Methods

This project measures how the growth of US data centers is affecting what residential electricity customers pay, with Massachusetts as its focus: the Findings page leads with Massachusetts and New England, and uses PJM, where data center demand is driving costs, as the comparison. It combines utility price and sales data, power market results, and data center locations, refreshes them automatically as sources publish, and publishes the results with the code that produces them.

Approach

1. Build one list of existing and proposed data center sites from four sources, with each site's county, utility, and estimated power.
2. Measure each county's and utility's data center load relative to its households.
3. Assemble a monthly panel of residential price, bill, and usage for about 375 utilities from 2015 to the latest EIA release, with inflation, weather, and measured commercial load growth.
4. Test whether local data center load raises local residential prices, and measure regional effects through PJM's capacity market.
5. Compare four markets with different designs: Dominion (Virginia), Maryland (BGE and Pepco), Texas (ERCOT), and Georgia Power.
6. For Massachusetts: the capacity charge in basic service, New England prices against the rest of the US, utility price trends, competitive supply and town aggregation, and the state's data center sites.

Data sources

How the site stays current

The Data status page and every page footer show how many sources are tracked, which brought new data recently, when each was added, and whether any is late.

Seventeen automated sources are checked on the schedule above (the PJM and market monitor entries are entered by hand). Whenever any source brings new data, the full analysis reruns and the Findings page figures and wording are regenerated from it. When a check finds new data, the pipeline stores a dated copy, rebuilds the tables and analysis, and republishes the site in the same run. Each page footer shows the latest data month. A freshness monitor opens an issue when a source goes quiet longer than expected. Temporary upstream outages (timeouts, rate limits, server errors) are retried on the next run; real errors stop the run.

Some sources publish with a lag: monthly utility sales arrive about two months after the month ends, annual utility data about ten months after year end, and outage history once a year. Capacity auction results and market monitor findings are entered by hand when they are published.

Data center sites

Result: 6,350 records merge into 2,792 sites.

Quality checks (every build; figures on the Data status page):

• Duplicates. Sources sometimes list one building twice (an owner LLC and an operator name, or a renamed listing). Nearby buildings of one operator are usually distinct (Stack NVA05 and NVA08), so pairs within 300 m with the same status are merged only when they share a building code (both "ATL1") or have nearly identical names and power within 25%. Pairs with different building codes (ATL01-1 and ATL01-3, Landbay A and B) are kept separate. Other nearby pairs with a shared operator or similar name are listed for review and kept, so some double counting can remain, most often a campus total listed next to its buildings. Reviewed pairs are recorded with their evidence in reference/datacenter_pair_decisions.csv and applied on every build; review starts with the pairs that carry the most MW.
• Network facilities. PeeringDB lists carrier interconnection points (for example Cogent or Lumen rooms in office buildings) alongside data centers. Unsized records from network carriers are typed as network facilities, hidden on the map by default, and left out of totals.
• Benchmark. Operating capacity is compared with Lawrence Berkeley National Laboratory's estimate of national data center electricity use (176 TWh in 2023); listed capacity at typical utilization exceeds it by 1.5-2x, consistent with capacity being planned build-out.
• Corrections. Anyone can report an error through the correction form. Power is planned capacity, often a full build-out, not average load. The operating total (about 62 GW, range 46-107 GW) is above published estimates of actual US data center load (about 20-45 GW).

Data center load by county and utility

• County: estimated MW by status per 1,000 households (ACS 2024 5-year).
• Utility: each site is placed in the retail service territories that contain it. Where territories overlap, those serving the site's state are preferred (territory boundaries are coarse at state lines), and the site's MW is split by each utility's customer density. Two corrections apply: "serving the state" uses the states EIA lists for the utility (the territory file records headquarters, so AEP Texas appears under Oklahoma), and where the territory file has no customer count (Texas wires utilities such as Oncor and CenterPoint) a company-reported count from reference/territory_customers.csv is used. Before this correction (2026-10-07), Texas data center load was mis-assigned to overlapping co-ops. Sites outside every territory (0.2% of MW) fall back to population-weighted shares of the utilities serving their county. EIA does not publish which utility serves a given site, so this is an approximation.
• Denominator: residential customers from each utility's most recent EIA-861 year within the last three. About 4% of operating MW sits at utilities with no residential count; it stays in totals, but those utilities have no per-customer figure.
• Tiers: operating plus pipeline MW per 1,000 residential customers: high 10 or more, medium 1-10, low under 1, none 0.

