Data Centers Are Now Using So Much Power They're Changing How America Is Built

AI data centers are driving unprecedented electricity demand that's raising household bills and reshaping American infrastructure priorities. Here is the specific scale and who pays.

PBS News' April 2026 reporting captured a specific shift in American energy infrastructure: "Skyrocketing power demand from massive data centers and rising household electric bills are injecting a wave of attention into who" — the incomplete sentence whose continuation describes the specific political, economic, and technical debate about who bears the cost of powering artificial intelligence's physical infrastructure.

Data center power consumption in the United States has grown from approximately 200 terawatt-hours annually in 2018 to an estimated 500+ terawatt-hours in 2026 — more than doubling in eight years, with the acceleration driven specifically by large language model training and inference workloads that the generative AI era has introduced. A single large training run for a frontier AI model consumes approximately as much electricity as 100 US households use in a year. OpenAI, Anthropic, Google, Microsoft, and Amazon are all simultaneously building and expanding the specific data center capacity whose aggregate electricity demand is reshaping American grid planning.

The specific geographic concentration of data centers — in northern Virginia (the largest data center market in the world), Iowa, Texas, and increasingly in states with available cheap power and land — creates particular local effects. Power companies in Northern Virginia have warned that the specific data center buildout has created electricity demand growth that their infrastructure cannot meet without substantial new generation investment. Average residential electricity bills in the specific utility zones serving major data center concentrations have risen as the cost of grid expansion to serve commercial customers is partially allocated to residential ratepayers.

The specific mechanism linking AI development to electricity demand involves the particular energy intensity of the specific hardware — Nvidia H100 GPU clusters, AMD Instinct accelerators, Google TPUs — whose operation requires continuous power at densities that no previous computing application produced at comparable scale. A single Nvidia H100 server draws approximately 700 watts. A large AI training cluster uses tens of thousands of these servers simultaneously. The cooling infrastructure required to prevent these systems from melting themselves requires additional power — typically 30-40% of the server power consumption — for the specific chillers, cooling towers, and air handling systems that maintain operational temperature.

The Iran war's specific energy cost impact adds a particular dimension to this story. Natural gas — whose price elevation from the Hormuz blockade increases electricity generation costs — is the primary fuel for the specific peaker plants and baseload combined-cycle generators whose operation produces the electricity that data centers consume. Higher natural gas prices translate directly into higher electricity costs for both commercial data center operators and residential consumers sharing the same grid.

Iron and steel prices — elevated by the same tariff dynamics that the broader economy is experiencing — affect the specific cost of the transmission infrastructure, transformers, and switchgear required to connect new data centers to high-voltage transmission networks. The specific construction timeline for this infrastructure — 3-7 years from planning to commissioning for large transmission projects — means that the electricity demand AI creates today may not have adequate grid infrastructure until the decade's end.

The specific policy debate is simple to state and complex to resolve: should the AI companies whose specific applications drive data center power demand pay the full cost of the grid expansion required to serve them, or should those costs be socialized across all ratepayers including households that don't use AI services but receive electricity from the same grid?

The current regulatory framework in most states treats large commercial customers and residential customers within the same rate-setting process, meaning grid upgrade costs are typically spread across all customers according to regulatory formulas that don't fully allocate new infrastructure costs to the specific new customers requiring it. The specific utility commissions in Virginia, Texas, and other major data center states are currently wrestling with exactly this question in specific rate cases and grid planning proceedings.

For the specific $500 billion Stargate initiative — whose data center buildout represents the largest single AI infrastructure investment in history — the specific electricity source question has geopolitical dimensions beyond domestic grid planning: Iran has threatened to strike the Stargate facility in Abu Dhabi, adding war risk to the specific energy and infrastructure challenges that AI's power appetite was already creating.