When Compute Commitments Become Claims

Electricity Demand, Customer Credits and Creditor Timing

8 September 2026

ABSTRACT

When does a commitment to compute infrastructure protect a payment, and when does it leave demand risk unresolved? AEP Ohio's current tariff supplies one precise result: its post-ramp capacity component cannot exceed 85% of aggregated affiliated new-load contract capacity and reaches that ceiling from approximately 116.7 MW upwards. The result applies to one billing component, not electricity use or forecast error. Only 43.3% of the same study cohort's requested capacity later reached signed contracts. An executed $3.3 billion project indenture supplies a second finding: principal timing changes with rent commencement, revenue-credit exhaustion and qualifying tenant replacement, while interest remains separately payable. Its initial annual amortisation rate represents $379.5 million of principal. AEP's current implementation guidance also allows conditional reassignment of up to 25% of a customer's capacity, releasing future obligations on the assigned portion. The two cases allow adaptation with different consequences for creditors, customers and other ratepayers; they do not support a simple ranking of flexibility. Accounting-life changes, including Amazon's 2025 shortening, and two public instruments provide bounded contrasts. Together, the evidence reconstructs ordinary infrastructure mechanisms applied to compute investment. It does not establish AI-specific rent, manipulation or an industry-wide causal effect.

FROM A REQUEST TO A CLAIM

In September 2025, AEP Ohio used 13.0 GW of paid data-centre study requests to construct an 11.1 GW demand forecast. The adjustment was 85%, drawn from its tariff, and the presentation expected the first contracts in November. By February 2026, the utility reported that 5,642 MW out of 13,022.7 MW of study requests had progressed to signed contracts. The tariff factor, study request and signed commitment mark different stages of the same anticipated expansion. (AEP Ohio 2025 [aeppjm2025]; AEP Ohio 2026 [aepupdate2026])

That progression matters to anyone expecting to capture an AI dividend. A scarce site may command an attractive contract, but an investor still needs to know what remains after delivery, maintenance, creditors and contractual promises. The public needs a different answer: did the project create benefits larger than the resources it displaced? A willingness to reserve capacity begins both investigations. It does not finish either.

Announced investment, contracted megawatts, chip purchases, lease commitments and reported margins each describe part of the expansion. They are not successive names for the same thing. Moving from a capacity announcement to profit requires separate evidence about availability, use, payments and costs. Establishing economic rent also requires the opportunity cost of the resources employed. A profitable intermediary can sit inside an investment chain that disappoints elsewhere.

The question is which complementary assets and contractual conditions turn compute deployment into collectable returns, and where the remaining obligations reside. A construction advance, a conditional capacity purchase, project notes, an electricity tariff, selected accounting changes and two forms of public participation all bear on it. These deliberately different mechanisms reveal what remains payable when hoped-for use does not arrive. Their amounts are not additive, and the cases do not constitute an industry sample.

Two results emerge. The tariff's tiered capacity component reaches its 85% ceiling only from approximately 116.7 MW of aggregated affiliated new load upwards. The financing, meanwhile, defers principal after defined tenant termination and sets a deadline tied to qualifying replacement. AEP's current guidance permits a conditional capacity assignment that releases future customer obligations on the assigned portion. Both mechanisms accommodate change while protecting different claims: noteholders face a timing adjustment, while the utility's assignment conditions protect other ratepayers from stranded costs. Contract length alone therefore identifies neither customer rigidity nor creditor protection. Those properties can be tested against the disclosed terms; the documents do not identify their net effect on investment or competition.

The mechanisms are familiar from other infrastructure sectors. AI and high-performance computing describe the assets' proposed use; they do not make the contracts novel or mean that all tariff load consists of AI hardware. What matters is how the instruments interact: one party's protected receipt can remain another party's conditional payment. Public benefits require a further comparison with the resources displaced.

WHAT COMPLEMENTARY ASSETS EXPLAIN—AND WHAT THEY DO NOT

An AI service requires more than an algorithm. Computing hardware, software, power, cooling, network access, operating capabilities and users must work together. Owning an input that is difficult to substitute can create a route to income even when another firm owns the most visible innovation. The complementary-assets literature provides the starting point: returns to innovation depend partly on who controls the assets required to commercialise it, not just on who invented it. (Teece 1986)

Specific investment adds a second difficulty. Some expenditure is worth much less outside a particular relationship. Once it has been sunk, the parties may face incentives to renegotiate how the resulting surplus is divided. An apparent high return relative to avoidable operating cost can therefore be a quasi-rent supporting recovery of an earlier investment, rather than evidence of monopoly profit over the whole life cycle. Ownership, long-term contracts and other safeguards can respond to this problem, but each also changes control and flexibility. (Klein et al. 1978)

A long contract should consequently be read in two directions. It may restrict the customer's ability to switch. It may also make the supplier willing to invest in capacity that would otherwise not be built. Empirical work on coal supply to electricity generators links contract duration to relationship-specific investment, giving a historical benchmark for interpreting duration as an economic choice rather than an automatic sign of market power. It does not establish which duration is efficient for today's servers, grid connections or model developers. (Joskow 1987)

The OECD's 2025 infrastructure paper identifies potential efficiency gains from partnerships alongside possible barriers to switching and foreclosure. A relationship can coordinate investment and still merit competition scrutiny. That distinction avoids treating every integration agreement as exclusionary or assuming that supplier concentration alone identifies rent. (OECD 2025)