Neighborhood costs (map)

The map shows household costs for all 50 states and DC: by county when zoomed out, and by census tract (about 4,000 people each; 84,119 tracts) when zoomed in. Tract data loads one state at a time as you zoom in.

• Utility: each tract is assigned the utility whose retail service territory contains the tract's representative point, using the same overlap rule as data center allocation.
• Average annual bill: that utility's residential revenue divided by customers for full-service customers (EIA-861, latest year). Customers who buy power from third-party suppliers are excluded, because the utility bills them only for delivery.
• State-average fallback: where the tract's utility has sold no power to homes in the last three years (most of Texas, where the wires utility only delivers, and about 2,900 other tracts nationwide), the state's average full residential bill is used and labeled "state average" (6,413 tracts, 3,511 of them in Texas).
• Bill change: inflation-adjusted (Census-region CPI), comparing the 2018-2019 average with the average of the latest two years. Two-year averages damp one-year swings such as co-op refunds and fuel true-ups.
• Energy burden: latest annual bill divided by the tract's median household income (ACS 2024 5-year). It is the only measure on the layer that varies within a utility, because income varies; bills are utility averages, so every tract a utility serves shares them. Nationally the middle 90% of tracts fall between 1.0% and 5.0% (median 2.2%).
• Capacity cost per home: shown for 2026/27 where default supply passes PJM auction prices through (Maryland and DC).
• County layer: household-weighted averages of its tracts' costs, plus estimated operating and pipeline data center MW per 1,000 households.

Prices, controls, and measured load

• Prices: EIA-861M, 2015 to the latest month, excluding EIA's state adjustment and behind-the-meter rows. Residential price is revenue divided by sales; the average bill is revenue divided by customers. In restructured states, utility figures cover customers on default supply.
• Inflation: BLS CPI-U all items for the utility's Census region, in dollars of the latest CPI month. BLS stopped publishing metro-area electricity indexes after December 2024, so regional indexes are used.
• Weather: NOAA county heating and cooling degree days, averaged over each utility's service counties and weighted by population (contiguous US only).
• Measured load growth: trailing-12-month commercial sales compared with the same months of 2019. Data centers usually bill as commercial customers, so this tracks load that has actually arrived. Commercial plus industrial is also kept; it picks up oil field and factory load as well.

Capacity cost per household (PJM, New England, MISO, and New York)

Figures and statements in this section update with each analysis run (last run October 28, 2026). See Data status for how they are checked.

PJM, the grid operator for 13 states and DC, buys generating capacity ahead of time in its Base Residual Auction. Prices rose from $28.92 per MW-day for 2024/25 to $269.92 for 2025/26, then reached the price cap in each of the 2026/27 to 2028/29 auctions ($329.17, $333.44, $325.00). Without the cap, the 2028/29 auction would have cleared at $554.72.^1,2,3

