The Demolition Is Priced. The Carbon Is a Study.

Two weeks ago this site compared how five companies cover the emissions from their electricity. Buried in its methodology was a guess:

operators putting turbines behind the meter would move emissions out of the purchased-power line entirely.

Two of those five now have dedicated gas running or planned, and a third large campus does too. The guess was half right, and the half it got wrong turns out to be the useful one.

Who will operate the machine

These are not three campuses chosen to represent the industry. They are the three cases in this set where a public filing establishes dedicated generation.

Two rows are not gas and settle nothing. In the three that are, the buyer is not the party that will run the plant.

Dedicated generation is still the exception. Most large data center projects connect to the grid rather than build a plant for the campus alone. Northern Virginia, the largest concentration anywhere, runs on the existing fleet and the capacity market. Texas is the same story even though two of these campuses sit there: ERCOT is tracking about 474 GW of large load waiting to connect to the grid, more than fifty times the roughly 8.3 GW of generation in this brief. Dedicated plants turn up where interconnection is slow and one tenant is big enough to justify its own. Three campuses is what this brief examined, not what the industry has settled on. All three are in the United States.

Nothing in the Louisiana filing says the structure was required. What it shows is a customer taking utility service, and once that choice is made in a vertically integrated state, the utility owns the steel. Abilene shows the other path: generation on the customer's own campus, owned by an entity at the site.

What the contract is exact about

Louisiana is the only one of the three where the whole structure is public, because a utility has to ask permission and a private developer does not.

Entergy Louisiana has applied to certify seven new 754 MW combined-cycle units, four near the customer's site in Richland Parish and three near Big Cajun, plus 400 MW of battery storage. The customer is named: Evest LLC, "a subsidiary of Meta Platforms, Inc." and an affiliate of Laidley LLC, whose adjacent project was an earlier docket.

Here is what the customer signed. A twenty-year initial term with automatic five-year renewals and three years of notice to leave. Minimum monthly charges "sized to cover the full incremental cost to serve," trued up against what the plants actually cost. On early termination, all unrecovered costs, or the net book value of the assets once they are in service. A guaranty from Meta Platforms, Inc.

And the row nobody writes about. If a plant is retired and cannot be sold, "Company decommissions at Customer's cost." If it sells within twenty-four months, the customer gets the proceeds. The minimum charges keep running through liquidation for up to thirty-six months.

The utility owns the plants. The customer pays to build them, pays if they stop being needed, and pays to take them down. Every dollar of that is assigned to somebody by name.

That precision is not the utility driving a hard bargain. A plant built for one customer raises the question of whose rates it lands in, and a commission reviews the arrangement with that question in front of it. The money terms are exact because somebody with authority was looking at them.

Which is the useful way to read what comes next. On carbon, the same filing says the utility is not asking the commission to certify carbon capture in this proceeding.

What the same contract is vague about

Before that half, the size of the thing. Seven units at 754 MW is 5,278 MW. Run at a 60 percent capacity factor, plants that size emit roughly 9 million tonnes of carbon dioxide a year. Louisiana's entire electric power sector emitted 46 million tonnes in 2022. So the generation proposed for one customer is about a fifth the size of everything the state's power sector already puts out.

Now read Section D of that agreement, "Designated Low-Carbon Option Resources."

The utility will use "commercially reasonable efforts" to procure resources that help the customer meet its mitigation goal. The stated priority is to "identify viable CCS opportunities at generators being constructed to serve Customer's anticipated load." The customer funds a study to evaluate capture at one of those generators, in 2026 or 2027. The cap on what it will fund is redacted. And the utility states that it is not asking the commission to certify carbon capture in this proceeding.

So the buyer is paying to study a retrofit on an asset it will never own, against a commercially reasonable efforts promise, for an amount nobody outside the deal can see.

That is further than most contracts go. It is also a long way from an obligation, and the contrast with the payment terms is the point. The payment obligations are quantified and enforceable, down to the demolition bill. The capture obligation is commercially reasonable efforts and a funded study.

Why the asymmetry decides the outcome

Buying power from a supplier is the oldest arrangement in the industry, and nothing here is unusual about that. Airlines lease aircraft. Apple does not own Foxconn's plants. Exact money terms sitting next to soft sustainability clauses is how supply contracts have always read.