The FTC's January 2025 staff study adds documentary detail. In the three partnerships it examined, investment arrangements were linked to cloud-spending commitments, resource sharing and various control or preferential rights. Those findings concern information supplied through September 2024 and public material through January 2025; they are not a description of every current contract. The report explicitly disclaims an assessment of illegal conduct. Its relevance here is institutional: a cash investment and subsequent cloud purchases may be linked parts of one relationship, rather than independent demonstrations of final demand. (Federal Trade Commission 2025)

Aghion and Bolton supply the strongest theoretical rival to the efficient-safeguard account: an incumbent's long contracts can prevent some lower-cost entry without excluding every entrant. Their original abstract supports that proposition; the complete scanned article was not verified. The discriminators are feasible alternative suppliers, exclusivity relative to investment recovery, capacity reassignment and safeguards relative to unrecovered costs. The documents here permit a test of written replacement routes. They do not reveal the complete alternative-supplier set or the efficient level of protection. (Aghion and Bolton 1987)

Indian infrastructure evidence makes the counterparty mechanism concrete. Ryan's study of solar procurement compares state and central intermediation for similar projects and finds that counterparty risk affects bid prices and investment. It supplies a reason to take credible safeguards seriously, not a transferable estimate for compute. The productivity J-curve literature separately explains why complementary investments can precede measured returns; neither literature licenses treating an unobserved AI dividend as already earned. (Ryan 2021; Brynjolfsson et al. 2020)

WHEN THE PROMISE PAYS

Three adverse events must be separated: a provider fails to deliver, a customer uses less than expected, or a customer cannot pay. A demand commitment conditional on available service addresses the second event without necessarily covering the first or third. A utility ready to supply a customer whose facility never opens occupies a different position from a compute provider whose own servers are unavailable. The performing party matters as much as the percentage promised.

A simple accounting identity keeps incidence separate from productive value. Suppose ordinary receipts are 20 units and costs 54. The operator has a 34-unit deficit. A collected support payment of 60 turns its cash position into a 26-unit surplus and costs the support provider 60, leaving their combined deficit at 34. This is an assumed example, not an empirical result. It prevents the operator's receipt from being counted as a joint gain.

In practice, a credible commitment can lower financing costs, enable construction or improve coordination; it can also weaken screening or encourage excess capacity. Those behavioural effects determine whether combined value changes, and none follows from the transfer identity. The relevant comparison is between the resources and service produced with the commitment and a credible alternative without it.

Collectability also depends on who makes the promise. A counterparty exposed to the same weak demand may be least able to pay when protection matters most. Collateral, priority, termination rights and alternative customers govern that risk. The horizon includes replacement investment and refinancing as well as current receipts: contract length by itself does not establish that the assets can recover their costs.

CUSTOMER CREDITS AND CREDITOR CLAIMS

Under its June 2024 arrangement with Core Scientific, CoreWeave would fund conversion for approximately 200 MW of high-performance-computing capacity. An estimated $300 million associated with infrastructure the operator would own was to be recovered through credits against no more than half the monthly hosting fees. The operator receives funding now and relinquishes later receipts. This is a prepayment within a service relationship, not evidence that the customer has made a gift. (Core Scientific 2024 [corz2024])

The accompanying investor presentation illustrates credits of $145 million, $145 million and $10 million across 2025--2027. The schedule demonstrates timing, not a measured financing gain. The 9% and 13% rates previously used as illustration endpoints belonged to different exit facilities repaid in 2024, before the first illustrated credit year. They are unsuitable as a transaction-specific valuation range. The appendix uses an explicitly assumed rate grid and keeps the initial funding and credit timing assumptions visible. (Core Scientific 2024 [corzdeck2024]; Core Scientific 2026 [corzamend2026])

That timing component cannot establish the arrangement's value on its own. Hosting prices, priority and remedies may compensate the funding party. The presentation also describes an eventual transfer of certain customer-owned assets for nominal consideration. Without their quantity, condition and remaining service potential at expiry, this right cannot be valued or ranked against the funding benefit. The asset right and the financing component require separate valuations.

Subsequent evidence makes that distinction tangible. The operator's amended 2024 accounts, filed in March 2026, corrected the treatment of assets committed to demolition for conversion. Net property, plant and equipment had been overstated by approximately $122.9 million at the end of 2024. This accounting correction did not alter the reported cash-flow categories. It nevertheless demonstrates why a customer-funded build cannot establish the economic value of everything replaced in the process. (Core Scientific 2026 [corzamend2026])

A May 2026 presentation describes an additional 70 MW arrangement in which Core Scientific funds $104 million of powered core and shell, with no associated capex credit. The contractual allocation differs from the original tranche. The documents do not isolate the reason: site scope, price, access to finance and bargaining conditions may all differ. Those differences prevent the terms alone from establishing who gained bargaining power. (Core Scientific 2026 [corzmay2026])

CoreWeave's separate NVIDIA order addresses a different uncertainty. Its September 2025 filing reports an initial $6.3 billion arrangement to purchase residual unsold capacity through April 2032, conditional on delivery, availability and termination provisions. The promise can support receipts for eligible capacity that lacks another buyer. It does not turn undelivered equipment into a working service. The executed order was not retrieved in the bounded exhibit review; the redacted master agreement gives orders precedence. Therefore this paper does not calculate an unconditional liability from the headline amount. (CoreWeave 2025; CoreWeave 2023)

The construction advance and residual-capacity order cover different moments. Construction funding changes who supplies initial capital and when the operator receives net cash. Residual-capacity purchasing changes receipts in a particular low-demand state after performance conditions are met. Neither substitutes for the other. Funding can arrive before usable capacity; a demand commitment can remain conditional on delivery. The sequence through both contracts is more informative than the label guaranteed.