• Cost per home per year = (auction total ÷ PJM annual energy) × (zone price ÷ PJM-wide price) × household kWh per year × residential peak factor.
• Peak factor: capacity is charged on each customer's share of the system peak, and homes peak harder than their share of energy. The central estimate uses 1.2, with 1.0 and 1.4 as the range. A cross-check using a household peak contribution of 2.0-2.5 kW gives $240-300 a year at the capped price.
• Data center share: the market monitor estimates that existing and forecast data center load added $9.3 billion to the 2025/26 auction, $7.3 billion to 2026/27, and $6.5 billion to 2027/28 ($23.1 billion in all). Each year's share of auction cost is applied to that year's household cost. No attribution has been published for 2028/29.
• Who pays: auction prices reach customers directly in restructured states (Illinois, New Jersey, Pennsylvania, Ohio, Maryland, Delaware, DC), where default supply passes them through. Regulated utilities mostly recover the cost of their own plants. Dominion supplied its own capacity outside the auction through 2024/25; AEP's regulated utilities still do.
• Upper bound: the auction total is cleared MW times price. PJM notes this overstates cost to customers, because some load is hedged or self-supplied.

Result: for a typical home in a restructured PJM state (about 10,401 kWh a year), capacity cost was $46 a year in 2024/25, $218 ($181-$254) in 2025/26, $238 ($199-$278) in 2026/27, and $243 ($202-$283) in 2028/29: about $197 a year more than 2024/25. The market monitor's findings attribute about $96-$138 a year of it to data center load. The 2028/29 level is the highest in the series, about 5 times 2024/25 and 1.5 times 2018/19.⁴

New England. ISO New England's Forward Capacity Auctions run three years ahead with the same June-to-May years. The region-wide price fell from $9.55 per kW-month for 2018/19 to $2.59 for 2026/27, then rose to $3.58 for 2027/28; the next auction, for 2028/29, is due in 2028 after a delay for market redesign. With the same method and New England's annual energy, a typical home's capacity cost fell from $287 a year to $92. All New England states but Vermont pass auction prices through default supply. No data center attribution has been published; the region has 0.7 GW of data centers planned or under construction, about 0.1 MW per 1,000 homes against 5.3 in PJM.

MISO. MISO's Planning Resource Auction is held each spring for the coming June-to-May year, with separate summer, fall, winter, and spring prices since 2023/24; seasonal prices are weighted by days. Most MISO utilities are regulated and supply their own capacity, so the site applies auction prices only where default supply buys capacity in the auction: Ameren Illinois (Zone 4). Its typical home's cost ranged from $3 to $224 a year (2022/23); for 2026/27 it is $123. No data center attribution has been published.

New York. NYISO buys capacity in seasonal and monthly auctions and a monthly spot auction, priced separately for New York City, Long Island, the lower Hudson Valley (the G-J Locality), and the rest of the state. The site averages the 12 monthly spot prices in each May-to-April capability year (complete years only, through 2025/26), read from NYISO's public auction viewer each week, and uses an approximate statewide requirement of 36 GW (forecast peak of about 31.5 GW plus the 18-24% installed reserve margin, less derating). All seven large New York utilities pass these costs through, including the Long Island Power Authority through its power supply charge. A typical home's cost ranged from $47 to $125 a year; Con Edison's peaked at $261 (2023/24). New York City's spot price reached $32.59 per kW-month in May 2026.

Massachusetts

Figures in this section update with each analysis run. Code: analysis/massachusetts.py.

