Who Pays the Bill for AI Infrastructure?
Data-center growth is creating new costs. The invoice that's coming due reaches technology companies, ratepayers, taxpayers, and host communities.
The artificial intelligence boom is generating an infrastructure bill. Its line items include electricity, new power plants and transmission, cooling water, tax incentives, backup generation, pollution controls, and health monitoring.
Technology companies pay part of that bill. Ratepayers, taxpayers, and host communities can be assigned the rest.
A year ago, I published The New Math of AI Infrastructure to make AI’s environmental cost tangible. Counting one prompt helped people see that a text request had measurable energy, carbon, and water impacts.
The 2026 update follows the invoice through the physical system. It asks four questions: What appears on the invoice? Who sets the terms? Who receives the benefits? Who pays?
These questions already touch nonprofit and foundation work on energy affordability, public health, climate, technology grants, workforce development, and environmental justice. They are entering boardrooms, grantmaking strategies, public hearings, and community campaigns.
The invoice reaches beyond the data center
United States data centers used an estimated 4.7 percent of the country’s electricity in 2024. Lawrence Berkeley National Laboratory projects a 2030 range of 9.5 to 15.3 percent.
That range covers all data-center workloads, including AI, and it remains a forecast. Its uncertainty is part of the governance problem because utilities and governments are making long-term commitments now.
To reduce the risk of shortages, regional grids pay power suppliers to be ready when electricity demand peaks. In the PJM region, which covers 13 states and Washington, DC, the price of that readiness is set through auctions.
An independent watchdog for the PJM market estimated that data-center demand accounted for $29.4 billion, or 46 percent, of the charges from its four most recent auctions. The $29.4 billion is a wholesale market figure. It can affect customer bills, but it is not the amount households alone paid.
Tax incentives create another line item: public revenue forgone. Virginia data-center operators reported $1.94 billion in state, local, and regional sales-tax savings in fiscal 2025. The figure covers reported sales-tax savings.
Total incentives and net fiscal impact require separate accounting. Public-return estimates change when analysts use different assumptions about what investment would have occurred without the subsidy.
Water shows how the same invoice can look small statewide and large locally. In 2023, data centers accounted for less than 0.5 percent of statewide water withdrawals in Virginia. At the six utilities reviewed by state analysts, their local share ranged from 0.2 to 21 percent after reclaimed water was excluded. Both figures are accurate. They answer different questions.
Any claim about cost should name the geography, the denominator, and the payer.
The invoice lands on an unequal map
New infrastructure costs arrive on top of older environmental burdens.
A 2026 working-paper tract analysis found no consistent national racial pattern in data-center siting. California and Washington inventories found regional concentration in communities with inherited environmental burdens or larger populations of color.
The studies counted different things and used different indicators. The two state inventories weren’t weighted by power capacity. Colocation doesn’t prove causation.
A national average can coexist with serious regional and neighborhood burdens. The difference between the national and state findings is part of the equity evidence because each method reveals a different scale of impact.
The updated article uses environmental racism as an inquiry into whether racialized policy, inherited pollution, and unequal decision-making power influence how new costs are assigned and whose objections carry weight. That inquiry is separate from a legal finding of intentional discrimination.
The relevant map includes the source of the electricity, existing pollution and health burdens, expected jobs and tax revenue, public investment, and who had influence while the terms could still change.
Memphis shows why those questions belong in the analysis. KeShaun Pearson and Memphis Community Against Pollution organized opposition to xAI’s turbine permit and partnered with researchers to install local air monitors. The dispute is site-specific, and its legal questions remain contested. The public permit fight followed the turbines’ arrival and operation.
Decide who pays before the bill comes due
AI already supports flood forecasting and wildfire response. Those benefits belong in the accounting too.
Boards, funders, and nonprofit leaders can demand fairer accounting before a permit hearing or incentive vote:
1. Audit every line item. Fund independent engineering, legal review, health monitoring, financial analysis, and resident organizing. Examine power, water, emissions, incentives, and cumulative burdens while sites, designs, and contract terms can still change.
2. Name every payer. Publish which costs go to the operator, ratepayers, taxpayers, and residents. Release the assumptions before officials approve incentives or contracts.
3. Put the price and protections into enforceable agreements. Assign infrastructure costs to the large customer. Include minimum payments, clawbacks, public reporting, and community benefits that go beyond basic permit compliance.
The full 2026 update follows the invoice through the infrastructure system, with 42 source notes, a glossary, case records, and specific policy options.
Before the next grant decision, utility case, or permit hearing, ask four questions:
What is on the invoice? Who receives the benefits? Who is expected to pay? Who had power to negotiate the terms?
Read the full analysis