Core Scientific Finance I's $3.3 billion notes, issued on 6 May 2026, mature on 15 May 2031. Their coupon is 7.750%, with interest payable semi-annually beginning 15 November 2026. Initial principal amortisation is 11.5% of issue-date principal per year: $379.5 million, or $189.75 million for a normal half-year before the specified adjustments. These are material payment amounts, but the filing does not establish an actual first principal-payment date. (Core Scientific 2026 [corznotes2026]; Core Scientific Finance I LLC 2026)

The issue was not $3.3 billion of fresh construction spending. The issuer estimated net proceeds of $3.24 billion and intended to fund a debt-service reserve and distribute the balance to the parent; the same filing reports actual repayment of the bridge facility on 6 May. Deferred amortisation therefore accompanies refinancing and an upstream distribution as well as the development cycle. That does not prove that the distribution caused the deferral or that all remaining cash left on the issue date. Subsequent reserve balances, construction spending and completion-guarantee draws would distinguish an adequately funded transition from pressure on the parent's support. (Core Scientific 2026 [corznotes2026])

The parent's completion guarantee is uncapped, excludes the Austin campus, and is supported by a security interest in specified Parent Collateral. It funds completion if available project funds are insufficient; it does not guarantee the notes. The five subsidiary guarantors do guarantee the notes. Omitting these distinctions would understate the parent's construction obligation while overstating its promise to repay noteholders. (Core Scientific 2026 [corznotes2026])

The executed indenture defines the first principal instalment by the first payment date at least fifteen days after the later of final rent commencement and the abatement of all revenue credits under each Datacenter Lease. Tenant is CoreWeave; the defined leases include the June 2024 agreements and other sites. This establishes the counterparty connection, but the trigger cannot be dated from the original $300 million tranche alone. The later supplement distinguishes construction credits, end-of-term credits and performance credits. Neither a single credit balance nor an assumed exhaustion schedule establishes when the indenture condition is satisfied. Interest remains separately payable. (Core Scientific Finance I LLC 2026; Core Scientific 2026 [corzsupplement2026])

The waterfall also qualifies what a creditor can infer from gross hosting revenue. Its first bucket includes capital expenditure and specified intercompany amounts as well as ordinary operating expenses; debt-service and reserve requirements follow. Receipts for power costs can bypass the controlled revenue account even if characterised as rent. Thus a secured claim on project receipts is not equivalent to an unrestricted first claim on every dollar labelled revenue. (Core Scientific Finance I LLC 2026)

The indenture expressly allows a qualifying tenant to assume or replace the lease, subject to eligibility criteria and rating confirmation. On a defined lease-termination event, scheduled principal is deferred until the earliest of a termination event of default, the fifteenth business day after a qualifying replacement lease, or maturity. Reserve replenishment is separately suspended until the specified default or replacement deadline. The agreement accommodates re-letting rather than requiring uninterrupted repayment from the original tenant in every state. (Core Scientific Finance I LLC 2026)

The route is selective. The qualifying-tenant definition names major technology companies and also admits specified large listed firms, suitably guaranteed providers or parties approved by a majority of note principal. Replacement is a documented option, not evidence that a willing substitute is available on equivalent terms. It defeats a blanket interpretation of the finance as an irrevocable lock to one tenant, but cannot establish efficient competition or full recovery after tenant failure.

The residual rights come with matching obligations. The April 2026 supplement describes a fifty-percent monthly base-fee offset for construction credits, capped at $1.5 million per MW for the applicable non-Austin premises. Its separate end-of-term credits can require remittance of unused balances. Those provisions support the interpretation of customer funding as part of an exchange; they do not supply a net valuation of the eventual asset-transfer right. (Core Scientific 2026 [corzsupplement2026])

ACCOUNTING LIFE IS ANOTHER CLOCK

Three disclosures show why both direction and measurement basis matter: Microsoft and Alphabet quantified 2023 extensions to server or network-equipment lives, while Amazon subsequently shortened some lives. The selection is purposive, not a census of changed and unchanged policies.

Microsoft's FY2023 extension from four to six years increased operating income by $3.7 billion and net income by $3 billion, measured on the opening carrying amount. The net-income effect was about 4.1% of reported annual net income of $72.361 billion. Alphabet's 2023 extension from four to six years for servers and from five to six years for certain network equipment reduced depreciation by $3.9 billion and increased net income by $3 billion, about 4.1% of $73.795 billion reported net income. Its effect covers the opening stock and equipment placed in service during the year. Similar ratios therefore come from different extensions and populations. (Microsoft Corporation 2023; Alphabet Inc. 2023)

Amazon moved in the opposite direction effective 1 January 2025, shortening a subset of servers and networking equipment from six to five years. It attributed this to faster technological development, particularly AI and machine learning. Its reported annual effect was $1.4 billion more depreciation and amortisation and $1 billion less net income, primarily in AWS; the affected basis included opening equipment and in-year acquisitions. The countercase rules out a general story of ever-longer accounting lives and favours testing operational explanations over presuming manipulation. It does not establish the appropriate life for another firm's fleet. (Amazon.com, Inc. 2025)

The two positive ratios measure disclosed effects relative to earnings that already include them; they are neither AI profit shares nor shares of earnings growth. Amazon's negative case makes the direction of each estimate change indispensable. A consistent industry study would also include unchanged policies and match asset populations, periods and replacement evidence. These three disclosures support a bounded contrast, not that wider inference.

Depreciation and contractual payments affect different accounts. Extending an allocation schedule changes when cost enters accounting profit. A construction credit changes cash received from a customer. A purchase floor changes an entitlement to receipts under stated conditions. These can all improve a reported or private financial measure through different channels. Only the latter two directly define transactions between counterparties; the depreciation change alone does not create a new payer. It cannot be added to contractual support as another source of funding.

