The data centers on your power bill mostly haven't been built yet

On June 1, 2026, the typical Pepco household in Washington, DC, using 614 kWh a month, started paying about $8.79 more a month, a 6.2% increase. The DC Public Service Commission’s explanation named higher regional demand from data centers and PJM’s “record-high capacity prices for the second consecutive year.” Most accounts of AI and electricity now say that the server halls arrive, the bills go up, and households pay for someone else’s chatbot.

I build public trackers on this, from a map of AI data-center sites to PJM capacity prices, and my read is that the story is right in one place and mostly wrong everywhere else. In PJM, the 13-state grid from Chicago to Washington, data centers are the main cause of a capacity-price shock: the market monitor puts them at 38 to 46% of recent capacity costs. Nationally, though, the rise in residential bills is mostly about gas, weather, wildfires and the cost of poles and wires. Most of the data-center capacity cost in PJM is for data centers that have not been built, and the main dispute is over who carries the risk if some of them never are. The utilities and states that make large loads put money down before the grid builds for them are getting this roughly right. PJM’s emergency backstop procurement shows what happens when that decision is left until the end.

Data centers are the main cause of PJM’s price shock

PJM’s capacity auction pays power plants to be available three years out, and the cost flows into retail bills. PJM’s auction reports put the RTO-wide price for the 2024/25 delivery year at $28.92 per megawatt-day. For 2025/26 it was $269.92, a ninefold jump, with a total cost of $14.7 billion, and after that the price ran into the collar that PJM and the governors had negotiated. According to PJM’s July 14, 2026 report, 2026/27 cleared at $329.17 ($16.1 billion), 2027/28 at $333.44 ($16.4 billion), and 2028/29 at exactly $325.00, the cap, again for $16.4 billion. (PJM defines the cap in installed-capacity terms; $325.00 is its 2028/29 equivalent in the units above.)

PJM’s own report on the latest auction uses careful words, saying that the reliability requirement rose by a “1,374.5 MW increase in forecasted load, largely due to additional Large Loads.” The press release was blunter and cited “the continued trend of the addition of large data center loads to the load forecast.” The auction still came up 6,831 MW short of the reliability target, leaving a 14.7% reserve margin against a 20% target.

The independent market monitor, Monitoring Analytics, has done the attribution that PJM does not. For the 2027/28 auction held in December 2025, it found data centers responsible for $6.5 billion, or 40%, of the $16.4 billion in costs, as Utility Dive reported, and wrote that “Data center load growth is the primary reason for recent and expected capacity market conditions.” After the July auction, it put the cumulative figure across the last four auctions at $29.4 billion of $63.6 billion, or 46%.

The strongest counterargument here comes from E3, in a May 2026 whitepaper funded by the Data Center Coalition. Its own analysis of the 2024/25 and 2025/26 auctions attributes about half of the capacity price increase to load growth, and the other half to market-design changes, plant retirements and reduced accreditation of fossil plants. I take that seriously, and I don’t think it rescues the industry’s position. Half of a ninefold increase is still a large number, the monitor’s estimate is in the same range, and in PJM the load growth in question is mostly data centers by the monitor’s accounting. On the narrow PJM capacity question, the people who complain about data centers have the better of the argument.

Most of your bill is paying for weather, gas and wires

EIA’s latest Electric Power Monthly, released September 24 with data for July 2026, shows a different national picture. It puts the US average residential price at 18.31 cents per kWh, up from 17.45 cents in July 2025, an increase of 4.9%. For the first seven months of the year, the average is 18.19 cents against 17.01 cents, up 6.9%. EIA’s October Short-Term Energy Outlook expects 18.2 cents for 2026 as a whole and 18.7 cents for 2027.

PJM states rank high by year-to-date increase, with the District of Columbia up 17.8%, Maryland 13.6%, Virginia 13.1%, Ohio 13.0% and Pennsylvania 12.5%. New Hampshire is up 15.2%, New York and Washington 13.0% each, Hawaii 12.7% and Idaho 12.0%, and none of them is in PJM. Meanwhile Texas, which has the biggest data-center pipeline in the country, is up 5.7%, below the national average. Georgia, where the utility’s estimate of new data-center load rose from 400 MW in 2022 to 8,500 MW in 2025, is up 4.5%, helped by a base-rate freeze through 2028 that the state commission ordered in July 2025. (The percentages are my calculations from EIA’s state tables; the state series is on /power/.)

