Data centers now drive over half of American electricity demand growth

The interconnection queue at the Electric Reliability Council of Texas grew so heavy this summer that Governor Greg Abbott did something no Texas politician wants to do: he stopped taking orders. On August 3, Abbott directed the Public Utility Commission of Texas and ERCOT to audit every new data center seeking grid access, effectively freezing approvals in the second-largest data center market in the country (Texas Tribune, Aug 3). A governor who called his state the epicenter of artificial intelligence paused its growth because the wires could not be promised honestly.

That pause sits on top of the number behind this headline. Grid Strategies' third annual load-growth study finds data centers driving roughly 90 gigawatts of the 166 gigawatts of expected new American peak demand, about 55 percent of it (Grid Strategies load growth report, 2026; Las Vegas Sun, Aug 22). For two decades American electricity use was flat, because efficiency savings canceled out population and economic growth. The data centers broke that pattern, and the whole machinery of American power, which had spent twenty years planning for no growth, is being rebuilt around them.

Name the actors. The hyperscalers, meaning Microsoft, Amazon, Alphabet and Meta, want gigawatts delivered fast and priced low, because compute capacity is now their capital-spending race. The utilities, like Georgia Power and the Texas municipals, want those contracts because large loads spread their fixed costs and justify new plants, which grow their regulated earnings. The state regulators sit between them, and their job, stated plainly by the Maryland Office of People's Counsel's David Lapp at a county conference this month, is to make sure infrastructure built for data centers does not land on residential bills (Conduit Street, Maryland Association of Counties, Aug 19).

The hyperscalers hold an option on cheap power and can walk away if AI economics disappoint; regulated utilities cannot walk away from a turbine once poured.

The trigger this month was Texas. The slow pressure underneath is older: a buildout of power plants, transmission lines and gas turbines sized for a demand curve that only exists if the AI spending keeps flowing. The EIA now projects American power consumption setting consecutive records in 2026 and 2027, driven by data centers and electrification (Reuters, Aug 11), but its August outlook also cut the Texas demand forecast sharply because of the pause there, expecting roughly 6 percent load growth in 2027 instead of the 14 percent projected before (EIA Short-Term Energy Outlook, Aug 2026). The forecasts move with the announcements. The turbines are ordered years ahead of either.

The history that fits is the Washington Public Power Supply System. In the late 1970s five regional utilities signed take-or-pay contracts for five nuclear plants, meaning they owed the full cost whether or not they ever needed or used the power, on the faith that Northwest electricity demand would double by 1990. It did not. In 1983 WPPSS defaulted on $2.25 billion of municipal bonds, then the largest default in American history, and the utilities turned to the courts to collect the shortfall from their own customers. The lesson survived: whoever signs first for capacity that may not be needed decides who eats the loss.

The counterexample argues the other way. This buildout has real customers with real money, not just forecasts: the hyperscalers are funding their own generation, signing power purchase agreements, and in some regions paying special tariffs designed to insulate households. A study by the consulting firm Energy + Environmental Economics published in May 2026 argued that blaming data centers for rising bills does not survive regulatory arithmetic, because large flat loads can dilute fixed costs rather than concentrate them (Energy + Environmental Economics, May 2026). Whether that holds depends entirely on the tariff design, which is exactly what regulators are fighting over.

And they are fighting. The Georgia Public Service Commission voted unanimously in July to investigate whether Georgia Power's largest industrial customers, including data centers, are shifting fuel costs onto residential ratepayers (Macon Telegraph, Jul 10). That same commission is reviewing a contract between Georgia Power and OpenAI for a proposed twenty-billion-dollar data center in Effingham County, and the utility took an extension into mid-August after staff raised objections (The Current, Aug 13). In Virginia's Henrico County, electricity rates rose nearly 25 percent starting July 1, adding about five million dollars a year across county government and school budgets, forcing schools to dim lights while thirty-seven data centers operate nearby (Inc., 2026).

Walk the consequences through. If regulators write strong large-load tariffs, the data centers pay for their own plants and transmission, utility shareholders absorb less risk, and household bills stabilize; the AI buildout slows at the margin but survives. If they do not, utilities finance plants against contracts with counterparties whose own revenue depends on AI demand holding up, and the risk migrates quietly to the people least able to carry it. Politicians have already noticed: a January poll by the Pew Research Center found nearly forty percent of adults blame data centers for higher utility bills (Pew Research Center, Jan 2026). Electric bills are becoming a midterm issue from Pennsylvania to Michigan (Bloomberg, Feb 2026).

What confirms this read: the Georgia PSC's cost-allocation investigation producing findings that large loads do or do not cover incremental costs, and more states following Texas in auditing interconnection requests before approval. What breaks it: wholesale power prices staying flat through the 2027 summer peak while residential rates also flatten, which would mean the buildout genuinely pays for itself and the backlash dissolves.

The judgment this piece earns is about who holds the option value. The hyperscalers hold an option on cheap power and can walk away if AI economics disappoint. Regulated utilities cannot walk away from a turbine once poured. Unless every state writes the Georgia question, who pays for capacity built on speculative demand, into law before the concrete sets, the answer will arrive the old way: on the bill.