Microsoft is paying about $112 per MWh for 835 MW from a reactor that shut in 2019, and it signed for 20 years. That is the Jefferies estimate for the Crane restart, published five days after the deal, and Bloomberg Intelligence put it at $100 (Bloomberg, September 25, 2024). In the same month Pennsylvania’s average industrial customer paid $91.20 per MWh delivered, and PJM’s Western Hub cleared $58.51 for the year (EIA-861M; EIA STEO, December 2025). The largest, best-rated power buyers in the country chose to pay a 70 to 90% premium to wholesale for firm, carbon-free, 20-year power. That number, not your utility’s tariff, is now the reference price for any long-term contract, and the Open Factory Large-Load Cost Table treats it that way.
We cover the eight hyperscaler nuclear agreements and what each discloses, the arithmetic that turns Talen’s $18 billion into $82 to $94 per MWh, Crane’s cost floor against its price, where the deals sit against industrial rates in 14 states and against bridge power, how much of the announced datacenter load they actually cover by 2032, and, behind the wall, what a factory negotiating 15 years should ask for and which utilities will sign.
Source: Open Factory Large-Load Cost Table (compiled from Constellation, Talen, Vistra and NextEra 8-Ks, Kairos and Amazon releases; as of December 2025)
The stack reaches 2,347 MW in 2027 when Crane, Clinton and Comanche Peak start, 5,771 MW in 2032 when Susquehanna and Comanche Peak hit full quantity, and 6,691 MW in 2035 with every SMR agreement counted at our assumed years.
Eight Agreements, 6.8 GW, One Notional Price
Between September 2024 and October 2025 five hyperscalers signed for nuclear output from six plants, and only one of the six sellers has put a dollar figure next to its contract. Constellation’s September 20, 2024 release calls the Microsoft deal “its largest-ever power purchase agreement”, 20 years, “approximately 835 megawatts”, and says nothing about price. Talen’s June 11, 2025 8-K gives 1,920 MW through 2042 and its investor deck adds “~$18bn notional revenue under 17-year contract” with “price escalators through 2042” (Talen deck, slide 3). Vistra’s September 29, 2025 8-K sells 1,200 MW from Comanche Peak for 20 years, with options for 20 more, to “a large, investment grade company” it will not name.
Source: Open Factory Large-Load Cost Table
Eight rows sum to 6,811 MW at full quantity, of which 5,691 MW is signed PPAs on existing or restarted reactors and the rest is Kairos’s 500 MW fleet, Amazon’s SMR development agreements, and Meta’s open request for 1 to 4 GW.
| Buyer / seller | Plant | MW | Term | Start | Price | Implied $/MWh |
|---|---|---|---|---|---|---|
| Microsoft / Constellation | Crane (TMI-1), PA restart | 835 | 20 yr | 2027 | n/d | 98 to 115 (analysts) |
| Amazon / Talen | Susquehanna, PA | 1,920 | to 2042 + options | ramp to 2032 | ~$18bn notional | 82 to 94 (ours) |
| Meta / Constellation | Clinton, IL | 1,121 | 20 yr | Jun 2027 | n/d | ~70 (Jefferies) |
| Undisclosed / Vistra | Comanche Peak, TX | 1,200 | 20 yr + options | Q4 2027 | n/d | n/d |
| Google, CIPCO / NextEra | Duane Arnold, IA restart | 615 | 25 yr | by Q1 2029 | n/d | n/d |
| Google / Kairos | SMR fleet | 500 | PPAs under MPDA | 2030 to 2035 | n/d | n/d |
| Amazon / X-energy, Dominion | Cascade WA, North Anna VA | 620+ | development | early 2030s | n/d | n/d |
| Meta / RFP | 1 to 4 GW new nuclear | 1,000+ | open | early 2030s | n/a | n/a |
Full rows with structure, announcement dates and sources: data/deal-book.csv. n/d = not disclosed.
