The Subsidy Architecture
Alongside the utility cost distributions, a parallel financial structure operates through state tax policy. Dozens of US states currently offer data center tax exemptions running into the billions of dollars annually. Virginia, the country's largest data center market, has extended exemptions that amount to nearly two billion dollars per year. Many of these programs are uncapped in both duration and dollar amount, running in some cases through 2050.
Only a minority of states with such programs publicly disclose which companies receive the benefits. The majority distribute foregone public revenue to unnamed corporate recipients, making accountability structurally impossible.
The argument for these subsidies is the standard economic development argument: investment, jobs, tax base. The counterargument is increasingly concrete: in Virginia, data centers pay no sales tax on electricity or equipment, while the state continues to fund roads, schools, and emergency services that serve the communities surrounding the facilities. In March 2026, the Virginia Senate moved to remove a significant portion of this tax break following sustained pressure from both progressive and conservative legislators — an unusual coalition driven by a shared recognition that the existing arrangement is not a subsidy for economic development but a transfer of public resources to private infrastructure at public cost.
The Louisiana case with Meta's Hyperion facility makes the same point in more concentrated form. The largest AI data center ever built, located in one of the poorest counties in the United States, financed through utility rate structures that will distribute infrastructure costs across the regional customer base for decades. The contract terms between Meta and Entergy Louisiana remain largely undisclosed. The communities within a fifty-mile radius have no mechanism to renegotiate what they did not negotiate in the first place.
The Transparency Gap
Fewer than three percent of new AI models disclosed any environmental data in 2024. This is not an oversight. Disclosure of energy consumption, water use, and emissions is not currently required by any major jurisdiction as a condition of deploying AI systems commercially. Companies that voluntarily publish sustainability reports do so on their own terms, using metrics of their choosing, audited by parties they select.
The practical consequence is that there is no reliable public accounting of the aggregate environmental cost of AI deployment at the system level. The total bill, in electricity, water, land, grid infrastructure, and foregone public revenue, is not something any institution currently measures. It is being paid, incrementally, across utility bills and tax shortfalls and depleted aquifers, by people who have no visibility into the accounting.
The Grid Risk Nobody Is Discussing Publicly
The stability implications extend beyond cost. Multiple analyses project a significant shortfall in grid capacity by the end of the decade: data center demand is expected to substantially outpace the supply of new generation infrastructure. In July 2024, a near-blackout in Northern Virginia was triggered when dozens of data centers simultaneously disconnected from the grid during a fault event, causing a surge that nearly cascaded into a regional outage affecting millions of people. Grid engineers described the incident as a warning. It has not produced visible regulatory response.
As with the hardware shortage driven by AI chip demand, the grid risk is a collective problem created by individually rational decisions made without coordination or accountability. Each company builds the data center that maximizes its competitive position. The cumulative load falls on infrastructure that was not designed for it, maintained by utilities that are not resourced to upgrade it fast enough, in a regulatory environment that has not caught up with the pace of buildout.
The political response is beginning to form. Roughly half of Americans expect data center costs to become a 2026 midterm campaign issue. Fourteen states have active moratorium movements. In an unusual alignment, both Bernie Sanders and Ron DeSantis have publicly criticized the data center expansion model on grounds of consumer cost and public burden, a bipartisan signal that the subsidy and cost-distribution architecture is politically unsustainable in its current form.
What a Different Policy Would Look Like
The interventions that would change this trajectory are not technically complex. They are politically complex, because the industries benefiting from the current structure have substantial resources to defend it.
Mandatory environmental disclosure for AI deployments above a certain compute threshold would at minimum make the aggregate cost visible. Cost-allocation mechanisms that shield low-income ratepayers from infrastructure rate increases driven by commercial data center demand would limit the regressive distributional effect. Binding water use commitments as conditions of permitting in drought-stressed regions would prevent the agricultural and municipal externalization of cooling costs. Capped and time-limited tax incentives, with public disclosure of recipients, would restore accountability to subsidy programs that currently operate without it.
None of these are novel regulatory concepts. They are standard tools of industrial siting policy, applied in many jurisdictions to many sectors. The gap is not conceptual. It is the absence of political will to apply existing frameworks to a sector that has, so far, been treated as exempt from ordinary accountability on the theory that its economic potential justifies extraordinary accommodation.
That theory is being tested. The people paying the electricity bills in Baltimore, the farmers drawing from depleted aquifers in Texas, the residents of Richland Parish in Louisiana. They are not testing it voluntarily. They are living the result of decisions made without them, by institutions that will not absorb the consequences. That is what externalisation means in practice. It means someone else pays. And so far, it is clear who that someone else is — in the labor market, the 180 workers Crypto.com eliminated to fund its AI infrastructure pivot did not vote on that capital allocation either.