Three things make this one different. Plenty of contracts push asset risk onto a buyer: tolling deals, take-or-pay, project finance. What is unusual here is how far the economic obligation travels without the decision right travelling with it. The buyer pays the full incremental cost, guarantees the asset and funds its demolition, and the supplier retains the plant-level decision rights unless the buyer contracts for them.

The plant is also dedicated, sized to one load and backed by one parent guaranty, so there is exactly one customer whose position could have bought the obligation. And the thing left as an effort is the buyer's own public promise. A supplier's emissions are rarely the buyer's headline commitment. These companies made the electricity itself the promise, with dates attached.

What it would take to close that gap is the last thing worth knowing. Direct emissions only fall when the stack changes. Burn less by running a combined cycle instead of a simple cycle. Burn something else that physically arrives at the plant. Or capture the carbon dioxide and store it. Nothing bought elsewhere does it. A certificate changes a column without changing what comes out of the stack.

All three sit with whoever owns the plant, unless the buyer has bought a say in them. A buyer cannot retrofit a machine it does not hold. It can only write the right into the contract, and in the one contract anyone outside can read, what it bought was a study.

Everything here belongs to somebody. Each plant has an owner. Each tonne gets reported, by whoever runs the plant. Each bill gets paid, by a named company with a guarantee behind it. The promises about the money are exact and enforceable. The promise about the carbon is commercially reasonable efforts and a funded study.

Meta reported 47,468 tonnes of Scope 1 for 2024, against the roughly 9 million above. That is not a discrepancy and reading it as one would be a mistake. A company that operates almost no combustion reports almost no Scope 1, which is exactly right under the rules. It is only the measure of what the ownership question moves.

One row already did it differently. Google's offtake from Broadwing is a gas plant built with capture rather than retrofitted for it. It is not campus-dedicated and it does not change that row's answer. But it is the one case here where a buyer put the abatement into the plant at the start instead of funding a study to consider it later.

None of this reads as anyone avoiding cost. Meta is funding plants it will never own at full incremental cost, plus a nuclear uprate through those same minimum charges. Microsoft stopped buying unbundled certificates because its own report called them non-additional, which made its reported number worse on purpose. Google runs the hardest matching standard anyone has adopted. And an oil major building a gas plant in a basin where it already holds subsurface assets is at least as plausibly positioning for capture. No document read here says which. This brief does not guess.

What changes

The economics of these deals are settled and written down. The abatement is not, and it sits with a party that carries no cost if it never happens.

So the question for anyone underwriting one of these, or negotiating the next one, is not what the power costs. It is what the owner is actually obliged to do about the emissions, and whether "commercially reasonable efforts" is the strongest word available.

Louisiana at least puts the answer in public, and only because a utility had to ask a commission for permission. The private versions are not simpler. They are just not filed anywhere.

That is not a gap in anyone's disclosure. It is what the disclosure is for. Oracle's latest annual report gives $260 billion of additional data center lease commitments on terms of fifteen to nineteen years, which is what a lease note is meant to give. The power behind those commitments is not in it, and would not be expected to be. The same is true across all five buyers. The generation shows up in the seller's filings or the regulator's docket. It does not show up in the buyer's.

Methodology and sources

How these were selected, and what that leaves out. The five companies are the set compared in the earlier carbon brief, chosen there because they publish electricity and emissions data rather than because they represent the buildout. The three campuses are the ones where a public document establishes dedicated generation: a Louisiana commission filing, a line in Chevron's 10-Q, and Texas air permits. A campus whose arrangement is private and unfiled does not appear here at all. And the only contract readable in full is a regulated one, because a utility has to ask permission and a private developer does not, so it may not resemble the private versions. None of this is a sample.

Louisiana. Entergy Louisiana, LLC, Direct Testimony of Laura K. Beauchamp, Public Redacted Version, March 2026, filed with the Louisiana Public Service Commission; the docket number is blank on the copy read here. Customer identity and the relationship to the earlier Laidley docket appear in the overview of the customer's project. The seven 754 MW units are the Proposed Generators in the project resources table and are not approved. Term, minimum monthly charges, termination provisions, stranded generator language and guaranty appear in the "Summary of Certain ESA Protections" table at page 18. The Waterford 3 uprate appears at page 99. The CCS commitments, the commercially reasonable efforts language, the engineering study, the redacted cap and the statement that certification is not sought appear at page 97. The filing carries 113 redactions.