The remaining rows are development deals rather than supply contracts. Google and Kairos signed a master plant development agreement on October 14, 2024 for 500 MW, first unit in 2030, complete by 2035, with PPAs “for energy, ancillary services, and environmental attributes” to follow (Kairos Power). Amazon’s October 2024 package is 320 MW with Energy Northwest, expandable to 960 MW, “beginning in the early 2030s”, plus “at least 300 megawatts” of SMR capacity at Dominion’s North Anna site (Amazon). Meta issued a request for proposals for 1 to 4 GW of new nuclear on December 3, 2024, delivery “starting in the early 2030s”, and as of this writing has announced no award (Meta). None of these carries a price, and none delivers a megawatt before 2030, which is why the Datacenter Electrical Roadmap 2026 to 2032 carries them as options, not supply.
The Susquehanna Arithmetic: $82 to $94 per MWh
Talen’s deal is the only one with a revenue number, so it is the only one where the implied price is arithmetic rather than an analyst’s guess. The contract ramps from a 300 MW behind-the-meter arrangement today to 840 to 1,200 MW in 2029 and 1,680 to 1,920 MW by 2032, then runs flat through 2042 (Utility Dive, June 11, 2025). We fill the 2026 to 2028 gap with 300, 420 and 660 MW, the 120 and 240 MW steps Talen’s deck and the original 2024 agreement describe (POWER, March 4, 2024), and run every unit at Constellation’s 2024 fleet capacity factor of 94.6% (Constellation Q4 2024 release).
Source: Open Factory Large-Load Cost Table
At the midpoint the contract delivers 25,020 MW-years, or 206 TWh, over 17 years; the low ramp gives 191 TWh and the high ramp 221 TWh.
Divide $18 billion by those volumes and Amazon is paying a flat-equivalent $87 per MWh at the midpoint, $82 on the high ramp and $94 on the low one. Every input and the division is in data/assumptions.csv and data/implied-price-susquehanna.csv. Two caveats cut in opposite directions. The deck says “notional”, which in Talen’s usage means the revenue at full contract quantity with escalators, so the year-one price is lower: a 2% annual escalator over 17 years puts the starting price about 15% below the flat equivalent, near $74. And the contract covers “carbon-free energy and capacity”, so a factory comparing it to an energy-only price should add PJM capacity, which for the 2027/28 delivery year cleared at $333.44 per MW-day, about $16 per MWh for a load running at 85% (PJM, December 17, 2025; PJM Capacity Charge Estimator).
The structure matters as much as the price. FERC rejected the original behind-the-meter expansion from 300 to 480 MW on November 1, 2024 by a 2-1 vote, with Commissioners Christie and See finding PJM “has not met its burden” and Chairman Phillips dissenting that “access to reliable electricity is the lifeblood of those data centers” (Utility Dive, November 4, 2024). The rewritten deal puts Susquehanna’s power onto the PJM grid, makes Talen the retail generation supplier and PPL Electric the wires company, needs no FERC approval, and eliminated Amazon’s option to cap at 480 MW (Talen deck, slide 4). A merchant generator acting as a licensed retail supplier under a state’s competitive-supply law is the template a Pennsylvania or Texas factory can actually copy; the Large-Load Interconnection Cost tool models both routes.
The 2024 version of the same deal shows how far the terms moved in fifteen months. Amazon paid $650 million for Talen’s Cumulus campus on March 4, 2024, a 960 MW site next to the 2.5 GW station, with a 10-year PPA, “two 10-year options to extend the agreement”, power ramping “in 120-MW increments over several years”, and “a one-time option to cap commitments to 480 MW” (POWER, March 4, 2024). The 2025 version doubled the volume, removed the cap, stretched the firm term to 17 years and attached a notional revenue figure. Every change ran in the seller’s favour, and the buyer agreed because the alternative, waiting for a grid connection in PJM, is measured in years in the Large-Load Cost Table.
Crane: $1,916 per kW and a $64 Premium
Crane is the cleanest cost case because the capital number is public. Constellation’s restart costs $1.6 billion, of which the Department of Energy is lending $1 billion at 0.375% over Treasuries, closed November 18, 2025, and the unit is now scheduled for 2027 rather than the original 2028 (Utility Dive, November 19, 2025; Constellation Q2 2025 release). Over 835 MW that is $1,916 per kW, against $2,240 per kW for a new gas reciprocating plant in EIA’s cost table and $2,670 per kW for a 2026-priced aeroderivative plant in our Datacenter MEP Cost Table.
Source: Open Factory Large-Load Cost Table
Capital recovery at 8% over the 20-year term is $23.70 per MWh at a 94% capacity factor, plant operating cost using Talen’s published Susquehanna all-in figure of under $24 as a proxy brings the floor to about $48, and the Jefferies estimate of $112 leaves a $64 premium.
