A data center's electricity bill has three lines, and almost every public comparison of "power costs" quotes only the first one. The first line is energy: kilowatt-hours times a $/MWh rate. The second is capacity (in organized markets) or demand charges (in vertically integrated ones): dollars per kilowatt or megawatt of reserved capability per month, payable whether or not a single electron flows. The third is transmission and the network upgrades your interconnection requires. Capacity is the line that behaves least like a commodity and most like an insurance premium — you are paying for the right to draw a megawatt that someone else keeps ready. In 2026 and 2027, that premium was repriced, and the right was converted from a monthly charge into a multi-year contract with collateral attached. This piece walks both moves with the operators' own numbers, then does the arithmetic the tariff pages leave to the reader.
The market layer is PJM, the largest wholesale electricity market in the United States, covering 13 states and the District of Columbia and about 67 million people by PJM's own count. Its capacity auctions are the clearest public price for "a megawatt standing by" anywhere in the world, and they are published in full. The contract layer is the state utility commission. In 2025 and 2026, both layers moved in the same direction at the same time, which is why the second page of the power bill is now the page that decides whether a project pencils.
How much did four auctions charge for the same megawatt?
PJM's own news release dated 2025-12-17 reports the 2027/2028 Base Residual Auction results: 134,479 MW of unforced capacity and demand response cleared, 145,777 MW including Fixed Resource Requirement resources, at $333.44 per MW-day (UCAP) for the entire footprint — the FERC-approved cap, and a 1.3% increase over the 2026/2027 record. Cleared supply times clearing price totals $16.4 billion, though PJM notes not all load pays it because of self-supply and bilateral contracts. The same release reports the outcome that matters more than the price: committed supply came in 6,623 MW short of the reliability requirement, leaving a 14.8% reserve margin against a one-in-ten standard that implies roughly 20%.
The published sequence, cross-checked between the Ohio Consumers' Counsel's case summary and Calpine Energy Solutions' client note, both of which tabulate the same four auctions:
| Delivery year | Clearing price ($/MW-day) | Note |
|---|---|---|
| 2024/2025 | $28.92 | the last auction before data-center load showed up in the forecast |
| 2025/2026 | $269.92 | record; $466.35 in the BGE zone and $444.26 in the Dominion zone |
| 2026/2027 | $329.17 | cleared at the cap set by the PJM–Pennsylvania agreement |
| 2027/2028 | $333.44 | cap again; PJM's uncapped RTO price would have been $529.80 |
Three things are worth extracting from that table. First, the multiple: 11.5× in three auction cycles, on a product whose physical definition did not change. Second, the cap: the agreement between PJM and Pennsylvania set a $325/MW-day cap and a $175/MW-day floor for the 2026/2027 and 2027/2028 auctions, and modelling details pushed the final clearing price slightly above the agreed cap; the $529.80 uncapped figure is the market's own answer, 59% above the printed one. The cap, in other words, is not a ceiling the market bumped into — it is a policy number that is currently setting the price, and by PJM's account it applies to these two delivery years only. Third, the shortfall: the highest capacity price in the market's history bought 774 MW of new generation and uprates, against a forecast peak that rose 5,250 MW for the delivery year, of which ~5,100 MW is data centers (PJM's release; the 774 MW figure as reported by Power Engineering's reading of the auction report). Price did its job on the demand side and could not do its job on the supply side within the auction horizon.
What does $333.44 per MW-day cost per MWh you actually use?