Lawrence Berkeley National Laboratory and Brattle have done the unglamorous work on why prices move, and their July 2026 update is the best document I know on it. Nominal residential prices rose 33% from 2019 to 2025, but only 5.6% after inflation. Distribution costs pushed prices up in 22 states between 2024 and 2025. Wildfire spending adds more than 4 cents per kWh in California, which at 33.61 cents in July is still the most expensive state in the contiguous US. The delivered price of natural gas is up 36% from its 2024 low and was a driver of year-on-year price changes in 21 states. In PJM, the authors find that capacity costs added 0.9 cents per kWh to average wholesale costs in 2025, with a further 0.6 cents expected in 2026, and they write that “Capacity prices are one but not the only contributor to retail price increases.” On the data-center question in particular, their verdict is that the “Presence of significant data center load does not appear to alter conclusions.”

The same pattern appears inside PJM this year. The market monitor found that PJM’s wholesale cost for January to July 2026 was $116.53 per MWh, up 46% from $79.57, and that data-center load accounted for $10.48 of it, or 9%, all through the capacity market. The increase was $36.96 per MWh. By my arithmetic, even if all of the data-center capacity charge were new this year (it isn’t), it would explain less than a third of that rise. EIA, for its part, forecasts that PJM wholesale prices will rise 41% in 2026, “the largest increase across all regions,” and attributes this year’s spikes mainly to weather, Winter Storm Fern and the July heat.

I concede two points against my reading. First, the monitor’s 9% counts only capacity. It excludes whatever data centers do to energy prices and transmission spending, and both effects are real even if no one has put a clean number on them. Second, the research showing that load growth lowers rates is weaker for households than its headlines suggest. LBNL finds that the states with the most load growth saw real all-sector prices fall by more than a cent per kWh, but for residential prices the authors say plainly that “The relationship between residential prices and load growth is not as clear as for all-sector average prices.” E3 makes the same point, noting that Texas and Virginia had the fastest load growth and the smallest rate increases over its study period. Virginia is up 13.1% so far in 2026, which is not small. That does not prove data centers caused it (the SCC’s November 2025 Dominion rate order alone added $11.24 a month to a typical residential bill for 2026), but it does mean the “load growth is good for rates” line has an expiry date somewhere in it, and northern Virginia might be where we find it.

The data centers you are paying for mostly do not exist yet

I think the most important number in this debate is in the monitor’s January report, and it is easy to skip past. Of the $6.5 billion in 2027/28 capacity costs it assigned to data centers, about $6.2 billion relates to data centers that have not been built yet but could come online by that delivery year. That share, roughly 95%, means that most of the data-center bill in PJM is for load that so far exists only in a forecast.

Forecasts of large loads have not been reliable. PJM’s January 2026 load forecast cut its 2028 summer peak by 4,414 MW, or 2.6%, from the previous year’s forecast, partly through stricter vetting of large loads, while still projecting peak load to reach 222,106 MW by 2036. In Ohio, after AEP’s new data-center tariff required customers to commit to paying for most of the capacity they requested, the utility cut its data-center forecast from 30 GW to 13 GW, as Utility Dive reported. Grid Strategies, which tallies utility filings nationally, counted about 90 GW of data-center demand through 2030 and judged that figure “likely overstated by roughly 25 GW.” And in Texas, ERCOT told the state Senate in July that, as of June, it was “tracking approximately 474 GW of Large Loads seeking interconnection, of which ~90% are data centers.” That one state’s queue is more than double PJM’s projected 2036 peak.

Nobody serious thinks 474 GW of data centers will be built in Texas. The queue is a list of options, many of them, I suspect, duplicates, filed by developers who are rationally applying everywhere because applying is cheap. Even so, a forecast built from those options sets real prices and real build plans, and somebody pays for them whether or not the buildings go up.

PJM’s price collar complicates this further. Its cap and floor, which FERC extended on April 28 to cover the 2028/29 and 2029/30 auctions, kept the last auction’s cost at $16.4 billion, against the $29.7 billion PJM simulates without it. I support the collar, but it also means that, at the margin, the forecast no longer moves the price. When Maryland’s People’s Counsel asked PJM in August to adjust the 2029/30 auction for new large-load information, PJM’s September 1 reply said “there will need to be substantial load reduction for the auction to clear at a price lower than the ‘collar cap’” and that the change “will not impact consumer bills due to the already implemented ‘collar cap.’” Under a binding cap, phantom load shows up in the shortfall instead of the price. For 2028/29 that shortfall was 6,831 MW, which the backstop procurement is meant to fill.

Large customers should post money first

The places that seem to me to handle this well all make the large customer put money at risk before the grid commits to serving it. None of them is perfect, and they differ a lot in detail.

AEP Ohio’s Schedule DCT, in effect since July 23, 2025, applies to new data-center load of 25 MW or more. It sets a minimum monthly billing demand of up to 85% of contract capacity, owed whether the customer uses it or not, and it sets a contract term in which “The initial term of the contract will equal the load ramp period (no greater than four years) plus eight years.” The 17 GW that vanished from AEP’s forecast after the tariff is, I think, the best evidence available of how much of a queue is real when real money is attached.