Later investment figures expand the scale of the question, but do not supply a new earnings-effect estimate. Microsoft's cash additions to property and equipment rose from $28.1 billion in FY2023 to $115.9 billion in FY2026, roughly 4.1 times as much. Total depreciation rose from $11 billion to $34.3 billion. These are different flows, and neither describes a uniform stock of AI servers. Finance-lease additions are outside that cash-spending line. Scaling the historical earnings effect by either growth rate would invent a counterfactual the disclosures do not provide. (Microsoft Corporation 2023; Microsoft Corporation 2026 [microsoft2026])

Microsoft said on its July 2026 earnings call that data-centre and office-building lives would extend from fifteen to twenty-five years starting FY2027. Management expected a minimal FY2027 operating-income benefit, with the larger impact on future lease classification and reported capital expenditure. That is prospective guidance about buildings, not a second measured extension of every server's life. (Microsoft Corporation 2026 [microsoftcall2026])

For capital recovery, the important evidence is how long equipment delivers economically useful service, the costs required to maintain that service and what replaces it. Technical survival is insufficient if running costs overwhelm the value of output; technological obsolescence is insufficient if the machine continues serving a profitable lower-demand workload. Both outcomes are possible. Contract renewals, retirement cohorts, maintenance and resale evidence would connect the accounting clock to the operating clock. The public disclosures used here do not complete that connection.

PUBLIC PARTICIPATION CREATES DIFFERENT CLAIMS

Norway's export-finance participation and Andhra Pradesh's incentive package create different public claims. In May 2026 Norway's minister reported Eksfin participation of $241.7 million in an approximately $790 million facility: about 30.6%. The minister described the agency as participating on the same terms as private banks. A later answer referred to categories such as margins and security requirements without publishing the actual margin, security package or enforcement ranking. (Norwegian Minister of Trade and Industry 2026 [norway2715]; Norwegian Minister of Trade and Industry 2026 [norway2864])

That percentage is a share of the committed financing, not verified disbursement, a grant rate or expected public loss. The answers describe financing with repayment obligations but do not establish whether Eksfin's participation is a direct loan, guaranteed lending or a combination. Equal terms with private banks are relevant without supplying the public claim's risk-adjusted value. Expected recovery, collateral, priority and concentration remain necessary, and unpublished rights cannot be assumed absent.

Andhra Pradesh's G.O.Ms.No.40 of 11 October 2025 records proposed investment of INR 87,520 crore and an overall incentive ceiling of INR 22,002 crore. The ratio is about 25.1%. Eleven quantified component limits add to INR 21,588 crore; the order separately states two concessions without rupee values. The INR 414 crore difference between the quantified limits and overall ceiling is not reconciled in the order. The gap does not demonstrate missing money. With two concessions unpriced, the quantified sum is not a complete valuation. (Government of Andhra Pradesh 2025)

Subsequent orders alter land allocation and implementation. An April 2026 order permits flexibility across incentive categories while preserving the overall amount and ties power-subsidy claims to actual consumption. These provisions affect when and how claims may arise. A ceiling is neither an amount already paid nor the discounted cost of future tax concessions, reimbursements and other support. The administrative test is how actual claims are checked against conditions and the overall limit. (Government of Andhra Pradesh 2026)

The denominators explain why 30.6% and 25.1% cannot rank the two public commitments. One divides public financing participation by the committed facility; the other divides an incentive ceiling by an applicant's proposed investment. Their durations, recovery rights and underlying projects differ. Putting both in one table clarifies disclosure and incidence; choosing the larger percentage cannot rank national generosity, fiscal risk or success.

Location adds another distinction. A building in a country need not be financed, owned or used solely by residents of that country. Foreign ownership does not make local value zero: wages, procurement, taxes, user benefits and infrastructure effects may matter. A policy evaluation needs those local benefits and costs relative to a credible alternative, including displaced uses of scarce power, land and public funds. The documents identify instruments and recipients more securely than they establish those counterfactual outcomes.

A MINIMUM BILL IS AN IMPERFECT FORECAST OF DEMAND

AEP Ohio's records expose the whole commitment sequence. Before the tariff, prospective data-centre customers expressed needs above 30,000 MW. Within the study deadline, 36 sites requested 13,022.7 MW. The September 2025 PJM slide rounded this to 13.0 GW and multiplied it by 85% to obtain 11.1 GW while contracts were prospective. By 12 February 2026, 5,642 MW had reached signed new-tariff commitments: 43.3% of study requests. That was added to 12,219 MW of pre-tariff contracts, making 17,861 MW contracted in total. Older contracts represented 68.4% of that total. (AEP Ohio 2025 [aeppjm2025]; AEP Ohio 2025 [aepletter2025]; AEP Ohio 2026 [aepupdate2026])

FIGURE 1: AEP Ohio's commitment stages and contracted composition. Expressed interest is shown at its reported lower bound; the actual value exceeds 30,000 MW. The right panel separates older contracts from new-tariff contracts, not protected from unprotected capacity. Source: 13 February 2026 release, with study-cohort timing from the November 2025 letter.
Asset: assets/aep_commitments.png

The older contracts are outside the new-load tier cohort, not outside Schedule DCT altogether. The current schedule includes Existing Load, a separate 60% billing-demand rule and collateral governed by existing agreements. New-load percentages cannot be applied to all 17,861 MW; neither does the older share measure wholly unprotected exposure. Expressed interest, studied load, signed capacity and actual use remain distinct quantities.