The premium is not a mistake by Microsoft. Jefferies’ Paul Zimbardo wrote that the deal “should provide Constellation with about $785 million in annual revenue by 2030”, which over 6.87 TWh a year is $114 per MWh and checks against his $112 (Bloomberg via Energy Connects, September 25, 2024). The same note put regional wind and solar PPAs near $60. Microsoft paid roughly double the renewable price for three things a factory also wants: firmness, a 20-year term, and a PJM location where the Lead-Time Monitor shows large power transformers at 128 to 144 weeks and new gas turbines sold into 2029. The premium is the market value of not waiting.
Clinton at $70 and the Undisclosed Half
Meta’s Clinton deal is the cheapest of the three priced contracts because it buys something narrower. Constellation’s August 7, 2025 release describes a 20-year PPA “for the full output of the Clinton Clean Energy Center” beginning June 2027, with a 30 MW uprate taking the plant from 1,092 to 1,121 MW by 2029. It is a virtual PPA for clean energy attributes, not physical delivery, so it needed no regulator’s signature. Jefferies valued it “around $70/MWh, or a roughly $20/MWh premium to Illinois’ energy-plus-capacity market compensation from 2028 and as much as $15/MWh higher than regional wind power purchase agreements” (Utility Dive, June 12, 2025). The same article records Morgan Stanley and Jefferies putting Crane at $98 to $115.
Clinton also shows who the seller was negotiating against. The plant ran for ten years on Illinois zero-emission credits that “saved Clinton from premature closure in 2017 after years of financial losses”, and Meta’s contract picks up when the ratepayer subsidy ends (Utility Dive, June 12, 2025). A buyer arriving as a subsidy expires gets a $70 price; a buyer asking a seller to spend $1.6 billion gets $112. The gap between them is the clearest single number here for a factory: the price of long-term nuclear depends on whether you are replacing someone else’s money or supplying it.
The other half of the contracted megawatts carries no price at all. Vistra’s Comanche Peak PPA begins delivery in the fourth quarter of 2027 and ramps to 1,200 MW by 2032 (Utility Dive, September 29, 2025). NextEra’s 25-year Duane Arnold PPA with Google restarts a 615 MW plant in Iowa “no later than the first quarter of 2029 and perhaps as early as the fourth quarter of 2028”, and the Central Iowa Power Cooperative “will purchase 50 megawatts of the plant’s output on terms and conditions consistent with the Google PPA” (NextEra Q3 2025 release). That last clause is the most useful sentence a smaller buyer will find in any of these filings: a 50 MW co-op got the hyperscaler’s terms because it was already at the table as a part-owner. We record both as “not disclosed” in the Large-Load Cost Table and do not guess.
The Price Ladder: Nuclear, Retail, Bridge Power
Put the three priced deals on one axis with what a factory actually pays and the picture is uncomfortable for anyone hoping a long contract means a low price. Industrial retail averages for September 2025 from EIA’s monthly sales file run $67.50 per MWh in Texas, $87.50 in Ohio, $91.20 in Pennsylvania, $90.20 nationally and $107.50 in Illinois (EIA-861M; rows in data/industrial-rates-sep-2025.csv). Our five-year bridge-power model for a 100 MW site lands at $96.00 per MWh for solid oxide fuel cells, $104.20 for gas reciprocating engines and $112.00 for aeroderivative turbines at $3.78 gas (Recips vs Turbines vs Fuel Cells; Datacenter MEP Cost Table).
Source: Open Factory Large-Load Cost Table (compiled from Bloomberg, Jefferies via Utility Dive, Talen, EIA-861M, EIA STEO December 2025, Open Factory Datacenter MEP Cost Table; as of December 2025)
Crane at $98 to $115 brackets Illinois retail; Susquehanna at $82 to $94 straddles Pennsylvania and Ohio retail; Clinton at $70 sits just above Texas retail and $11 above PJM West’s 2025 average of $58.51.
