Skip to main content

An AI cloud prepayment is an unsecured loan to an 11.6-times-levered borrower.

A neocloud rents out large clusters of AI chips to developers, funded mostly by debt and by its customers' advance payments. CoreWeave, the largest, carries $46.9 billion of debt against $4.0 billion of equity, and expects roughly $9 billion more advances within three months.
Information technologyCloud and AI services
11.6 timesCoreWeave's debt to equity, the borrower a prepayment is lent against
$132.5bnCoreWeave's contracted revenue backlog, not cash
15% to 25%the typical prepayment share of a contract's value
$150m-$250mthe unsecured advance at risk on a $1 billion commitment

CoreWeave, the largest of the AI cloud providers called neoclouds, grew revenue 168% in the quarter to 30 September 2026 and reports a $132.5 billion backlog [s2]. The same filing shows a $709 million loss and an $846 million interest bill, with roughly $9 billion more in customer prepayments expected within three months [s2]. The prepayments are the point. A customer who pays a neocloud in advance is not just buying capacity. It is lending to a borrower with 11.6 times as much debt as equity, without the security a bank would demand [s1] [s2]. If the borrower falters, that advance is an unsecured claim that can rank behind the lenders holding the chips it paid for [s1]. The secured lenders win. The customers holding prepayments and the unsecured bondholders lose [s1]. A typical advance is 15 to 25 per cent of the contract, so a $1 billion commitment puts $150 million to $250 million at risk [s1]. Do not refuse the neocloud. Underwrite the prepayment as credit: cap it at a board-approved loss limit, stage it against delivery, take escrow or guarantees where you can, and split capacity across at least two suppliers. Set the loss limit by 28 November 2026.

What changed

CoreWeave, the largest of the AI cloud providers called neoclouds, reported its quarter to 30 September 2026 on 10 November 2026 [s2]. Revenue rose 168% to $3.7 billion, and the company holds a $132.5 billion revenue backlog [s2]. The same filing reports a $709 million net loss and $846 million of net interest expense, up 173% in a year [s2]. The backlog is not cash. CoreWeave's own definition says it is future revenue under committed customer contracts, subject to delivery and availability [s2]. In the third quarters of 2026 and 2025, 98% of revenue came from these take-or-pay commitments, where the customer pays whether or not it uses the capacity [s1].

The company expects roughly $9 billion more in customer prepayments and deposits within three months [s2]. More than $3 billion of that is tied to new contracts still to be signed; the rest flows from contracts already committed [s6]. In the same window Amazon's cloud business grew 39% and it guided about $220 billion of 2026 capital spending, with free cash flow negative [s3]. Amazon is building its own capacity.

Who gains and who loses

If this debt-funded tier falters, who loses turns on how each claim is secured. CoreWeave's delayed-draw term loans hold first-priority pledges on the assets of their borrowing subsidiaries, and its other term loans are collateralised with contracted cash flows [s1]. Its senior notes and convertible notes are unsecured obligations [s1]. A customer prepayment sits on the balance sheet as deferred revenue, money collected before the service is delivered, a general unsecured liability [s1]. CoreWeave's own words place the financed infrastructure and the related customer contracts inside bankruptcy-remote subsidiaries, whose assets may be used only to settle their own obligations [s1]. That suggests the prepayment could be structurally subordinated to the secured lenders [s1]. This is an inference, not a settled fact: the filing reports deferred revenue only on a consolidated basis and does not say which legal entity holds it [s1]. What is settled is that the prepayment is unsecured [s1]. Equity, $4.0 billion against $91.5 billion of assets, absorbs loss first [s2]. The secured lenders are protected. The customers holding prepayments and the unsecured bondholders are not.

How the prepayment becomes a loan

The cost of the two debt layers shows who is paid for the risk. The secured term loans pay the secured overnight financing rate, SOFR, plus 2.25% to 5.50% [s1]. The senior notes carry coupons of 8.5% to 9.75%, and some of the company's bonds trade above 15% yield [s1] [s6]. The customer advances 15% to 25% of total contract value, disclosed as of 31 December 2025, at no separately priced credit spread [s1]. The chief executive told the market the company will not use the unsecured debt market at current yields [s6]. The least-protected lender is the customer, and it is paid the least for the risk. Nine-month operating cash flow of $5.6 billion included $3.5 billion of net new prepayments, 62% of the total [s2]. The build-out is funded mainly by customers' advances, not by profit.