• Capacity charge per home. Median across the Massachusetts default-service utilities (Eversource, National Grid's Massachusetts Electric and Nantucket Electric, Unitil), from ISO New England Forward Capacity Auction prices for their zones, by the same method as every market (see Capacity cost per household). Basic service passes the auction price through to customers who have not chosen a supplier.
• New England vs. the rest of the US. The same event-study regression as PJM's (utility and month fixed effects, weather controls, reference year 2020, standard errors clustered by utility), with New England utilities against the rest of the US outside PJM. PJM is left out of the comparison group because its own rise would make New England look lower. The coefficient measures how much New England prices changed relative to that group since 2020, not the difference in price levels.
• Utility price trends. Residential revenue per kWh from EIA-861M by full calendar year, deflated by the Northeast CPI. Utility figures cover homes on the utility's own basic service. Municipal utilities are combined, weighted by sales.
• Competitive supply and town aggregation. Customer counts by utility from the Massachusetts DOER customer choice data (residential rate classes, including low-income and heat-pump rates). Prices from EIA-861 annual: the supply part of basic service is investor-owned utilities' full-service revenue per kWh minus their delivery-only revenue per kWh; the supplier price is the energy-only revenue per kWh of competitive suppliers selling to Massachusetts homes. The supplier average mixes town aggregation (which buys in bulk) with individual contracts, so it does not show what a household on an individual contract pays. The dollar difference multiplies the price gap by suppliers' residential sales, an estimate of what those homes would have paid on basic service at the average basic service price.
• Data centers. Massachusetts sites from the merged site list (see Data center sites), with New England totals by state.
• Known data issues. reference/data_issues.csv lists reported figures that cannot be right, with the evidence. Those figures are left out of prices, bills, statements, and the regressions; customer and sales counts stay. Current entry: Eversource (NSTAR Electric) residential revenue in EIA-861M from March 2025, which implies about 20 cents per kWh when its basic service supply charge alone was 13 to 15 cents.

Panel regressions

Figures and statements in this section update with each analysis run (last run October 28, 2026). See Data status for how they are checked.

The outcome is the log of inflation-adjusted residential price for 326 utility service areas, January 2015 to August 2026, with utility and month fixed effects, heating and cooling degree days, and standard errors clustered by utility.

• Local load: comparing each utility only with others in the same regional grid in the same month, a 40% rise in its own commercial load goes with a 1.0% lower residential price (95% CI 0.4% to 1.6% lower). Local load growth has not raised local residential prices relative to regional peers; more sales spread fixed costs.^5,6 EPRI's state-level causal study, using a different method (an instrumental variable for where data centers were built), finds the same direction: data center growth from 2019 to 2024 lowered average residential rates about 6%.⁵
• PJM vs. the rest of the US (2020 = 0): no difference from 2017 to 2021; +3.3% in 2022, +11.1% in 2023, +11.1% in 2024, +12.6% in 2025, +15.7% in 2026. The gap opened before capacity prices spiked, consistent with 2022 natural gas prices being locked into default-supply contracts, and widened as capacity costs reached bills from mid-2025.
• Within PJM, restructured vs. regulated: +18.4% in 2026, but the two groups were not on parallel paths before 2020, so this comparison is weaker evidence.
• New England vs. the rest of the US outside PJM, relative to 2020: +22.0% in 2023 and +9.5% in 2026 (95% CI +2.8% to +16.7%).

Reliability

Figures and statements in this section update with each analysis run (last run October 28, 2026). See Data status for how they are checked.

• Utility test. EIA-861 SAIDI (outage minutes per customer per year) and SAIFI (outages per customer), excluding major event days, 2013-2025, for 1,342 utility-state units. The IEEE 1366 figure is used where reported, otherwise the utility's state-method figure (25% of utility-years, including Dominion); each utility and method gets its own fixed effect. Regressed on the log of the utility's commercial sales with balancing authority x year fixed effects. A 40% rise in commercial load goes with −2.2% outage minutes (95% CI −4.6% to +0.4%) and −1.8% outages; including major storms, −4.3% (−7.2% to −1.3%). IEEE-only check: −0.6% (−3.8% to +2.8%). No sign that reliability has worsened where commercial load grew.
• County check. ORNL EAGLE-I customer-hours without power per customer, 2018 onward. The 32 counties with at least 10 MW of operating data center capacity per 1,000 households had the same or fewer outage hours than other counties in their state each year (2024: −35%, the year Hurricane Helene hit rural areas). Descriptive: these counties are mostly suburban, with more underground lines.
• Limits. These measure local distribution outages that have already happened. Bulk-supply risk is forward-looking: PJM's 2027/28 capacity auction cleared about 6,600 MW below its reliability requirement.

Case studies

Figures and statements in this section update with each analysis run (last run October 28, 2026). See Data status for how they are checked.