The November 2025 study supplies the delivery horizon. It found regional upgrades necessary for all 36 sites and estimated service in Q4 2031 for Clusters 1 and 3, and 2033 for Cluster 2. These were conditional planning dates, including an assumed PJM regional solution and parallel local works. Generation adequacy and wholesale-price effects were not evaluated. The study population is clear, but the public documents do not identify the precise cluster mix or individual identities of the later signed subset. A roughly six-year interval to the earlier planning date is not observed delay for each signed project. (AEP Ohio 2025 [aepletter2025])

The current tariff permits a sharper mathematical result. In the September 2026 book, sheets effective 28 August 2026 give a post-ramp capacity component M(C) for affiliated new loads aggregated across the territory. For 25<C≤75 MW, M(C)=0.85C-6.25. Above 75 MW it is the smaller of C-17.5 and 0.85C. Thus M(C)≤0.85C, with equality from C=350/3, or approximately 116.7 MW, upwards. At 50, 100 and 200 MW, the component is respectively 36.25, 82.5 and 170 MW. (Ohio Power Company 2026)

FIGURE 2: The post-ramp capacity component reaches an 85 percent plateau at 116.7 MW of aggregate affiliated new load. The dashed line is a component ceiling, not a ceiling on monthly billing demand. Source: DCT sheet 223-4, effective 28 August 2026.
Asset: assets/tariff.png

For any partition into eligible affiliated groups, applying a flat 85% to the sum cannot fall below the sum of these components. Equality requires every group to reach the plateau. This conditional comparison establishes no signed error in AEP's demand forecast. Actual metered demand and the historical ratchet can exceed the component; ramp rules and failure-to-energise billing differ. Unless mutually agreed otherwise, failure to energise is billed on Contract Capacity. The current sheet also postdates the forecast, and a cohort mean above 116.7 MW would not establish that every affiliated group exceeds it. The formula bounds one component. It does not measure actual use, the full bill or the appropriateness of a forecast multiplier.

The chronology does establish that an 85% billing-based adjustment at the study stage did not eliminate subsequent applicant selection: less than half the studied capacity reached the reported contract stage. It cannot identify how much the tariff caused that change, and February contracts are not realised electricity consumption against the September forecast. Screening, delivery and utilisation remain separate tests.

AEP's arrangements also permit adaptation. The current tariff sheets set an initial term of at least the ramp period plus eight years, continue minimum charges during customer curtailment or termination, and provide an exit route after five post-ramp years for the equivalent of 36 months' minimum charges. The sheets do not describe the full implementation arrangement. AEP's current guidance separately allows a customer to request assignment of up to 25% of its contracted capacity to another DCT customer in lieu of some or all exit fees on that portion. Successful assignment releases future obligations on the reassigned capacity; the original customer remains responsible for unused capacity it retains. (Ohio Power Company 2026; AEP Ohio 2026 [aepguidance2026])

The assignment is conditional. AEP must find it electrically feasible; the recipient must sign a DCT agreement, meet collateral requirements and pay incremental transfer costs. The transfer must leave no stranded investment for other ratepayers, or the assigning customer must pay that cost. Both parties must attest that no compensation beyond the specified costs is exchanged. AEP promises a good-faith effort, not a guaranteed substitute, and restricts the assigning customer's return for replacement capacity until circumstances change or a reasonable period passes. These are terms in the utility's implementation guidance, not an executed customer agreement. They make the remaining cost allocation explicit. (AEP Ohio 2026 [aepguidance2026])

Replacement has different consequences in the two cases. The indenture adjusts creditor payment timing after a tenant-termination event; AEP describes partial release of the departing customer's future obligations after a successful capacity assignment. Neither document establishes that an eligible substitute will exist, and neither supports a general ranking of flexibility or efficiency.

Collateral is also conditional on the applicant's financial position. A customer must provide qualifying support equal to half the full-term minimum charges unless it satisfies both the A-/A3 ratings condition and audited cash and equivalents above ten times the collateral requirement. A qualifying parent or affiliate guarantee is one permitted form. The requirement reduces with each energised year of timely payment and can be reassessed as financial conditions change. Neither investment-grade status alone nor being a hyperscaler establishes exemption. This is evidence of differentiated credit screening, not proof that security equals unrecovered infrastructure cost. (Ohio Power Company 2026)

These provisions are consistent with recovery of specific investment, but duration and collateral alone cannot prove proportionality. The study letter links local-service plans to cancellation reimbursement and construction contributions; it does not demonstrate recovery of every regional-system cost. Historical take-or-pay research supplies an efficiency explanation and circumstances in which private and social incentives diverge. Neither result can be transferred wholesale to this cohort. (AEP Ohio 2025 [aepletter2025]; Masten and Crocker 1985)

The next test is to follow the original 36-site study population, record which projects remain, and compare later energisation and demand with the November 2025 cluster estimates and each signed contract's path. Publication of the individual project mapping would permit a stronger project-level test. Infrastructure spending and recovered charges must then be matched separately. An overstatement of physical use and successful recovery of specified costs can coexist; one outcome would not settle the other.

EVIDENCE DESIGN AND THE LIMITS OF THE COMPARISON

Table 1 codes each mechanism by the event that creates a payment claim and the coverage the documents do not establish. The cases were chosen to examine different parts of the investment chain for which original evidence was accessible. The tariff, contracts and administrative instruments each change an obligation; accounting changes reported cost without creating another payer. The selection is not a representative survey, a count of independent projects or a measure of prevalence.

