Your utility bill: what's actually driving it up

Fuel & purchased power — what you use

The fuel and wholesale power a utility buys to serve you — the largest single expense (~36%) and the most volatile. It drove the 2021–22 bill spike as natural-gas prices surged (gas sets ~40% of U.S. generation), and eased as they fell back.

The grid — what must exist

Wires, substations, and their upkeep, sized for the single highest moment of demand, not the average. Distribution is now the largest and fastest-growing slice of utility investment — capital spending on it grew ~50% from 2019–2023, even as spending on power plants fell.

The long-run shift

Over two decades the balance has tilted from producing power toward delivering it: EIA finds power production fell from 69% (2006) to 54% of utility costs while delivery climbed to 46%. New large loads accelerate that grid spending — and it's recovered from everyone.

Deregulated markets (PJM, ISO-NE, NYISO) — Capacity is procured through auctions. The cost is allocated to utilities based on their total load during system peak hours, then passed through to customers. Residential customers see it as a line item or bundled into the default service rate. C&I customers see explicit demand and capacity charges and can manage them.

Regulated / vertically integrated markets (Duke, Southern, Entergy) — No separate capacity market. The utility owns generation and recovers costs through base rates set in rate cases. When it builds new capacity to serve data center growth, the capital is rate-based and recovered from all customers through higher per-kWh rates.

ERCOT (Texas) — No capacity market at all. Texas uses an energy-only market where scarcity pricing during peak hours is supposed to incentivize generation investment. Residential customers on variable-rate plans are directly exposed to wholesale spikes (which is why bills exploded during Winter Storm Uri). Data centers can negotiate bilateral PPAs that lock in low prices, shifting scarcity cost to the remaining pool.

In all three structures the pattern holds: large industrial loads have tools, tariffs, and negotiating power to manage their exposure. Residential customers absorb socialized costs with no visibility and no control.

Capacity obligation — grid operators run auctions years in advance so power plants commit to being available during peak. Those commitments cost money whether or not the plants run.

Peaker plants — gas plants that run only 50–200 hours a year but must be maintained year-round. Their per-MWh cost is enormous.

Transmission upgrades — wires are sized for the peak, not the average. Building for a 2 GW peak instead of 1.5 GW can mean billions in new lines.

The "coincident peak" trap — many utilities set your capacity charge from your usage during the single highest-demand hour across the grid that year. A large new load raises that peak, and every customer's allocation increases, even those whose own usage didn't change.

Your utility buys capacity in bulk. It must procure enough to cover every customer's share of the system peak, plus a reserve margin. For a utility serving 1 million homes in PJM, the 2025/26 auction cost roughly $800M–$1.2B in capacity obligations alone — before a single electron flows.

Allocation via your capacity tag. In PJM every customer gets a Peak Load Contribution — your usage during the five highest-demand hours of the prior summer, averaged. Ran the AC hard on those afternoons? You carry a bigger share. Most customers have no idea which hours set it.

Line item — or buried. Deregulated utilities (Pepco, BGE, PECO) show an explicit capacity line item; regulated ones (Duke, Dominion, Southern) roll it into a blended supply rate; co-ops bury it in the base rate. Pepco customers got a notice in spring 2025 that capacity charges were rising ~$10/month.

Time-of-use pricing. Some utilities charge more during peak hours (2–7 PM summer weekdays) to push usage off-peak — your bill depends on when you use power, not just how much.

• Scope: 22 of the largest U.S. balancing areas (~80% of demand)
• Method: production cost modelling with incremental load additions and curtailment constraints
• Key innovation: the curtailment-enabled headroom metric — how much load can be added before reliability standards are violated, at a given curtailment rate
• Result: at 0.5% curtailment, headroom ranges from 2–15 GW per balancing area, ~98 GW nationally
• Comparison: existing demand response programs already curtail at comparable rates (FERC Order 2222 resources average 1–3%)

Norris, T. et al. (2025). "Curtailment-Enabled Headroom: How Flexible Large Loads Can Accelerate Grid Integration." Nicholas Institute, Duke University.

Misaligned incentives

When a data center drives up the system peak, the resulting capacity charges are spread across all ratepayers — the incremental cost it imposes is socialized. There's no price signal saying "your load this hour just cost the grid $50M in capacity obligations." And curtailing saves the operator nothing on its own bill, since auction prices are set months in advance, while every hour of curtailment risks SLA penalties and lost revenue.

Uptime SLAs are contractually sacred

Cloud contracts guarantee 99.99–99.999% uptime. Voluntary curtailment triggers SLA breach penalties (millions per incident), customer churn, and liability exposure. The irony: AI training is actually flexible — runs can pause and resume — but operators bundle training and inference on shared infrastructure and apply the strictest SLA to everything.

No regulatory mandate

Unlike power plants, data centers have no obligation to participate in demand response — they're treated as ordinary load.

Lawrence Berkeley National Lab, 2024 US Data Center Energy Usage Report — 325–580 TWh (6.7–12% of US electricity) by 2028, up from 176 TWh (4.4%) in 2023

• U.S. data center electricity climbed from 58 TWh (2014) to 176 TWh (2023)
• Projected 325–580 TWh by 2028 (6.7–12% of total U.S. electricity)
• Demand growth has tripled over the past decade and is projected to double or triple again by 2028
• In some regions AI-driven demand is outpacing capacity, forcing companies to install inefficient on-site generators

Reliability incident (Jul 2024): a voltage fluctuation in Northern Virginia triggered simultaneous disconnection of 60 data centers — a 1,500 MW surplus requiring emergency grid adjustments to prevent cascading outages.

Harvard Belfer Center — AI, Data Centers, and the U.S. Electric Grid: A Watershed Moment (Feb 2026)

• Traditional load forecasting is failing — AI demand is growing faster than any historical precedent
• Regional concentration creates localized reliability risks that national statistics obscure
• Recommends mandatory demand response for large loads and reformed interconnection processes

E3 (for Amazon) — Tailored for Scale: Designing Electric Rates and Tariffs for Large Loads (Dec 2025). Utility-funded: commissioned by Amazon, and concludes its data centers are not cross-subsidized

• Studied Amazon facilities across four utility territories (PG&E, Umatilla, Dominion, Entergy)
• Found data centers generate $3.4M surplus revenue per 100 MW facility (2025), rising to $6.1M by 2030
• Concludes data centers are net contributors, not subsidized

Important context: the study examines individual facilities in isolation and does not model the system-wide capacity market and transmission effects PJM's market monitor attributes to data center growth. Both findings can be true — a facility can pay more than its direct cost-to-serve while still driving up socialized system-wide costs.

Columbia CGEP — Electricity Affordability and Load Growth: grid-enhancing technologies, demand response and data-center load flexibility as near-term capacity (Jun 2026)

• Dynamic line ratings, power flow controllers, and topology optimization could release 20–40% more capacity from existing transmission without new construction
• Combined with demand response, could ease price pressure through 2030
• Estimated to defer $10–30B in transmission investment nationally

UC Berkeley Energy Institute — What Will Data Centers Do To Your Electric Bill? (2025)

• Investor-owned utilities sought $18 billion in rate increases in 2025 — the most since the mid-1980s
• Residential prices rose 6% nominal (2× inflation)
• Capacity market costs are the fastest-growing bill component in RTO markets, with data centers the primary demand driver
• Recommends large loads bear their full marginal cost of service, not just embedded average costs