Research article · 7 September 2026

Who Financed the AI Site?

CoreWeave’s $300 million construction advance reveals a financing mechanism. Its value depends on timing and the rest of the bargain; neither the advance nor a later takeover offer measures economic rent.

A customer pays to convert its supplier’s data centre, then takes the money back through lower monthly bills. Who benefits? The answer depends on the years between those two payments, on what the customer receives in return, and on who carries the risks that remain.

In June 2024 Core Scientific announced twelve-year hosting agreements with CoreWeave for approximately 200 MW dedicated to high-performance computing, at sites with roughly 280 MW of gross capacity. CoreWeave would fund the conversion. An estimated $300 million associated with infrastructure Core Scientific would own would be recovered through credits of no more than half the monthly hosting fees. The announcement projected over $3.5 billion of revenue over the initial terms, before those prepaid build-out costs. [1]

The customer supplies funding and the operator gives up future receipts. This provides financing, although the disclosure describes a prepayment within a hosting arrangement, not a conventional standalone loan. A headline contract value does not reveal how much cash the operator ultimately keeps.

A calculation with an explicit clock

The June 2024 investor presentation supplies an illustrative credit schedule: $145 million in 2025, $145 million in 2026 and $10 million in 2027, following funding in 2024. It records the prepayment as deferred revenue. [2] That supports a timing calculation, but not an observed financing gain: the presentation does not supply exact cash-transfer dates or establish that the illustrated credits occurred.

Assume all $300 million arrives at time zero and credits fall at the ends of years one, two and three. At a chosen annual discount rate, the timing value is $300 million minus the present value of those credits. The table uses rates from Core Scientific’s January 2024 exit financing to illustrate sensitivity. Those instruments had different collateral, maturities, payment options and risks. Their coupons are context for scenarios, not a matched estimate of the customer’s opportunity cost or the operator’s marginal borrowing rate. [3]

Indiconomics scenario: 300 − 145/(1+r) − 145/(1+r)² − 10/(1+r)³, in USD millions. Full advance at time zero; year-end credits; no default, tax or other contract cash flows. Annual credits come from the company’s illustrative schedule, not a realised repayment record.

Under those assumptions, the timing value is $37–51 million. Spreading credits through each year would lower it relative to year-end credits; staging the advance would also change it. Delaying only the credits by one year raises the values at 9% and 12.5% to $58.9 million and $77.4 million. That sensitivity is not a claim that delay benefits the whole project: later delivery can also postpone revenue and trigger penalties.

The central limit is contractual. The schedule contains no separately stated interest charge, but the customer could receive compensation through the hosting price, priority, performance rights or other terms. A financing concession cannot be valued independently of those offsets without a credible alternative bargain. The calculation identifies a time-value component, not a free transfer, a price of impatience or economic rent.

What the customer eventually hands over

The investor deck describes another right: conversion capital above $1.5 million per HPC MW would be funded and owned by CoreWeave, then transferred to Core Scientific at contract expiry for nominal consideration. [2]

That may matter to the operator, but its size is unknown. Original expenditure is different from residual value after years of use. Without an asset schedule, remaining service potential and any associated obligations, the eventual transfer cannot be ranked above the financing component or counted as profit today.

Later disclosure demonstrates why asset value deserves its own test. In March 2026 Core Scientific amended its 2024 accounts after finding that assets committed to demolition for HPC conversion had remained capitalised when they should have been impaired. The cumulative overstatement of net property, plant and equipment at December 2024 was about $122.9 million. The company said the correction did not change revenue or operating, investing and financing cash flows. [3] This is an accounting correction, not a net present value of the conversion. It does show why customer funding alone cannot establish the operator’s economic return.

One relationship, different terms

Core Scientific’s May 2026 presentation describes an additional 70 MW agreement under which it funds $104 million for the powered core and shell, with no associated capex credit. [4] The original tranche used customer funding; this tranche uses operator funding. The comparison establishes contractual variation within the same relationship. It does not establish why the terms changed: project scope, prices, financing access and bargaining conditions can all differ.

The April 2026 financing supplement separates construction credits from performance remedies. The former return agreed funding, capped at $1.5 million per MW. Late-delivery credits and outage credits compensate different failures; the disclosed delay schedule includes $20,000 a day for days 31–60 and $30,000 thereafter, with exceptions. Commencement more than 120 days after the target can permit termination on notice. [5] These mechanisms may all reduce fees, but netting them together would conceal which risk each addresses.