Inflation-adjusted residential price, January to August 2026 compared with the same months of 2025:

Resident bills

Residents can add the numbers from a recent electric bill through a Google Form (linked from the Your utility page). The form asks for ZIP code, utility, bill date, kWh, total amount, and optionally days billed, supplier, and electric heat. It does not collect names, emails, addresses, account numbers, or files, and does not require sign-in.

• Responses are stored in a private Google Sheet. A small public endpoint (Google Apps Script) returns only aggregates: the exact number of valid bills per ZIP code from the past 12 months, and, for ZIP codes with at least 10, the median monthly bill, median monthly kWh, and median price per kWh. Individual responses are never published.
• Supply vs. delivery. Optional questions split the bill into supply charges (the electricity itself: generation or energy charges, or a retail provider's charges), delivery charges (distribution, transmission, riders, and the fixed customer charge), the fixed customer charge on its own, and separately listed taxes and fees. A bill's split is used only if supply + delivery + taxes is within 8% of its total. Once a ZIP code has at least 10 bills with a usable split, the site shows median supply and delivery cost per kWh, the median fixed charge, and taxes and fees as a share of the bill. Data center growth can raise either part: supply through capacity and energy prices, delivery through new lines and substations.
• Bills are normalized to a 30-day month. A bill counts as valid if its date is within the past 12 months, normalized usage is 50-8,000 kWh, and the implied price is 5-80¢ per kWh.
• Pending counts are exact, so in a ZIP code with few homes a count of one or two can reveal that someone nearby submitted.
• Submissions are not verified; one person can submit more than once. Counts are labeled as pending until a ZIP code reaches the threshold, and medians limit the effect of outliers.

How findings stay current

Each statement on the Findings page, and each result statement on this page, is stored in analysis/claims.yaml with the condition it rests on: for example, that a confidence interval lies entirely below zero, that one market's price rose more than another's, or that a figure is the highest on record. Figures fill in from each analysis run; facts entered by hand (rate orders, auction shortfalls, the number of PJM states) live in reference/context_facts.csv with their source and a review-by date.

• Every analysis run checks every condition. A statement whose condition no longer holds is replaced by its current figures without interpretation, marked "under review", and opened as a GitHub issue. Some statements have wording already reviewed for more than one outcome (for example, a local price effect that is negative, or indistinguishable from zero); the matching one is used.
• Each key figure also has a plausibility range. A figure outside it is treated as a possible data error rather than a new finding: the statements that use it are held for review, and the issue asks for the source data to be checked first.
• Rewritten wording is drafted with Claude and reviewed by the project owner before it is published; the change history is public in the repository.
• Hand-entered facts past their review date are listed on the Data status page and opened as an issue.

Feedback

Every page links to a "Report an issue" form (Google Forms, no sign-in). It asks which page, what kind of issue, and what was seen, with an optional place and an optional email for a reply. Reports go to a private sheet and are never published; the pipeline only counts them so the owner is notified.

What the monthly data cover

EIA's monthly survey (EIA-861M), the source of utility prices and bills, reports each utility's residential customers, sales, and revenue for homes on the utility's own (basic or default) supply. Two groups of homes are left out:

• Competitive supply and aggregation. In restructured states, many homes buy supply from a competitive supplier, a retail electric provider (Texas), or a town or community choice aggregation program. The utility still delivers their power, but EIA's monthly survey counts them only in a statewide total, not by utility. In 9 states (CA, CT, IL, MA, NH, OH, PA, RI, TX), at least 10% of homes, about 20,704,874 in all, buy supply from competitive suppliers or aggregation programs and are not in the utility-level prices.
• Smaller utilities. Co-ops and municipal utilities outside EIA's monthly sample report only once a year. They are not shown on the site.