Texas took a different route when, under Senate Bill 6, the Public Utility Commission adopted its large-load rule on September 18. It applies to loads of 75 MW or more and requires a customer to post $50,000 per MW of requested peak demand before ERCOT will even study the connection, plus a $100,000 study fee. The state’s Office of Public Utility Counsel asked for $100,000 per MW. The commission declined, saying $50,000 “appropriately balances the statutory objectives of PURA §37.0561 to support business development while minimizing the potential for stranded infrastructure costs.” I would have set the amount higher, but the posted money is real.

Virginia’s regulator has moved in stages, starting with the November 2025 Dominion order, which created a new class for customers of 25 MW or more who must pay “a minimum of 85% of contracted distribution and transmission demand and 60% of generation demand,” starting January 1, 2027. On July 31, 2026, the SCC went further and directed Dominion to require a mandatory contribution in aid of construction from large loads for the transmission lines and substations that connect them. The same order shifted transmission costs among customer classes, which Utility Dive reported cut the monthly increase for a typical residential customer under Dominion’s transmission rider from $2.90 to $0.94. California joined in September with SB 886, signed on September 21, which requires the CPUC to adopt data-center tariffs by 2028, with a 10-year early-termination fee, according to Ballard Spahr’s summary. E3 counts at least 38 large-load tariffs since 2018, 30 of them adopted in 2025 and 2026.

A minimum bill protects other customers from wires and plants built for someone who leaves, but it does nothing about the market-wide effect in PJM, where a data center that faithfully pays its 85% minimum still raises the capacity price for everyone in the region. Utilities also make generous claims for the model. Georgia Power says its large-load contracts, including a 3,200 MW contract with OpenAI approved in August, will save a typical household “at least $15 per month beginning in 2029.” That’s a projection from the company selling the power, and I’ll believe it in 2029.

PJM’s backstop followed the opposite order. On January 16, 2026, the White House’s National Energy Dominance Council and the governors of all 13 PJM states issued a statement of principles calling for an emergency auction with 15-year contracts, with costs falling first on utilities that add data centers without bringing their own capacity. The PJM board announced its own backstop the same day. By July 31, PJM had filed a Reliability Backstop Procurement to buy up to 6,831 MW on contracts of up to 15 years, at a weighted-average price of up to $555 per MW-day. That price is 70% above the $325 cap on the regular auction, though it is a levelized price over a long contract and so not a like-for-like comparison. The board said “new large loads should bear the costs they cause,” and, in the same decision, conceded that “PJM does not have jurisdiction to allocate retail costs directly to individual data centers.”

On September 29, FERC accepted the procurement but suspended it “for five months to become effective February 28, 2027, subject to refund.” It approved the $555 cap and sent three issues to a paper hearing: cost allocation, the rules for transmission owners that leave PJM, and collateral for load-serving entities. All three issues concern who pays. FERC Chair Laura Swett, in a concurring statement reported by Utility Dive, said PJM’s late filing left no time “to rehabilitate the mess we received.” One co-op, Northern Virginia Electric Cooperative, said it would have had to post about $2 billion in collateral under the proposal. PJM postponed the auction that was due to open the next day.

I don’t think FERC was wrong, because a 15-year procurement whose cost allocation cannot survive review is worse than a five-month delay. Still, the episode shows the cost of building the buying mechanism first and arguing about the bill afterwards, which is the reverse of what Ohio and Texas did.

What I’m watching between now and February

EIA publishes August 2026 state prices on October 23, and I want to see whether the PJM states keep pulling away from the national average once the June capacity-year change has passed through a full summer of bills. Under the FERC order, PJM’s first brief in the paper hearing is due 45 days after September 29, which puts it in mid-November, unless PJM refiles first. The 2029/30 capacity auction closes on December 15, and given PJM’s own letter, I’d bet on another result at the cap. The backstop’s earliest effective date is February 28, 2027. In Congress, the Ratepayer Protection Act, which would make large loads of 100 MW or more pay the full incremental cost of the infrastructure that serves them, passed the House 417 to 3 on September 16 and was blocked in the Senate the next day when a senator objected to passing it by unanimous consent.

The first result that would change my mind is August and September EIA data that show Texas and Georgia rising as fast as Maryland and Ohio. That would make my claim that the bill story is mostly regional wrong, and it would mean data-center demand acts on prices through channels the capacity market does not capture. The second is a 2029/30 auction that clears well below the cap, which would mean the load forecast has been cut so hard that I have overstated the phantom-load problem. The third is a final backstop allocation that spreads the bill over all customers by usage instead of over the utilities whose data-center forecasts created the gap. The January principles would then have produced the opposite of what they promised, and the households behind those $8.79 Pepco increases would be paying for capacity bought for buildings that, as of October 2026, mostly do not exist.