By June 2026 about 395 MW of an approximately 590 MW leased portfolio had commenced billing: roughly two-thirds. The inputs refer to the expanded relationship at the same date. This is progress from signed capacity to billing, not GPU utilisation, cash collection or profit. The unbilled portion cannot be labelled late without matching sites to contractual schedules. [6]

Core Scientific also states that colocation electricity costs pass through to its customer without markup, so price changes move revenue and cost together without changing colocation gross profit. [6] That addresses electricity-price exposure; it does not establish that enough power is available or remove service obligations.

A buyer for capacity still needs a working site

A separate agreement, further up the supply chain, addresses another risk. CoreWeave’s September 2025 filing describes an NVIDIA order with an initial value of $6.3 billion: NVIDIA must buy residual unsold capacity through April 2032, subject to delivery, availability and termination conditions. [7] That can support demand for eligible capacity. It cannot make an unfinished facility usable.

CoreWeave’s April 2026 proxy reports NVIDIA payments of approximately $326.3 million during 2025 and $59.6 million in the first quarter of 2026 under their master agreement and related orders. It does not identify those receipts as purchases under the September order. [8] Nor is the initial order value directly comparable with company-wide remaining performance obligations, which deduct estimated variable consideration, including potential customer credits. [9]

The executed September order was not retrieved in the earlier bounded exhibit review. The filed master agreement is redacted and gives order forms precedence. [10] The public record therefore supports the purchase mechanism more securely than an estimate of enforceable exposure, eligible capacity or realised benefit.

What the failed takeover can tell us

In July 2025 CoreWeave agreed to acquire Core Scientific at 0.1235 CoreWeave shares for each Core Scientific share. The announcement put the implied equity value at approximately $9 billion and highlighted elimination of more than $10 billion of cumulative future lease overhead. It also described other assets, expansion opportunities and operating savings. In October, the deal failed to receive the necessary shareholder approval and was terminated. [11] [12]

Ownership could internalise a contractual payment stream, but the bid was for an entire company. An equity valuation and twelve years of undiscounted lease payments are different quantities. Their comparison cannot isolate the value of an income right, still less the return above the opportunity cost of all resources employed. Shareholders’ rejection is evidence that the transaction failed, not a measurement of scarcity. Later standalone share prices also cannot reveal what the merged company would have been worth.

The strongest commercial explanation remains coordination. A site owner needs construction funding, a compute provider needs timely capacity, and a supplier may value access to spare computing. Contracts can align those needs without a monopoly return. The rival concern is that commitments can sustain construction whose eventual receipts fail to justify its costs. Neither account is settled by the size of a prepayment or a takeover headline.

Customer funding, residual asset rights, performance credits and conditional capacity purchases allocate different risks, even when they appear in the same chain of bills. Pricing the time-value component makes that distinction concrete. Establishing rent requires the next step: actual project cash flows and costs, together with credible alternatives available to both sides.

Sources and calculation notes

1. Core Scientific, hosting announcement, 4 June 2024. Capacity, funding, credit ceiling, term and projected revenue; management announcement, not the complete contract.
2. Core Scientific, Investor Day, 12 June 2024, slides 22–23. Illustrative annual credit schedule and end-of-term asset transfer.
3. Core Scientific, amended 2024 Form 10-K, 2 March 2026, Note 3 and the debt disclosures. Restatement and historical debt terms. The 9%, 10%, 12.5% and 13% scenarios correspond to disclosed exit-facility, convertible cash-pay, secured-note and equipment-facility rates; alternatives and collateral differ. They are not a representative borrowing-cost range.
4. Core Scientific, Q1 2026 presentation, slide 13 note 4. Additional 70 MW funding arrangement.
5. Core Scientific, financing supplement, 21 April 2026. Summary of construction and performance terms; no claim that a remedy was triggered.
6. Core Scientific, June 2026 Form 10-Q. Same-date billing comparison: 395/590 = 66.95%, rounded to two-thirds because inputs are approximate; electricity pass-through.
7. CoreWeave, 15 September 2025 Form 8-K, Item 1.01. Conditional NVIDIA order.
8. CoreWeave, April 2026 proxy, NVIDIA related-party transactions. Relationship payments are not order-specific receipts.
9. CoreWeave, September 2025 Form 10-Q, Note 2. RPO measurement definition.
10. CoreWeave–NVIDIA master services agreement, executed April 2023. Redactions and order precedence. Earlier review inspected subsequent annual/interim exhibit lists without retrieving the September order; this is not an exhaustive no-disclosure claim.
11. Joint merger announcement, 7 July 2025. The $9bn was an implied equity value, not a price for leases alone.
12. Core Scientific, termination announcement, 30 October 2025.