TABLE 1: Different instruments protect different events. The rows are mechanisms, not independent observations for statistical inference.
Mechanism | Trigger and parties | Addressed exposure | Remaining boundary
Construction advance, original 200 MW tranche | CoreWeave to Core Scientific. Conversion funding, then hosting credits | Early financing requirement | Performance and remaining project return
Residual-capacity order | NVIDIA to CoreWeave. Eligible available capacity remains unsold | Some weak-demand receipts | Completion failure; unconditional note repayment not established
Executed project notes | Issuer to noteholders. Interest and principal under separate terms; principal tied to rent and credits | Creditor priority and receipt timing | Qualifying replacement changes principal timing; parent completion support excludes Austin
DCT new-load tariff | Data-centre customer to utility. Service agreement and applicable demand/ramp rules | Recovery of specified minimum charges; conditional partial assignment | Remaining capacity obligation; physical use and systemwide cost recovery not established
Equipment-life revisions | No new counterparty payer. Extensions and a shortening alter cost allocation | Reported period expense changes in either direction | No new cash funding or proof of AI profit
Eksfin financing participation | Public participant and financed project. Committed financing with repayment obligations | Project financing | Tranche loan/guarantee structure, subsidy equivalent and expected loss not established
Andhra Pradesh incentive package | State and eligible claimant. Qualifying conditions and administrative claims | Incentive entitlement within ceiling | Actual expenditure and additional public benefit not established

Sources: original documents cited in Sections 4--7. Source identifiers and clause locators accompany the revised comparison data; the DCT row also uses AEP's current implementation guidance. No monetary amounts are pooled.

The records distinguish document values, calculated ratios, qualitative clause coding and assumed scenarios. Per-row locators accompany the matched ratios, while a separate register holds the main additional magnitudes. These are selected auditable inputs; not every number in the narrative belongs to one complete dataset. Source status distinguishes retained originals, retained extracts and documents known only through limited descriptions. A missing security term marks an evidence boundary, not the absence of security.

The literature review is focused and narrative. Readable originals of Teece, Klein and co-authors, Joskow, Masten and Crocker, and Ryan support the specified claims. Aghion and Bolton are used only for the proposition verified in their authors' institutional abstract; the retained scan's text extraction was unusable. Two recorded 2026 works by Hepp and by Yang, Qiu and Huang overlap the finance and tariff questions, but their full texts were not obtained, and a fresh retrieval again failed. The contribution is therefore the bounded clause comparison and calculation reported here, without an exhaustive novelty or priority claim. (Hepp 2026; Yang et al. 2026)

Four questions remain for both routes: available substitutes, exclusivity relative to recovery, reassignment and proportionality. The executed indenture and AEP's current implementation guidance each describe conditional replacement. Their consequences differ: creditor principal timing changes in the financing, whereas successful utility-capacity assignment releases the customer's future obligations on that portion while allocating transfer and stranded-equipment costs. The guide supplies implementation terms absent from the tariff sheets inspected and is not presented as an executed customer agreement. Actual substitutability, efficient duration and net welfare remain unresolved. The other table rows illuminate different mechanisms rather than add observations to this two-case comparison.

The reconstruction revisits the source material behind Indiconomics articles A01--A03 and adds the electricity and creditor chronology. Earlier conclusions are not treated as authority, and a missing retrieved clause is not treated as universal contractual absence. A causal study would need comparable projects, choices before contracting and outcomes after service commencement. The result here concerns written obligations and measured commitment stages.

WHERE THE AI DIVIDEND WOULD HAVE TO APPEAR

The contracts protect different clocks. Ohio's 85% figure describes neither a universal minimum percentage for each applicant nor a realised-demand observation. Its precisely bounded capacity component sits alongside separate credit screening, exit charges and a long conditional transmission timetable. The project notes separately attach material principal payments to commencement, credit exhaustion and qualifying tenant replacement. Stronger commitment at one stage therefore does not establish uninterrupted cash at the next.

This changes the interpretation of the headline measures. A noteholder needs the timing and permitted uses of net receipts, not just contracted megawatts. An Ohio planner needs applicant survival and service dates as well as minimum charges. For India, the Andhra Pradesh orders identify incentive limits and claim conditions; additional domestic income and productivity still require evidence on realised benefits, fiscal cost and alternative uses of scarce resources. None can be inferred by adding those limits to foreign financing commitments or accounting earnings effects.

The next decisive observation is the fate of the 36-site Ohio study population relative to its Q4 2031 and 2033 planning vintages. If survivors energise and pay but consume less than expected, the tariff may protect specified claims while the demand forecast disappoints. If substitutes occupy abandoned capacity, adaptation may reduce the loss; the terms determine who shares that benefit. For the project notes, the comparable observations are the actual first principal payment and what happens after any tenant replacement. These tests connect protection to delivered service and retained returns without assuming that a contractual claim is already a productive dividend.

APPENDIX

REPRODUCTION AND DEFINITIONS

Evidence was checked through 8 September 2026. Historical amounts and planning dates retain their original reporting periods. The companion package contains the coded comparison, observed numerator and denominator inputs, calculated ratios, tariff reconstruction and separate scenario data. The scenario code does not estimate default probabilities or company values. Dollar and rupee amounts are never added across instruments.

For aggregate contracted new-load capacity C in MW under common ownership or affiliation, the post-ramp capacity component reconstructed from DCT sheet 223-4 is

M(C)={
15+0.85(C-25),  if 25<C≤75,
min{57.5+(C-75),0.85C},  if C>75.
}

This component does not replace metered demand, the historical ratchet or other tariff provisions. It is not an electricity-consumption forecast. At the threshold, eligibility and the tariff's kW bands govern; the chart starts above 25 MW.