West Texas. Chevron Corporation, Form 10-Q for the quarterly period ended June 30, 2026, accession 0000093410-26-000167: "Signed an agreement to develop a power facility in West Texas designed to provide approximately 2.67 gigawatts of behind-the-meter dedicated electricity capacity to Microsoft under a 20-year power purchase agreement." The project name, developing subsidiary, turbine suppliers and investment decision timing are company statements from Chevron's newsroom, not the filing.

Scale context. The PJM figures are from PJM's Largest Data Center Market Is Relying on the Existing Fleet, Not New Gas, sourced there. The ERCOT figure is ERCOT's own: "ERCOT is tracking approximately 474 GW of Large Loads seeking interconnection, of which ~90% are data centers," from its Senate panel presentation of July 29, 2026, on data as of June 2026. The 8.3 GW comparison adds the seven proposed Louisiana units, the West Texas capacity in Chevron's 10-Q and the Abilene figure, which is secondary. It sets proposed and planned generation against requested load and is an order-of-magnitude comparison, not a like for like.

Abilene. TCEQ, Statement of Basis of the Federal Operating Permit, Longhorn Data Center, permit O4721, prepared October 15, 2025, and the TCEQ Air NSR Change of Ownership for the same site, authorizations 177262 and 177263, primary business "ELECTRICITY GENERATION FOR DATA CENTER," owner operator Abilene Data Center Campus Master Association. No filing read here names a tenant.

California. No document read here establishes a specific generation counterparty for Google's campus load, which is why that cell names none. The Broadwing offtake, Google's hourly matching approach and Meta's 2024 Scope 1 of 47,468 tonnes are from the carbon instrument brief, sourced there.

Susquehanna. Talen Energy Corporation, Form 10-K for fiscal 2025: 1,920 MW of carbon-free nuclear through 2042 for the adjacent AWS data campus, transitioning to a front-of-the-meter arrangement.

Oracle. Oracle Corporation, Form 10-K for fiscal 2026. The absence of the words megawatt, generator and natural gas is from a full-text search of that filing.

Scope treatment. GHG Protocol Scope 2 Guidance, section 6.11.2: where certificates are issued "the certificates themselves serve as the emission factor for the market-based method"; where they are not issued for the technology or jurisdiction, a contract "may nevertheless convey generation attributes," and where it is silent, "the contract for power can be used as a proxy for delivery of attributes." So a dedicated gas contract can arrive in the market-based column near the plant's own rate rather than at zero. Scope 1 and 2 boundaries follow owned or controlled operations under each company's chosen consolidation approach. No buyer's contract, control analysis or accounting conclusion was reviewed here, and none of these classifications is settled from outside.

State comparison. Louisiana electric power sector CO2 of 46.46 million metric tons for 2022 is from EIA, State Energy-Related Carbon Dioxide Emissions, electric power sector table, the most recent year in that series. The roughly 9 million tonnes is computed here at an assumed 60 percent capacity factor and appears in no source. The comparison states relative size. It is not a forecast of what state emissions would become, since new generation may displace existing output.

Intensity figures. Carbon dioxide per megawatt-hour is heat rate multiplied by 53.06 kilograms per million Btu, the EPA emission factor for natural gas. Two combined-cycle heat rates exist in this corpus and they come from different EIA publications: 6,226 Btu/kWh, a 1x1 H-class unit from the AEO 2025 capital cost estimates, giving 0.330 tonnes per megawatt-hour; and 7,239 Btu/kWh, EIA's generic new combined cycle from Assumptions to the AEO 2026, giving 0.384, the figure used in the earlier carbon brief. Neither is a with-capture rate. Simple cycle at 9,447 Btu/kWh gives about 0.50. Grid factors of 0.313 to 0.347, the instrument prices, each company's stated approach and the Broadwing offtake are from the carbon instrument brief, sourced there.

Excluded pending verification. 45Q credit values, Class VI permitting primacy, and the treatment of book-and-claim renewable gas under the GHG Protocol. Differentiated-gas certificates exist for upstream methane claims; no buyer here is shown using one.

Not established. Who takes title to the West Texas plant. Whether Laidley LLC is itself a Meta entity rather than an affiliate of one. The redacted magnitudes. Whether any buyer treats any of these arrangements as a lease. How any buyer reports these megawatt-hours under the market-based method. No executive is named and no motive is attributed.