Data Centers: The Public Cost of the AI Build-Out

Data centers cool their servers with water. A single 100-megawatt facility can consume up to 2 million liters — roughly 528,000 gallons — per day. The hyperscalers now building AI infrastructure at unprecedented scale are doing it in the wrong places. Since 2022, more than two-thirds of new data centers have been sited in water-stressed regions, including Texas, Arizona, Saudi Arabia, and India.

In the Phoenix metropolitan area — Maricopa County — data centers from Microsoft, Google, Meta, and others have drawn from an aquifer system operating under federally-acknowledged drought conditions. The Arizona Department of Water Resources has designated portions of the region as groundwater-stressed. The Arizona Republic documented data center water draws in Maricopa County under contracts that were negotiated without public input. The county water authority executed those contracts.

The global water footprint of data centers is estimated at 560 billion liters annually — with projections that it will double by 2030. U.S. data centers alone consume roughly 1.7 billion liters per day, 57 percent of it drawn from potable supplies. Microsoft, Google, and Meta have pledged to become "water positive" by 2030, meaning they will replenish more water than they use in aggregate. Aggregate water credits do not restore what's been removed from a specific aquifer in a specific drought-declared county.

In 2024, Elon Musk's xAI opened Colossus — a supercomputer campus in a former Electrolux factory in South Memphis, one of the most economically disadvantaged and predominantly Black neighborhoods in Tennessee. The facility covers 785,000 square feet. To power it while waiting for grid connection, xAI deployed 35 portable methane gas turbines with a combined capacity of 422 megawatts.

The Southern Environmental Law Center identified the Colossus facility as likely the largest industrial emitter of nitrogen oxides in Memphis, estimating between 1,200 and 2,000 tons of NOx annually — smog-forming pollution linked to respiratory disease. The turbines also emitted formaldehyde, a known carcinogen. Residents reported chronic respiratory complaints. Shelby County Health Department records documented the emissions.

The permit question: Shelby County regulations at the time required air quality permits only after equipment had been at a location for more than 364 days. xAI classified its turbines as "portable" — a classification that exempted them from permitting requirements regardless of output. The facility operated under this classification for months without a Clean Air Act permit. A January 2026 EPA ruling changed the compliance framework, requiring permits for temporary large methane turbine operations. The emissions during the unpermitted period are not recoverable.

Residents were not notified before construction began. No formal community review occurred. The neighborhood — already living with the legacy of industrial siting decisions that historically concentrated pollution in Black communities — had no opportunity to weigh in before the turbines went online.

Loudoun County, Virginia is home to "Data Center Alley" — more than 60 massive data center campuses covering 25 million square feet, the world's largest concentration of compute infrastructure by density. The electricity demand from these facilities has placed significant strain on the Dominion Energy transmission grid.

A December 2024 audit by the Virginia State Corporation Commission found that the grid investment required to service data center demand could increase residential electricity bills by $14 to $37 per month by 2040. The mechanism is structural: utilities must build and maintain transmission capacity for peak demand. When industrial customers — data centers — drive that peak, the fixed-cost recovery gets allocated across the full ratepayer base, including residential customers who had no say in where the data centers were built.

U.S. data centers consumed 4.4 percent of national electricity in 2023. The IEA projects that figure will reach between 6.7 and 12 percent by 2028. Nationally, one energy modeling study projects an 8 percent increase in U.S. electricity prices by 2030 attributable to AI compute demand. That price increase would fall on every residential ratepayer in the country. The hyperscalers whose compute demand drives it have negotiated wholesale industrial rates that do not reflect the full cost.

The AI compute build-out is reversing utility decarbonization timelines. Power plants previously slated for retirement — including natural gas and, in some cases, coal facilities — are being kept online to service data center load that would otherwise strain grids. New gas plants are being permitted in states with active hyperscaler construction. State PUC dockets in Virginia, Texas, and Georgia show utilities citing data center demand as justification for generation capacity that contradicts their own published decarbonization plans.

The rate-design mechanism: data centers negotiate below-cost wholesale industrial rates, often through state economic development agreements that treat favorable electricity pricing as an incentive. The fixed costs of generation, transmission, and distribution that those rates do not recover are then allocated to residential and small-commercial ratepayers. State public utility commissions — the regulatory bodies nominally protecting ratepayer interests — have approved these rate structures after processes in which hyperscaler lobbying was the dominant voice.

Amazon, Microsoft, Meta, and Google are among the largest political spenders in state legislatures where data center infrastructure decisions are made. OpenSecrets data on technology sector lobbying, combined with state-level lobbying disclosures in Virginia, Texas, Tennessee, and Georgia, show consistent presence in utility regulation proceedings where rate design and generation permitting are at stake.

Data centers are capital-intensive and labor-light. A facility consuming hundreds of megawatts and housing billions of dollars in equipment may employ fewer than fifty permanent workers — security, maintenance, and a small technical staff. The construction workforce is temporary. The economic footprint in the local tax base is minimal once the property tax abatement is in effect.

Good Jobs First's Subsidy Tracker database documents data center deals across states where Microsoft, Meta, Google, and Amazon have received property tax abatements — sometimes full exemptions for ten years or longer — in exchange for "job creation" commitments. State economic development filings from Wisconsin, Iowa, and Tennessee record per-job subsidy costs that can reach hundreds of thousands of dollars for facilities that deliver a fraction of the promised employment. Ratepayer-counsel reports in several states have documented the gap between jobs promised and jobs delivered.

In some deals, contract terms are concealed from the public under non-disclosure agreements between the company and the local government. Communities learn what is being built — and what was given away — only after the agreement is signed. FOIA requests have extracted contract excerpts in several jurisdictions; local reporting has documented the pattern. The abatements reduce the local tax base available for schools, roads, and emergency services in the same communities that bear the infrastructure burden of the facility.

Elon Musk, Sam Altman, Mark Zuckerberg, Jeff Bezos, Larry Page, Sergey Brin, Larry Ellison, and their respective companies are building AI compute infrastructure at a scale that requires public water from drought-stressed aquifers, public grid capacity whose costs fall on residential ratepayers, public tax forgiveness over decade-long abatement windows, and regulatory accommodation that bypasses normal permitting and environmental review. None of this is "building" in the sense that earns a fortune. It is extracting from public infrastructure that they did not build and are not paying for.

The "AI race" frame — national competitiveness, falling behind China, the urgency of speed — functions to convert this extraction into a public interest story. The frame is effective because it is not entirely false: there are genuine competitive dynamics in AI development. It is effective because it uses true premises to reach a conclusion the true premises do not support: that the subsidy, the water draw, the pollution, and the rate increase are necessary costs for the country rather than necessary profits for specific owners of capital.

The people paying the cost of Data Center Alley are Loudoun County residential ratepayers. The people paying the cost of Colossus are the residents of South Memphis who breathed NOx for months while xAI classified its turbines as portable. The people paying the cost of Maricopa County data center water draws are the downstream users of an aquifer system that does not refill on a quarterly earnings cycle. The owners of xAI, Microsoft, Meta, Google, and Amazon are not paying those costs. They are compounding capital. That is the transaction. The race framing is the disguise.