How AI data centers are reshaping electricity prices and energy politics across the US

Summary

Matthew Zeitlin, a correspondent at Heatmap News covering climate and energy policy, reports that AI data centers are becoming the dominant driver of electricity demand growth in the US and reshaping both energy prices and regional politics.

Zeitlin dates the shift to around 2023. He was initially reporting on electrification from electric vehicles and heat pumps when GPT-3's launch refocused the story. Data centers, not cars or heating, became the dominant load growth.

Electricity pressure by region

Price spikes are not uniform. California and the Northeast have seen steep increases but have little data center development, partly because prices there are already too high. The real pressure appears in the Mid-Atlantic and Midwest, where data centers cluster in the PJM interconnection covering Virginia, Ohio, and Indiana. Utilities must procure capacity in advance. The auction mechanism has malfunctioned under sudden data center demand, generating billions in revenue for existing generators and raising residential rates. New Jersey residential prices have risen roughly 20% as a result.

Infrastructure costs and pricing

When hyperscalers say they will "pay for their own electricity," they don't mean just the electrons. A 1-gigawatt data center consumes as much power as 800,000 homes. That scale requires system-wide upgrades including new transmission equipment and grid reinforcements. Historically, those costs spread across all users. The pricing model traces back to Samuel Insull's utility framework in the 1920s. In theory, new large demand should lower prices over time. In practice, upfront infrastructure costs are immediate and steep. Microsoft and Anthropic have committed to absorbing system costs in their own rate structures rather than passing them to residential customers. Press releases cannot undo formal utility rate cases, though.

Water versus electricity

Water usage from data centers has dominated public concern online, but systemic impact is negligible. Individual data centers may use 100 million gallons annually, but household water consumption far exceeds it. Golf courses use more. Hyperscalers lean heavily into "water positive" messaging in part because it's easier PR than confronting electricity costs. The material constraint sits with electricity. Ten to fifteen percent of a $50 billion data center's lifetime cost is electricity, plus hundreds of millions or possibly billions in extra grid infrastructure costs those companies volunteer to absorb to secure state approval.

Caterpillar and Williams as unlikely winners

Two companies have emerged as surprise beneficiaries. Caterpillar operates Solar, a business that builds small gas turbines in the 20 to 50 megawatt range. These units are normally used for remote oil and gas operations but have become essential for data centers desperate for fast power. Elon Musk's xAI recognized that grid interconnection studies take years and stacked small, inefficient turbines to avoid that delay. Caterpillar supplies units to Meta's Ohio facility. Williams, a pipeline company, has similarly pivoted to assembling turbines and reciprocating engines to power data centers, a business now performing "incredibly well."

These solutions are less efficient and dirtier than large-scale turbine systems from GE Vernova or Siemens, but desperation for capacity trumps environmental standards. xAI navigated EPA rules by operating temporary turbines only 364 days per year, then moved to Mississippi before certification was finalized.

Energy mix and hyperscaler hedging

Hyperscalers maintain public 100% renewable commitments while simultaneously embracing natural gas to plug immediate gaps. Google acquired renewable energy developer Pattern for $4.5 billion at year-end 2024 and bought solar projects in areas where Meta is simultaneously deploying gas turbines. All four major hyperscalers now have nuclear fusion or fission projects underway. Google invested in Commonwealth Fusion Systems and holds stakes in enhanced geothermal. This represents a major strategic shift. As their power needs have grown, they've become willing to fund frontier energy technology they would have avoided five years ago.

Political opposition without left-right division

Data center opposition defies typical left-right divides. Zeitlin's reporting at Heatmap Pro found that people distrustful of institutions broadly, whether Republican or Democrat, resist local data center projects. The same constituency that opposes renewable energy projects now opposes data centers using identical tactics: objecting to rezoning agricultural land for industrial use. At the federal level, the Biden administration is heavily pro-data-center. Sam Altman has announced multibillion-dollar investments at the White House. Yet most data centers land in Republican exurban areas where space exists, triggering intense local resistance in Indiana, Kentucky, and Virginia.

Governors and unions favor data centers because tax revenue and union construction jobs are tangible. The real battle happens in community boards and zoning meetings, where hyperscalers must convince residents of local benefit. The structural problem is that data centers generate no lasting local employment after construction. A data center offers no local product or service. Compute is exported. Residents cannot credibly claim reduced-latency AI inference as a neighborhood amenity.

The power plant bottleneck

Zeitlin argues that energy growth is the next hard constraint on AI scaling after chip supply. US electricity production has grown 0.5 to 2% annually, with occasional 5% years. Reaching 10% is "tough." Grid connection studies last years. Permitting solar takes longer than it should. Gas turbine opposition will dwarf wind and solar opposition. Even data center construction, just a warehouse with computers, faces years of local resistance. Building power plants at data center scale faces exponentially harder permitting. Litigation risk is real. Litigation-powered NEPA delays could cripple infrastructure timelines.

Davos and concrete tradeoffs

Zeitlin observes that Davos has shifted from theater to reality. The conference no longer tolerates vague sustainability commitments. Hyperscalers now discuss concrete tradeoffs and costs. For a reporter covering climate policy, that's welcome. People are dealing in facts rather than symbols. The irony is sharp. Climate politics remain deeply partisan. The Northeast enforces strict decarbonization standards while Florida, the most climate-vulnerable US state, does little. Elon Musk, history's largest solar investor and second or third largest Republican donor, demonstrates the paradox. Republican efforts to dismantle EV tax credits don't change his behavior, suggesting party affiliation overrides material incentives.

Solar manufacturing and subsidy tradeoffs

Domestic solar panel makers such as T1 Energy and First Solar face structural disadvantage. Chinese manufacturers produce cheaper panels faster. US tariffs redirect supply through Malaysia, Cambodia, Thailand, and Indonesia nominally to avoid Chinese ownership, but tariffs follow those factories too. Developers prefer Chinese panels. Political pressure forces American ones. The supply chain is now capital-intensive and policy-dependent, similar to legacy automakers before Tesla. Subsidies are necessary but create their own problem. If you subsidize American solar, you also subsidize American supply chain jobs, which can conflict with rapid clean energy deployment. Zeitlin cites colleague Casey Hammer's counterintuitive position that if subsidies exist, buy every panel available. Otherwise you waste policy spend on manufacturing that won't scale.