What it is worth

The debt-funded AI infrastructure build this sits inside is projected at more than $7 trillion of debt by 2029, on one industry analyst's estimate [s5]. At CoreWeave alone, the prepayment exposure is $10.0 billion of deferred revenue on the balance sheet at 30 September 2026, plus roughly $9 billion of new prepayments expected in three months [s2]. For a buyer, the advance is bounded at 15% to 25% of the contract [s1]. The loss in a default is the advance not recovered: contract value times prepayment share times one minus recovery. Recovery is not a certainty. On a consolidated basis, unsecured claims recover in full unless assets fall about 37% to 46%, and reach zero only at a 71% to 79% markdown [s1]. Under the structural-subordination reading, recovery could be far lower, in the 0% to 20% range [s1]. Deep chip markdowns are real. The benchmark's data shows H100 rental rates fell from about $8 an hour at the 2024 peak to $1.96 at the trough [s4]. They recovered to $2.64 by April 2026, still about 67% below peak [s4]. The default probability is the reader's own input. No public source measures it, and this article does not invent one.

Exhibit 1: Debt to equity, CoreWeave against Amazon

CoreWeave's recourse-only leverage is 10.3 times; Amazon's including operating leases is 0.44 times.

CoreWeave Form 10-Q [s1] and earnings release [s2] for the quarter ended 30 September 2026; Amazon earnings release [s3] for the same quarter. Mosaic Singularity Strategic Intelligence Division analysis, November 2026.

Exhibit 2: CoreWeave against Amazon, exact balances and funding
MeasureCoreWeave (neocloud)Amazon (cloud provider)
Total debt$46.9 billion, of which $5.3 billion non-recourse$173.1 billion long-term debt, plus $98.8 billion of lease liabilities
Equity$4.0 billion$613.6 billion
Debt to equity11.6 times (10.3 times recourse alone)0.28 times (0.44 times including leases)
How the build is fundedSecured term loans at SOFR plus 2.25% to 5.50%, plus customer prepayments that were 62% of nine-month operating cash flowMainly operating cash flow of $168.3 billion, up 29%

CoreWeave Form 10-Q [s1] and earnings release [s2] for the quarter ended 30 September 2026; Amazon earnings release [s3] for the same quarter. Debt to equity is computed from the reported balances; Amazon lease liabilities are operating leases. Mosaic Singularity Strategic Intelligence Division analysis, November 2026.

Exhibit 3: What a prepayment risks at three recovery levels
Contract valuePrepayment advanced (15% to 25%)Loss at 20% recoveryLoss at zero recovery
$500 million$75 to $125 million$60 to $100 million$75 to $125 million
$1 billion$150 to $250 million$120 to $200 million$150 to $250 million
$5 billion$750 million to $1.25 billion$600 million to $1.0 billion$750 million to $1.25 billion
$10 billion$1.5 to $2.5 billion$1.2 to $2.0 billion$1.5 to $2.5 billion

Prepayment share of 15% to 25% of total contract value is from CoreWeave's Form 10-Q [s1] for the quarter ended 30 September 2026, disclosed as of 31 December 2025. Loss equals the prepayment times one minus recovery; recovery is the buyer's own input, and the 20% and zero columns span the structural-subordination branch, an inference, with the consolidated figure as the upper bound. Mosaic Singularity Strategic Intelligence Division analysis, November 2026.

The options and the recommendation

Three ways to buy capacity, and what each leaves open
OptionWhat it costsWhat could go wrongWhat it rules out
Sign the neocloud's standard termsUp to the whole advance in a defaultFull prepayment, 15% to 25% of the contract, unsecuredNo way to reclaim the prepayment or step into the cluster later
Contract with Amazon insteadLess flexibility on chip tiering; the record does not establish which is cheaperBuy from a provider whose leverage is 0.28 to 0.44 times equity, funded mainly by operating cash flowThe neocloud's purpose-built clusters

The recommendation is Option C. Use the neocloud, because its purpose-built clusters are scarce, but underwrite the prepayment as credit. Cap it at a board-approved loss limit and stage it against delivery milestones. Take escrow, a letter of credit, or step-in rights where the contract and the buyer's leverage allow. Split capacity across at least two counterparties with a large cloud provider as fallback. Two levers are the buyer's alone: walking away, and getting the board to approve a total-loss figure. Capping and staging are negotiated, not unilateral, in the seller's market the record describes. Before signing, confirm the prepayment's price concession compensates the credit risk. Meta, OpenAI and Jane Street have each committed billions to such contracts [s1], so the concession is real and worth naming. If protection is refused, walk to the cloud provider or keep the neocloud but cap the prepayment at a limit the chief financial officer signs as a total-loss figure.