For the construction-credit illustration, V(r)=300-145/(1+r)-145/(1+r)^2-10/(1+r)^3, in millions of dollars. The revised sensitivity grid uses assumed rates of 0, 5, 10 and 15 percent; none is a matched project discount rate. Full funding is placed at time zero and credits at year ends. Slower credit exhaustion is not independently favourable: it can accompany weaker hosting receipts and changes the principal-start condition. No standalone timing component values the complete transaction.

The current tariff book was inspected, but the latest appellate disposition was not established. This paper describes the filed schedule and historical implementation documents, without asserting judicial finality. The complete NVIDIA September 2025 order and detailed public-financing security terms were not obtained. Their absence constrains the relevant claims rather than all historical documentary analysis.

The source audit states which originals or extracts are retained and records retrieval failures separately. The amended filing and April supplement were newly retained on 8 September 2026; that does not retroactively repair earlier archive claims. The distribution package supplies source records and original analysis, not a collection of third-party copyrighted papers.

REFERENCES

[aeppjm2025] AEP Ohio (2025). AEP Large Load Forecast Update. PJM Load Analysis Subcommittee, September 16, slide 4. https://www.pjm.com/-/media/DotCom/committees-groups/subcommittees/las/2025/20250916/20250916-item-04f---aep-large-load-request.pdf

[aepupdate2026] AEP Ohio (2026). AEP Ohio Updates PUCO on Data Center Load: Figures Show Tariff is Working. February 13 release; contracts as of February 12, study cohort and older contracts. https://www.aepohio.com/company/news/view?releaseID=10753

[teece1986] Teece, David J. (1986). Profiting from technological innovation: Implications for integration, collaboration, licensing and public policy. Research Policy, 15, 6. 285--305. Sections 3.1,3.3 and 5.1; printed pp. 287--293. https://doi.org/10.1016/0048-7333(86)90027-2

[klein1978] Klein, Benjamin and Crawford, Robert G. and Alchian, Armen A. (1978). Vertical Integration, Appropriable Rents, and the Competitive Contracting Process. Journal of Law and Economics, 21, 2. 297--326. Section I pp. 298--299; concluding comment p. 325. https://doi.org/10.1086/466922

[joskow1987] Joskow, Paul L. (1987). Contract Duration and Relationship-Specific Investments: Empirical Evidence from Coal Markets. American Economic Review, 77, 1. 168--185. Section III; Tables 3--6; Section VI. https://masonlec.org/site/rte_uploads/files/Joskow%20class%202%20reference.pdf

[oecd2025] OECD (2025). Competition in artificial intelligence infrastructure. Competition Policy Papers 330, Sections 3.4--3.6. https://www.oecd.org/content/dam/oecd/en/publications/reports/2025/11/competition-in-artificial-intelligence-infrastructure_69319aee/623d1874-en.pdf

[ftc2025] Federal Trade Commission (2025). Partnerships Between Cloud Service Providers and AI Developers. Office of Technology staff report, January; pp. 19--20 and 31--32. https://www.ftc.gov/system/files/ftc_gov/pdf/p246201_aipartnerships6breport_redacted_0.pdf

[aghion1987] Aghion, Philippe and Bolton, Patrick (1987). Contracts as a Barrier to Entry. American Economic Review, 77, 3. 388--401. Author institutional abstract only; retained full scan is not text-readable. https://business.columbia.edu/faculty/research/contracts-barrier-entry

[ryan2021] Ryan, Nicholas (2021). Holding Up Green Energy. Cowles Discussion Paper 2294, August 2021 version; abstract and introduction, printed pp. 1--4. https://economics.yale.edu/sites/default/files/ryan_n_-_holding_up_green_energy_d2294.pdf

[brynjolfsson2020] Brynjolfsson, Erik and Rock, Daniel and Syverson, Chad (2020). The Productivity J-Curve: How Intangibles Complement General Purpose Technologies. NBER Working Paper 25148, revised January 2020; original issued 2018. https://www.nber.org/papers/w25148

[corz2024] Core Scientific (2024). Core Scientific to Provide Approximately 200 MW of Infrastructure to Host CoreWeave’s High-Performance Computing Services, Capturing Significant AI Compute Opportunity. June 4 announcement: approximately 200 MW and construction credits. https://www.sec.gov/Archives/edgar/data/1839341/000162828024026483/corzcoreweavehpc.htm

[corzdeck2024] Core Scientific (2024). Investor Day Presentation. June 2024, slides 22--23; annual credit illustration. https://investors.corescientific.com/sec-filings/all-sec-filings/content/0001628280-24-027821/corescientificinvestorda.htm

[corzamend2026] Core Scientific (2026). Amended Annual Report for the Year Ended December 31, 2024. Filed March 2, 2026; Note 3 restatement; Note 9 Convertible and Other Notes Payable, instrument rates and repayment dates. https://investors.corescientific.com/sec-filings/all-sec-filings/content/0001628280-26-013232/core-20241231.htm

[corzmay2026] Core Scientific (2026). First Quarter Fiscal 2026 Earnings Presentation. May, slide 13 note 4; incremental 70 MW. https://investors.corescientific.com/sec-filings/all-sec-filings/content/0001628280-26-031246/q1fy26earningsdeck.htm

[crwvorder2025] CoreWeave (2025). Current Report on Form 8-K: NVIDIA Order. Filed September 15, Item 1.01; order dated September 9, ending April 13, 2032. https://www.sec.gov/Archives/edgar/data/1769628/000176962825000047/crwv-20250909.htm

[crwvmsa2023] CoreWeave (2023). Master Services Agreement with NVIDIA. Redacted agreement filed 2025, order precedence; not the September 2025 order. https://www.sec.gov/Archives/edgar/data/1769628/000114036125036118/ny20053122x6_ex10-31.htm