Where a state publishes monthly switching counts by utility, those replace EIA's annual split: the Illinois Commerce Commission (ComEd, Ameren Illinois), the Pennsylvania PUC's PAPowerSwitch statistics, and the New Jersey Board of Public Utilities (switching and government aggregation). Massachusetts publishes the same data (Department of Energy Resources customer choice data, monthly from 2020, split into basic service, individual competitive supply, and municipal aggregation), but its website blocks automated downloads, so it is added by hand each quarter. Where a utility is missing from EIA's newest annual file (an early release), its latest earlier year is used for delivery and supply prices. The Data status page lists every state below 90% coverage and splits its missing homes by reason, using EIA's annual survey (2025). The Your utility page and the map flag affected utilities. Where EIA's annual survey reports a utility's delivery-only customers, the Your utility page estimates what those homes pay: the utility's delivery revenue per kWh plus the statewide average price charged by competitive suppliers. The statewide supplier average mixes aggregation programs and individual contracts, so it is a rough guide.

Known limitations

• Site lists are incomplete, and most sites don't disclose power; estimates carry wide ranges.
• Site-based load is a current snapshot. Measured commercial load growth is the time-varying measure.
• Utility assignment of sites is approximate where territories overlap.
• Price differences have many causes. The regressions control for weather, inflation, and region, but the capacity-cost channel is the most direct causal link.
• Connecticut's counties differ between Census (planning regions) and EIA (legacy counties).
• Smaller utilities that report to EIA only annually are not in the monthly panel.