Three adjustments keep the comparison honest. The retail rates are delivered, and include distribution, transmission and capacity; the PPA figures are energy, or energy and capacity in Talen’s case, at the plant. A Pennsylvania factory paying $91.20 delivered is paying roughly $60 for energy and $30 for everything else, so Susquehanna’s $87 is not a 5% premium to its bill, it is a 45% premium to the energy line. The bridge-power figures are all-in at the site but exclude the grid connection the site still needs for backup and for the day the bridge ends (Behind the Meter, In Front of the Queue). And the wholesale reference is moving: EIA’s December outlook has PJM West at $40.75 in 2024, $58.51 in 2025 and $63.69 in 2026, a 56% rise in two years (EIA STEO, December 2025, Table 7a).
The decomposition is the part most factory bills hide. On a Pennsylvania industrial invoice near $91 per MWh, roughly $60 is energy at a price that follows PJM West, about $16 is capacity at the 2027/28 clearing price of $333.44 per MW-day for a load running 85% of the hours, and the balance is distribution, transmission and riders (PJM, December 17, 2025; split is our estimate in data/assumptions.csv). Only the first two lines are what a nuclear PPA replaces, and only the first is what a retail supplier’s fixed-price offer replaces. A factory that compares a $75 energy-plus-capacity quote to a $91 delivered bill and concludes it is saving 17% has compared two different products; the PJM Capacity Charge Estimator exists to separate them before the comparison is made.
Source: EIA-861M via Open Factory Large-Load Cost Table
Twelve jurisdictions span $60.60 in Louisiana to $107.50 in Illinois, a 1.8x spread, before New Jersey’s $139.40 and California’s $254.70 are even on the chart.
PJM States Rose 16 to 25% in a Year
The year-over-year move is why the 15-year question is being asked at all. From September 2024 to September 2025, Ohio’s average industrial rate went from 7.01 to 8.75 cents per kWh, up 24.8%; Pennsylvania from 7.87 to 9.12, up 15.9%; Illinois from 9.36 to 10.75, up 14.9%; Virginia from 8.96 to 9.69, up 8.1% (EIA-861M). Texas rose 10.7% to 6.75 cents and Georgia 4.8% to 7.47. The states with the largest increases are the states whose capacity price went from $28.92 to $269.92 per MW-day in one auction, a mechanism we walk through bill by bill in A Factory Owner’s Guide to the PJM Capacity Charge.
The mechanism is not a mystery. PJM’s capacity price for the rest of the region went from $28.92 per MW-day for 2024/25 to $269.92 for 2025/26, then to the $329.17 cap for 2026/27 and $333.44 for 2027/28, and the December 17, 2025 auction that set the last of those cleared 6,623 MW short of its reliability requirement with about 5,100 MW of the year’s forecast growth attributed to data centers (PJM, December 17, 2025). At 85% load factor $333.44 per MW-day is $16 per MWh; at a 50% load factor it is $27. An industrial customer with a poor load factor in Ohio or Pennsylvania has absorbed most of that in the twelve months on the chart, and the PJM Capacity Charge Estimator shows what the next two delivery years add.
Source: EIA-861M via Open Factory Large-Load Cost Table
Ohio gained 1.74 cents, Pennsylvania 1.25, Illinois 1.39 and Virginia 0.73 in twelve months, while North Carolina fell 0.19 and Georgia gained 0.34.
A PJM factory that signed a five-year fixed price in September 2024 at that month’s rate is 16 to 25% ahead today, and the hyperscalers locked their number for 17 to 25 years. That is the asymmetry a factory should carry into any negotiation: the counterparties that know the most about where power prices are going have all chosen very long fixed-price contracts at levels above today’s retail. The Large-Load Cost Table records the term each utility now demands of large loads, 12 years at AEP Ohio and Indiana Michigan Power, 14 at Dominion, up to 15 in Georgia; the utilities are asking for the same thing the hyperscalers asked for, only with the price floating.
5.8 GW Against 90: The Forecast
Now the demand side of the first chart. Lawrence Berkeley National Laboratory’s December 2024 report puts US data center consumption at 176 TWh in 2023, 4.4% of the country’s electricity, and at 325 to 580 TWh in 2028, 6.7% to 12.0% (LBNL, p. 5). As average load that is 37 to 66 GW. Grid Strategies’ November 18, 2025 report attributes “about 90 GW, or roughly 55%” of a 166 GW five-year peak-load growth forecast to data centers by 2030, a forecast that was 38 GW in the 2023 edition (Latitude Media, November 18, 2025). BloombergNEF’s December 1 forecast reaches 106 GW of data center demand by 2035, 36% above its April number (Utility Dive, December 3, 2025).