[corznotes2026] Core Scientific (2026). Current Report on Form 8-K: Senior Secured Notes. May 6, Item 1.01; completed offering and completion guarantee. https://investors.corescientific.com/sec-filings/all-sec-filings/content/0001193125-26-208905/d47435d8k.htm

[corzindenture2026] Core Scientific Finance I LLC (2026). Indenture: 7.750 Percent Senior Secured Notes Due 2031. Executed May 6; Sections 1.01, 3.09, 4.23 and 14.01. https://investors.corescientific.com/sec-filings/all-sec-filings/content/0001193125-26-208905/d47435dex41.htm

[corzsupplement2026] Core Scientific (2026). Supplemental Information Regarding Core Scientific, Inc.. April 21, 2026; CoreWeave Licensing Agreements: Development Costs and Expenses; Credit; End of Term Credit; performance credits. https://investors.corescientific.com/sec-filings/all-sec-filings/content/0001193125-26-165121/d149019dex992.htm

[microsoft2023] Microsoft Corporation (2023). Annual Report 2023. Year ended June 30, Note 1 and financial statements. https://www.microsoft.com/investor/reports/ar23/index.html

[alphabet2023] Alphabet Inc. (2023). Annual Report on Form 10-K for the Year Ended December 31, 2023. Note 1, useful lives and consolidated net income. https://www.sec.gov/Archives/edgar/data/1652044/000165204424000022/goog-20231231.htm

[amazon2025] Amazon.com, Inc. (2025). Annual Report on Form 10-K for the Year Ended December 31, 2025. Note 1, Use of Estimates, printed p. 41; year-end 2025 financial statements. https://www.sec.gov/Archives/edgar/data/1018724/000101872426000004/amzn-20251231.htm

[microsoft2026] Microsoft Corporation (2026). Annual Report on Form 10-K for the Year Ended June 30, 2026. Property and equipment and cash-flow statements. https://www.sec.gov/Archives/edgar/data/789019/000119312526323660/msft-20260630.htm

[microsoftcall2026] Microsoft Corporation (2026). FY2026 Fourth Quarter Earnings Conference Call. July 2026; FY2027 guidance on data-centre and office-building lives. https://www.microsoft.com/en-us/investor/events/fy-2026/earnings-fy-2026-q4

[norway2715] Norwegian Minister of Trade and Industry (2026). Written Answer on Eksfin Participation in Nscale Financing. Question 15:2715; May 2026; dollar amounts and participation terms. https://www.stortinget.no/no/Saker-og-publikasjoner/Sporsmal/Skriftlige-sporsmal-og-svar/Skriftlig-sporsmal/?qnid=123118

[norway2864] Norwegian Minister of Trade and Industry (2026). Written Answer on Export Criteria and Eksfin Financing Terms. Question 15:2864; export eligibility, margins and security categories. https://www.stortinget.no/no/Saker-og-publikasjoner/Sporsmal/Skriftlige-sporsmal-og-svar/Skriftlig-sporsmal/?qnid=124334

[apgo40] Government of Andhra Pradesh (2025). G.O.Ms.No.40: Incentive Package for Raiden Infotech India. ITE and C Department, October 11; investment, component caps and total ceiling. https://goir.ap.gov.in/dgo.ashx?gid=561152&fileType=E

[apgo20] Government of Andhra Pradesh (2026). G.O.Ms.No.20: Implementation Flexibility for the Incentive Package. ITE and C Department, April 17; total ceiling and power-consumption basis. https://goir.ap.gov.in/dgo.ashx?gid=573527&fileType=E

[aepletter2025] AEP Ohio (2025). Data Center Tariff Load Study Letter. November 7, pp.1--6; 36 projects and generation adequacy assumption. https://www.aepohio.com/lib/docs/ratesandtariffs/ohio/AEP-Ohio_DCT_Load_Study_Letter_25.11.7.pdf

[aeptariff2026] Ohio Power Company (2026). Schedule DCT, Data Center Tariff. September tariff book, sheets 223-1--223-8, effective August 28; PDF pp.113--120. https://www.aepohio.com/lib/docs/ratesandtariffs/Ohio/September_2026_AEP_Ohio_Tariff_Book.pdf

[aepguidance2026] AEP Ohio (2026). Data Center Tariff: Implementation Guidance. Current guidance checked 8 September 2026; Assigning Retail Capacity to Another Customer; 2026 Study Tranche Two. https://www.aepohio.com/company/about/rates/data-center-tariff/

[masten1985] Masten, Scott E. and Crocker, Keith J. (1985). Efficient Adaptation in Long-Term Contracts: Take-or-Pay Provisions for Natural Gas. American Economic Review, 75, 5. 1083--1093. Sections I and II; Table 2 and p.1091 footnote 21. https://masonlec.org/site/rte_uploads/files/Masten%20and%20Crocker%20class%203%20Efficient%20Adaptation%20reference.pdf

[hepp2026] Hepp, Stefan (2026). The AI Infrastructure Debt Complex: Where the Leverage Behind Artificial Intelligence Sits, How It Connects, and Who Would Bear the Loss. Original abstract/landing page only; full text not obtained. https://papers.ssrn.com/sol3/papers.cfm?abstract_id=7161938

[yang2026] Yang, Fuli and Qiu, Tingting and Huang, Jing (2026). Who Pays When the Data Center Never Comes? Contract Design and Stranded-Cost Risk in U.S. Large-Load Electricity Tariffs. SSRN, written August 8; original abstract only; full text inaccessible. https://papers.ssrn.com/sol3/papers.cfm?abstract_id=7250402
