The AI Boom Just Hit a Wall Made of Electricity Bills
• 3 min read • Ahmed Daoudi
For years, a new data center was a win for a town: construction jobs, tax revenue, a ribbon-cutting photo. In 2026 the mood flipped. From Virginia to Arizona, residents are showing up at hearings to stop them, and politicians from both parties are listening.
How big the backlash has become
- Politico reported in October that 16 state legislatures have considered outright bans on data center development, where a few years ago states competed with tax breaks.
- Research cited by Quartz estimates US wholesale electricity prices could be 6% to 29% higher by the end of the decade depending on data center growth, with generation costs in Virginia rising as much as 57%.
- The Washington Post described a shift inside the industry's own political allies, with new voices floating ideas like pausing data center construction and requiring operators to shield residents from higher bills.
It is about electricity, not AI
MIT Technology Review argued on November 9 that the backlash is less a culture war over technology than a fight over power. Drawing electricity from the grid raises prices for everyone on it. Building on-site generation, often gas turbines, brings noise, emissions and local opposition of a different kind. Neither problem goes away with better public relations.
Water is the other pressure point. Cooling large facilities in dry regions puts them in direct competition with farms and households.
Why developers should care
It is easy to treat this as someone else's problem. It is not. If power gets more expensive and new capacity gets harder to build, AI inference gets more expensive, and that cost flows into API prices and cloud bills. The era of treating model calls as nearly free is ending.
Some practical habits help both your budget and the grid:
- Use the smallest model that does the job. Many classification and extraction tasks do not need a frontier model.
- Cache aggressively. Identical prompts should not hit the model twice. Most providers now offer prompt caching for repeated context.
- Batch non-urgent work and run it when it is cheapest. Batch APIs are typically discounted for exactly this reason.
- Measure tokens per feature the way you measure database queries per request.
My take
The AI industry assumed the physical world would keep up with its roadmap. The physical world has opinions. The companies that do well from here will be the ones that treat efficiency as a feature, because the cheapest kilowatt-hour is the one you never needed.