Changelog

• 2026-10-13: References batch 2 (Massachusetts): the Attorney General's 2026 competitive supply report, ISO New England's first large-load forecast (CELT 2026), and DOER's customer choice data, now cited in the Massachusetts sections.
• 2026-10-13: The site now focuses on Massachusetts. Findings leads with Massachusetts and New England: the capacity charge in basic service, New England prices vs. the rest of the US outside PJM (new regression), Massachusetts utility price trends, competitive supply and town aggregation, and the state's data center sites. PJM and the other markets follow as the comparison. The map opens on Massachusetts, and Your utility on National Grid with Eversource one click away. New table of known data issues: Eversource's EIA-861M revenue from March 2025 is left out pending EIA revision; dropping those months moves the national results slightly (local effect −0.9% to −1.0% at 40% more load; PJM 2026 gap +15.7% to +15.5%).
• 2026-10-09: Review of the remaining 404 flagged data center pairs: 54 more records removed (about 5.7 GW), mostly building-level OpenStreetMap or PNNL records on campuses whose FracTracker record already covers the campus total (Meta, Google, Microsoft, Switch), and projects listed twice (Stream Atlanta West, Keel Panther Creek, Microsoft East Point). 340 pairs confirmed as separate, most of them different county parcels in Northern Virginia. Six pairs remain for review, none with more than 30 MW at stake. PJM's planned and under-construction data center load changes from 133 to 132 GW.
• 2026-10-08: New York (NYISO monthly spot capacity prices by locality, 2018/19 to 2025/26) added for seven utilities; collected weekly from NYISO's public auction viewer.
• 2026-10-08: Correction: the rule that prefers utilities serving a site's or tract's own state had been using each utility's headquarters state only, because of a data-type error. Multi-state utilities now count as in-state wherever EIA lists them. Ameren Illinois, for example, now covers 824 Illinois tracts instead of 116, and about 20 GW of planned data center load moved from rural co-ops to the utilities that serve those areas (Southwestern Public Service in New Mexico, AEP Texas, Ameren Illinois, SWEPCO). National totals are unchanged. Glossary added.
• 2026-10-08: MISO (Planning Resource Auction, 2018/19 to 2026/27) added for Ameren Illinois, the MISO utility whose default supply buys capacity in the auction.
• 2026-10-08: New England (ISO-NE Forward Capacity Auctions, 2018/19 to 2027/28) added to capacity cost per household, the Findings chart, the map, and the Your utility page.
• 2026-10-08: Massachusetts DOER customer choice data added (through June 2026): 82% of Eversource and 69% of National Grid residential customers are on competitive supply or municipal aggregation. Utility-level coverage now uses each utility's latest EIA annual year (58 utilities with a customer split).
• 2026-10-08: State switching statistics (IL, PA, NJ) added for utility-level coverage, checked weekly.
• 2026-10-08: Coverage flags: states and utilities where EIA's monthly figures miss homes on competitive supply or at annual-only utilities are flagged on the Your utility page, the map, and the Data status page, with an estimated price for competitive-supply homes where EIA reports the split.
• 2026-10-08: References page added (52 sources: research, filings, market monitor reports, news). Corrections found while compiling it: PJM's 2027/28 shortfall is 6,623 MW (PJM's release; previously about 6,500 from a secondary source), and the data center share of 2026/27 and 2027/28 capacity cost now uses the market monitor's published figure for each auction ($7.27B and $6.50B) instead of splitting the three-auction total by auction cost.
• 2026-10-08: Findings and Methods wording generated from checked statements with conditions and plausibility ranges; hand-entered facts moved to a sourced reference table with review dates; "Report an issue" form added to every page.
• 2026-10-07: Bill form adds optional supply, delivery, fixed charge, and taxes questions; published medians split supply vs. delivery.
• 2026-10-07: Resident bill form launched with per-ZIP pending counts and medians at 10+ bills.
• 2026-10-07: Reliability analysis added (utility SAIDI/SAIFI and county outages); outage minutes per customer added to the Your utility page.
• 2026-10-07: Manual review of the 30 flagged pairs carrying the most MW: 14 duplicate or component records removed (about 6.9 GW, mostly proposed campuses listed twice), 16 pairs confirmed as separate.
• 2026-10-07: Data center quality checks: duplicate merging (9 listings, 927 MW), network facilities typed and hidden by default, pairs pending review published, correction form added.
• 2026-10-07: Household cost layer extended to all 50 states and DC (84,119 tracts). Data status page and footer added. Texas utility assignment corrected (see "Data center load by county and utility"), which moved about 4.6 GW of operating and 22.5 GW of pipeline data center load in Texas to Oncor.
• 2026-10-07: Map replaced with a zoomable map (MapLibre, OpenFreeMap basemap) with clustered data center sites, a neighborhood cost layer for Virginia, Maryland, and DC, and a national county layer.
• 2026-10-06: Public site launched.
• 2026-10-05: First descriptive comparison (customer-weighted, no controls): residential prices rose 49.6% in PJM vs. 33.7% elsewhere between Aug 2020-Jul 2021 and Aug 2025-Jul 2026. Superseded by the panel regressions above.
• 2026-10-05: AEP's regulated PJM utilities confirmed as self-supplying capacity (Utility Dive, 2026). Maryland case added.

Data downloads and licenses

Built by analysis/exports.py after every analysis run. Do not edit by hand.

Keys in the site files are abbreviated: x/y lon/lat, g status group, mw/lo/hi estimated MW and range, b MW basis, n name, o operator, c/s county/state, src sources, eta expected online.

Files: capacity_household_cost.json, case_studies.json, counties.json, counties_geo.json, coverage.json, crowd.json, event_studies.json, massachusetts.json, meta.json, reliability.json, sites_fractracker.json, sites_open.json, utility_exposure.json, utility_reliability.json, utility_usage_bill.json

Glossary

95% confidence interval (CI)

The range a result could plausibly fall in given the data's noise. When the interval excludes zero, the effect is statistically distinguishable from no effect.

Capacity auction

An auction in which a grid operator pays power plants, batteries, and demand-response providers to be available to run during the highest-demand hours, usually one to three years ahead. Utilities and suppliers pay the cost and recover it on customers' bills. PJM's is the Base Residual Auction, ISO New England's the Forward Capacity Auction, and MISO's the Planning Resource Auction.

Capacity price ($/MW-day, $/kW-month)

The price of making one megawatt of capacity available for one day. ISO New England quotes prices per kilowatt per month; the site converts them (1 $/kW-month is about $32.88/MW-day).