This is the conversion almost nobody performs, and it is the whole argument. Capacity is billed per reserved megawatt per day; a data center budgets per megawatt-hour delivered. Multiply by 365 and you get $121,706 per MW-year — equivalently $10,142 per kW-month, which is the unit colocation landlords quote, and that coincidence is useful later. Divide that annual figure by the hours you actually run, and the capacity charge per delivered MWh depends entirely on your own load factor (all figures below are our arithmetic on PJM's published clearing price):
| Load factor | Hours per year | Capacity cost per MWh delivered |
|---|---|---|
| 90% | 7,884 | $15.44 |
| 75% | 6,570 | $18.52 |
| 60% | 5,256 | $23.16 |
| 30% | 2,628 | $46.31 |
Read the table as a statement about workloads, not about markets. The same megawatt, in the same zone, in the same auction, costs one MWh-buyer $15 and another $46 — a 3× spread created entirely by utilization. At any energy price between $40 and $70/MWh, capacity is therefore somewhere between 18% and 28% of the delivered cost at a 90% load factor, and between 40% and 54% at 30%. For scale: the Ohio Consumers' Counsel, summarizing the same auction results for residential customers, puts capacity at "roughly 20 percent" of the supply price ComEd customers pay. A hyperscale inference hall running flat-out and a training cluster that bursts for two weeks a quarter are not in the same cost regime, even when they sign the same tariff.
There is a second, less comfortable reading. $121,706 per MW-year is 1.1% of the $10.7M per MW that JLL's 2026 outlook says a new data center costs to build (our division). A capacity payment that has risen 11.5× is still a rounding error against the capital stack of the building it serves — which is precisely why the price could rise 11.5× without stopping a single project. What stops projects is not the price of capacity. It is the contract that now comes with it.
What does the 85% take-or-pay floor actually cost?
Ohio got there first. The Public Utilities Commission of Ohio adopted the AEP Ohio data-center tariff on 2025-07-09 in case 24-0508-EL-ATA, and the Ohio Consumers' Counsel's own summary of the order is unusually plain: the tariff "requires new, large data centers to pay for at least 85 percent of the electricity capacity they contract for, for up to 12 years, even if they use less." Data-center industry groups appealed to the Supreme Court of Ohio in November 2025; the consumer advocate intervened to defend the settlement. AEP's schedule covers new load above 25 MW, with affiliates aggregated, and pairs the floor with a ramp period and an exit fee — the shape the whole industry is converging on.
Virginia wrote the same idea in longer form. Dominion Energy Virginia's GS-5 class, approved by the State Corporation Commission in November 2025 (final order 2025-11-25 in case PUR-2025-00058) and effective 2027-01-01, applies to customers of 25 MW or more at a high load factor and imposes 14-year contract terms, minimum demand charges covering 85% of contracted transmission and distribution demand and 60% of generation demand, and collateral of $1.5 million per MW — figures that appear identically in EPRI's regulatory digest citing the SCC order and in Dentons' August 2026 client alert, which adds that roughly 131 data centers, about 30% of Dominion's data-center accounts, fall into the class.
Now price the floor. Take a 100 MW subscription that actually runs at a 30% load factor — a plausible shape for a training-first campus in its first two years. The floor bills capacity on 85 MW while the facility consumes the equivalent of 30 MW around the clock: 85 × $121,706 = $10.35M a year of capacity charges spread over 262,800 MWh, or $39.36 per MWh consumed (ours). Had the same 100 MW run at 90%, it would have paid $12.17M over 788,400 MWh — $15.44. The floor does not make power expensive in the abstract. It makes underutilization expensive, and it prices that risk at signing rather than at the meter.
What does a $1.5M-per-MW deposit cost?
Collateral is the part of the tariff that never appears in a $/MWh comparison and always appears in a board memo. At $1.5M per MW, a 500 MW campus posts $750M before any credit adjustment (ours; the multiplication is the whole point). Against a build cost of 500 MW × $10.7M/MW = $5.35B, that is 14% of the project, committed for up to 14 years and released, per the order as reported, one-fourteenth a year. At a 6% cost of capital the carry on that deposit is about $45M a year — against an annual capacity charge of 500 × $121,706 = $60.9M. In other words, the financing cost of the deposit alone is roughly three-quarters of the capacity bill it is securing. Nobody's spreadsheet has that column, because it does not come from the utility; it comes from the treasurer.
