Capacity charges, peak load auctions, and the coincident-peak trap: a plain-language guide to the parts of your electric bill you've never heard of — and why data center growth is making them explode.
If your electricity bill has been climbing and you can't figure out why — you're not using more power, you haven't added appliances — the answer is probably hiding in a line item you've never looked at.
It's not the kilowatt-hours. It's the capacity charges.
Most people think of electricity as a single rate: use more, pay more. In reality, your bill has three major components, and the one growing fastest is the one almost nobody understands:
| Component | What it pays for | Typical share | Growing? |
|---|---|---|---|
| Energy charges | The actual electricity you consumed (kWh) | 40–60% | Slowly |
| Capacity charges | Keeping enough power plants available to meet the highest hour of demand all year | 15–30% | Fast |
| Transmission & distribution | Wires, substations, and poles to move power from plant to home | 20–30% | Moderately |
The energy charge is intuitive — run the AC more, pay more. But capacity charges are different. You pay them whether you use the power or not, because the grid must maintain enough generation to handle the worst-case peak — even if that peak lasts only a few hours each summer.
Think of it like a fire department: your taxes pay for fire stations that sit idle most of the year, because they need to be there the one day your house catches fire. Capacity charges are the grid's fire department. And right now, someone is building a factory that requires its own fire station — and sending you part of the bill.
Here's where it gets personal. In most deregulated markets, your capacity charge isn't based on your individual peak usage. It's based on your usage during the system coincident peak (CP) — the single highest-demand hour across the entire grid that year.
If a new 200 MW data center comes online in your utility's territory and raises the system peak, your capacity allocation goes up even though your behavior didn't change. The new load raises the waterline, and everyone pays more to keep the system above it.
This isn't hypothetical. In PJM — the grid operator serving 65 million people across 13 states from New Jersey to Illinois — it just happened:
That's not a rate increase driven by fuel costs, inflation, or your usage. It's a rate increase driven by someone else's load.
In February 2025, researchers at Duke University's Nicholas Institute for Energy, Environment & Sustainability published a study that reframed the entire debate. Led by Tyler Norris, the analysis introduced a concept called "curtailment-enabled headroom" — how much new load the existing grid can absorb if that load agrees to briefly reduce consumption during the handful of hours each year when the system is most stressed.
The findings were striking:
In plain English: if data centers agreed to dim the lights for less than two days a year — spread across the summer's hottest afternoons — we wouldn't need to build tens of billions of dollars in new infrastructure, and your capacity charges would stay flat.
This is the question everyone asks, and the answer reveals a structural failure in how we regulate large electricity consumers.
1. The costs they impose aren't the costs they pay.
When a data center raises the system peak, the resulting capacity charges are socialized across all ratepayers. The data center pays its share, but the incremental system-wide cost it imposes — the billions in new capacity procurement triggered by its load — is spread across millions of customers. There's no price signal telling the operator: "Your consumption during this hour just cost the grid $50 million in capacity obligations."
Under current rate design, a data center that curtails during peak hours saves almost nothing on its own bill. The capacity auction was settled months ago; the price is already set. Curtailment is a cost with no reward.
2. Uptime SLAs are contractually sacred.
Cloud and colocation contracts guarantee 99.99–99.999% uptime — "four nines" to "five nines." Five-nines means a maximum of 5.26 minutes of total downtime per year. The Duke study's 44-hour curtailment, even if it's just a partial load reduction, would blow through any existing SLA by orders of magnitude.
Renegotiating these contracts means: - SLA breach penalties (often millions per incident) - Customer churn risk (if AWS curtails but Azure doesn't, customers switch) - Insurance and liability exposure (financial, healthcare, and government workloads have legal uptime requirements)
Here's the irony: AI training workloads are actually highly flexible. A training run can pause, checkpoint, and resume — it doesn't care about latency or real-time availability. But operators run training and inference on shared infrastructure and apply the strictest SLA to everything. Separating these workloads is technically straightforward but commercially inconvenient.
3. There's no regulatory mandate.
Unlike power plants, which must bid into capacity markets and face penalties for non-performance, data centers have no obligation to participate in demand response. They're classified as ordinary load. They consume what they want, when they want, and the grid must accommodate them.
Several mechanisms could change this but don't exist at scale:
The Duke study didn't land in a vacuum. A growing body of academic and policy research is wrestling with the same questions:
Lawrence Berkeley National Lab (LBNL, Jan 2025) found U.S. data center electricity surged from 58 TWh (2014) to 176 TWh (2023) and projects 325–580 TWh by 2028 — potentially 12% of all U.S. electricity. In July 2024, a voltage fluctuation in Northern Virginia triggered simultaneous disconnection of 60 data centers, causing a 1,500 MW surplus that required emergency grid adjustments.
The Harvard Belfer Center (Feb 2026) called AI-driven load growth a "watershed moment" for grid planning, noting that traditional forecasting methods are failing because demand is growing faster than any historical precedent.
E3, funded by Amazon (Dec 2025), studied four Amazon facilities and concluded data centers generate $3.4 million in surplus revenue per 100 MW facility — paying more than their direct cost to serve. This is the industry's primary counterargument to the "ratepayers are subsidizing data centers" narrative.
The critical nuance: both things can be true simultaneously. A data center can pay more than its direct cost-to-serve while also driving up system-wide capacity costs that are socialized to everyone. E3's facility-level analysis and PJM's system-level market monitor are measuring different things. The surplus at the meter doesn't capture the externality at the auction.
Columbia University (2025) showed that grid-enhancing technologies (dynamic line ratings, power flow controllers) could release 20–40% more capacity from existing transmission — deferring $10–30 billion in new construction.
The research converges on a handful of structural reforms:
Cost-causation rate design — charge large loads for the capacity and transmission costs they cause, not the system average. FERC and several state PUCs are investigating this, but no major market has implemented it yet.
Mandatory demand response for large loads — if you consume more than 10 MW, you participate in curtailment programs, period. At least five state legislatures introduced versions of this in 2026.
Load flexibility contracts — utilities offer lower rates in exchange for contractual curtailment rights during peak hours. Duke Energy and Dominion are piloting programs, but participation is voluntary and uptake is low.
Interconnection reform — FERC Order 2023 is speeding up queue processing and requiring deposits to prevent speculative capacity hoarding, but implementation is slow.
On-site generation requirements — require large loads to provide their own peaking capacity (batteries, on-site generation) so the grid doesn't have to overbuild for them. Proposed in North Carolina, Virginia, and Georgia.
The Duke University research proves the technical solution exists: brief, modest curtailment — less than two days a year — could avoid tens of billions in new infrastructure and keep your capacity charges from spiraling. The barrier isn't engineering. It's a regulatory and commercial framework that lets the largest electricity consumers externalize their peak-load costs onto everyone else's bill.
Until that framework changes, residential ratepayers bear the cost of keeping the grid ready for loads that refuse to flex.
Explore more: Use the 💡 Your Utility Bill tab for an interactive breakdown of bill components, the 🕐 Grid Timing tab to see real-time grid stress in your region, and the 🏛️ Officials tab to contact your legislators about rate reform.