Competitive supplier

A company other than the utility that sells the electricity supply portion of a bill in a restructured state (called a retail electric provider in Texas). The utility still delivers the power and sends the bill.

Default service (basic, standard offer)

The supply a customer in a restructured state gets from the utility without choosing a supplier. Names vary by state: basic service (Massachusetts), standard service (Connecticut), standard offer service (Maryland, Maine), default service (Pennsylvania). The utility buys the power in wholesale markets and passes the cost through.

Delivery year

The 12 months a capacity auction buys for, June 1 through May 31 (written 2026/27).

Energy burden

The share of household income spent on energy bills. The site's measure covers electricity only.

Grid operator (RTO or ISO)

A regional transmission organization or independent system operator: a nonprofit that runs the high-voltage grid and wholesale power markets for a region. It schedules power plants and plans transmission but does not own plants or sell electricity to homes.

ISO New England (ISO-NE)

The grid operator for Connecticut, Maine, Massachusetts, New Hampshire, Rhode Island, and Vermont. It runs the region's wholesale markets and the Forward Capacity Auction.

kWh, MW, GW, TWh

Units of electricity. A kilowatt-hour (kWh) is the energy a 1,000-watt appliance uses in an hour; a typical US home uses about 10,000 kWh a year. A megawatt (MW) is 1,000 kilowatts of power, a gigawatt (GW) 1,000 MW, and a terawatt-hour (TWh) a billion kWh. Large data centers draw 100 MW or more.

Market monitor

An independent firm that watches a grid operator's markets for problems and reports on them. PJM's is Monitoring Analytics, whose reports estimate how much data center load added to capacity costs.

MISO

The Midcontinent Independent System Operator, the grid operator for 15 states from Manitoba's border to Louisiana, including most of Illinois outside Chicago, Michigan, Minnesota, and Iowa. It runs a yearly Planning Resource Auction for capacity, with separate prices for each season since 2023/24.

Municipal or community choice aggregation

A program in which a city or town buys supply for all of its residents from a competitive supplier, unless a resident opts out. Common in Massachusetts, Ohio, Illinois, and California.

NYISO

The New York Independent System Operator, the grid operator for New York State. It buys capacity in seasonal, monthly, and monthly spot auctions, priced separately for New York City, Long Island, the lower Hudson Valley, and the rest of the state, on May-to-April capability years.

Peak factor

How much more of the system's peak a home is responsible for than its share of yearly electricity. Capacity is charged on contribution to the peak, and homes use relatively more at peak times (summer evenings). The site uses 1.2, with 1.0 and 1.4 as the range.

PJM

PJM Interconnection, the grid operator for 13 states and Washington, DC, from New Jersey to Illinois (about 67 million people). It runs the high-voltage grid, the wholesale electricity markets, and a capacity auction. It owns no power plants and does not sell to homes. The name comes from the Pennsylvania-New Jersey-Maryland Interconnection, the power pool it began as in 1927. Northern Virginia's data center cluster is inside PJM.

Real (inflation-adjusted) price

A price expressed in today's dollars using the Consumer Price Index for the utility's Census region, so changes reflect more than general inflation.

Regulated (vertically integrated) utility

A utility that owns power plants as well as the wires and sells both, with rates set by the state regulator. It mostly recovers the cost of its own plants rather than paying capacity auction prices.

Restructured state

A state where the local utility delivers power but customers can buy the electricity itself (supply) from competitive suppliers. Customers who do not choose get the utility's default service. In these states, capacity auction prices pass directly into bills.

SAIDI and SAIFI

Standard reliability measures: SAIDI is the average minutes without power per customer per year; SAIFI is the average number of outages per customer. Usually reported with and without major storms.

Supply and delivery charges

The two main parts of a bill in restructured states. Supply is the electricity itself (generation and capacity); delivery is the cost of the wires, poles, and substations that bring it to the home, set by the state regulator.