This is the actual content of the 2026 policy shift, and it is not really about price. The floor and the collateral move the risk of a stranded substation off the utility's other ratepayers and onto the data center's balance sheet. Whether that is fair is a political question; whether it is expensive is arithmetic, and the arithmetic says the deposit is not a rounding error — it is the second-largest line item after the rent.
Why is flexibility now the product?
Because the regulators have started pricing the alternative. On 2026-06-18, FERC issued tailored show-cause orders under section 206 of the Federal Power Act to all six of its regional grid operators — PJM, MISO, SPP, CAISO, ISO-NE and NYISO — giving each 60 days to justify or reform its large-load rules, and naming five categories of reform. Two of the five are the interesting ones: "accommodating co-location agreements and behind-the-meter generation," and "providing new transmission services for flexible large loads." Commissioner Rosner's accompanying remarks spell out the logic: customers "willing and able to limit withdrawals from the grid" can reduce the network upgrades and generating capacity needed to serve them, which speeds interconnection and lowers costs. The same order set "escalating readiness requirements" to deter speculative requests — the queue was being clogged by projects that had not decided to be projects.
Put that beside the capacity auction and the contract layer and a coherent price signal appears. A megawatt you can interrupt is worth more to the system than a megawatt you cannot, because it needs less standby capacity and fewer upgrades. Ohio's tariff charges you for 85% of what you reserve; PJM charges the whole system $333.44 per MW-day for capacity that mostly sits idle; FERC is now directing the markets to create a service that rewards loads willing to go away for a while. Our read (analysis, not a finding): from 2027 the cleanest way to cut a data center's power cost is not a cheaper state or a smarter PPA — it is a load shape you can sell back. That is the same conclusion our behind-the-meter battery piece reached from the other direction: at Colossus 2 the battery is not a UPS, it is a machine for moving load, and it exists because flexibility is worth more than ride-through.
Who ends up paying if the load doesn't show?
The short answer: whoever signed, which is the change. Under the pre-2024 model, a utility built to a forecast and recovered the cost from all ratepayers; a project that evaporated left a substation on everyone's bill. Under the 2025–2027 tariff model, the 85% floor, the 14-year term, the exit fee and the collateral all point the same way — the signatory carries it. That is why the policy fights are so loud and so local: the Ohio Consumers' Counsel defended the settlement on exactly this ground, and data-center groups appealed it on exactly this ground.
What the tariffs cannot fix is the reliability side, and PJM says so itself in the same release that announced the record price: capacity "accounts for a fraction of wholesale electricity costs," the auction cleared short of the reliability requirement, and the mitigating factors listed — lower forecast peak, retirements not happening, winter-only resources — are all things that might happen rather than things that were bought. Price rationed demand; it did not summon supply. For a buyer, that is a scheduling risk, not a rate risk: the megawatt you contracted may be paid for and still not be there on the afternoon it matters.
What are the caveats?
Every figure above is a published number from an operator, a regulator or a commission order, checked 2026-11-11, and each one is jurisdiction- and date-specific. PJM's $333.44 applies to the 2027/2028 delivery year and to UCAP accounting; your zonal price can differ, as the 2025/2026 auction showed when BGE cleared at $466.35 and Dominion at $444.26 while the RTO cleared at $269.92. The AEP and Dominion terms are the headline provisions of orders whose details — ramp schedules, exit-fee formulas, credit thresholds, legacy-customer exemptions — decide the actual bill, and both are under appeal or subject to further proceedings. The uncapped $529.80 is a modelling counterfactual published by PJM and relayed by a market participant, not a price anyone paid. The load-factor table and the $750M/$45M collateral arithmetic are ours, built on those published inputs with the assumptions stated in line; a different cost of capital moves the carry proportionally. Finally, we did not print the several widely circulated estimates of aggregate rate impact (per-household or per-state) because we could not verify them in two independent primary sources — the rule that governs every number on this forum. Capacity is a real bill, it is now a contract, and it is measurable. The rest is politics, and we would rather show you the arithmetic than the argument.
Comments