Source: Open Factory Large-Load Cost Table, LBNL (December 2024), Grid Strategies via Latitude Media (November 2025), BloombergNEF (December 2025)
Contracted hyperscaler nuclear is 2.8 GW in 2028 against 37 to 66 GW of demand, 4.6 GW in 2030 against 90 GW of datacenter peak growth, and 6.7 GW in 2035 against 106 GW.
Our forecast: by 2032, hyperscalers will have 5.8 GW of nuclear under contract, 5.6 GW of it from reactors that were already built, and it will cover between 6% and 9% of the datacenter load every published forecast expects by then. The other 91 to 94% comes from the grid as it stands, from the gas turbines and engines whose slots are sold through 2029 (The $60 Million Deposit; Cummins and Caterpillar Are Sold Out Through 2028), and from fuel cells that still carry a 30% investment tax credit (The 30% Fuel-Cell ITC). Nuclear’s contribution before 2035 is capped by a simple count: the US has two shut plants with credible restarts (Crane and Duane Arnold, 1,450 MW between them, both now contracted), one more restarting for two co-ops with $2.8 billion of federal loans and grants (Palisades, Utility Dive, October 1, 2024), and a merchant fleet whose uncontracted output is being sold plant by plant. Every reactor with a hyperscaler contract in 2032 has already been named.
Three consequences follow for a factory. First, the nuclear PPA is not a template a 50 MW buyer can sign; the smallest counterparty in the eight rows is the 50 MW co-op that already owned part of the plant. Second, the price it sets is a ceiling that utilities will use, not a floor a factory can demand: a regulated utility that can sell firm power to a hyperscaler at $87 to $112 through a special contract has no reason to write a factory a 15-year fixed price at $70. Third, the megawatts not covered by nuclear are the megawatts the Lead-Time Monitor and the Lead-Time-Adjusted Schedule exist to track, because they will be built from transformers, switchgear, turbines and engines, and every one of those queues is where a factory’s order sits behind a datacenter’s.
The one firm, carbon-free source that is being contracted at sub-gigawatt scale is geothermal, and it is following the nuclear price logic rather than undercutting it. Fervo’s Cape Station in Utah plans 100 MW of firm power online in 2026 and 500 MW in total, and the company raised $462 million on December 10, 2025 with Google in the round, bringing its total to $1.5 billion since 2017 (Latitude Media, December 10, 2025). Its first offtake structure with a utility, NV Energy’s Clean Transition Tariff approved in May 2025, delivers 115 MW to Google at a premium that Google alone pays (Latitude Media, May 15, 2025). No Cape Station price is public. A 100 MW first phase for a single buyer’s premium tariff is the same pattern as the reactor deals one order of magnitude down, and it is the pattern we expect the Large-Load Policy Tracker to record in Utah, Nevada and Texas before any SMR delivers a megawatt-hour.
What to Do Monday
Price your energy line, not your bill. Take twelve months of invoices, strip distribution and transmission, and add the capacity charge the PJM Capacity Charge Estimator computes from your peak-load contribution; the result for a Pennsylvania plant is roughly $75 per MWh for energy plus capacity, which is the number to hold against Susquehanna’s $87 and Clinton’s $70. Then ask your utility or retail supplier for three quotes on the same day: three-year fixed, seven-year fixed, and the longest fixed term they will write. The spread between them is the utility’s own forecast, and it is the one document in the negotiation you will not have to pay for.
If the long quote comes back above $95 per MWh delivered in a PJM state, price the alternative before signing: a 100 MW bridge plant in our model runs $96 to $112 per MWh for five years and lands 20 to 30 months after the purchase order (Genset TCO; Datacenter MEP Cost Table), and a factory that can shed 20% of load at PJM’s five peak hours cuts its capacity line by the same fraction. Behind the paywall: the eight utilities and two merchant generators that have written long-term large-load contracts as of December 2025, with the threshold, term, minimum bill and whether the price is fixed for each, and the arithmetic showing that at 50 MW the gap between Meta’s price and Microsoft’s is $15.6 million a year, from the Open Factory Large-Load Cost Table.