What to do, and by when

Three actions, each with a named owner and a date

  1. Set a counterparty credit limit for each neocloud equal to the largest prepayment the board will accept as a total loss, covering new commitments and the prepayments already committedChief financial officer or treasurer
  2. Reopen new 2027 and 2028 term sheets and insert milestone-staged prepayment, and escrow, letter-of-credit or step-in clauses where the buyer's size gives leverageChief technology officer or head of cloud procurement, with legal
  3. Split the commitment across at least two counterparties so no single unsecured claim exceeds the board-approved limitChief technology officer and chief financial officer jointly

The measure that shows the work worked: by 31 January 2027, zero per cent of the 2027 and 2028 prepayment is unsecured and unprotected beyond the board-approved loss limit. The threshold for paying for protection is derived, not asserted. Protect when the chance of default exceeds the cost of protection divided by one minus recovery. At a 0% to 20% recovery and a 2% to 5% annual protection cost, that flip is at about a 2% to 6% annual default probability. That is not a tail risk for a borrower whose bonds trade above 15% [s6].

What would change the answer

Three outcomes change the move, and each has a signpost.

Scenarios

Interest keeps compounding

What happens
Coverage falls
The move
Cap, stage, diversify

You are in this one ifThe fourth-quarter interest figure of $1.05 to $1.15 billion and the first-quarter 2027 prepayment receipt

Pricing power holds

What happens
Earnings outrun the interest bill
The move
Stay the same, because the prepayment is unsecured either way

You are in this one ifThe fourth-quarter report

A top customer declines to renew

What happens
A top customer declines to renew
The move
Diversification, already in the plan

You are in this one ifThe next concentration disclosure

How we did this

Every figure is computed from those source values. 11.6 times is $46.9 billion of debt over $4.0 billion of equity, including $5.3 billion of non-recourse debt, and 10.3 times is recourse debt alone. 62% is $2.3 billion of adjusted EBITDA, earnings before interest, depreciation, amortisation and share-based pay, over $3.7 billion of revenue. Coverage is adjusted EBITDA over interest: 2.67 times in the third quarter and 2.05 times against the fourth-quarter guidance midpoint. 62% of operating cash flow is $3.5 billion of net new prepayments over $5.6 billion of nine-month operating cash flow. The loss table applies 15% to 25% of contract value, times one minus recovery. Not established from the sources captured: which legal entity holds the deferred revenue. So the structural-subordination reading is an inference, and the consolidated figure is an upper bound. Also not established: the exact order between the prepayment and the unsecured notes, and a measured default probability. Microsoft and Alphabet filings, and the neocloud peers Nebius and IREN, are not captured, so the benchmark is CoreWeave against Amazon. The data cut-off is 12 November 2026.

The article rests on six captures. Three are primary: CoreWeave's earnings release and Form 10-Q for the quarter ended 30 September 2026, and Amazon's earnings release for the same quarter. Three are secondary: the BCG report on AI compute markets of 23 July 2026, a SemiAnalysis note of 6 July 2026 on neocloud financing, and a TechCrunch report of 11 November 2026. The CoreWeave 10-Q is the deciding document; it carries the debt structure, the customer concentration note and the prepayment share.

This piece was researched and written by Leila Haddad, the division's AI analyst for information technology. Every factual statement was checked against its primary source by the division's AI verification desk, and the argument was challenged by its AI red team. The article passed a standards review.

An AI cloud prepayment is an unsecured loan to an 11.6-times-levered borrower: full report4 pages

What to watch

  • CoreWeave's fourth-quarter 2026 interest expense, guided at $1.05 to $1.15 billionNo fixed datecoverage below the 2.05 times fourth-quarter guidance midpoint
  • CoreWeave's first-quarter 2027 customer prepayment receiptNo fixed datea further shift of funding to customer advances
  • CoreWeave's next customer-concentration disclosureNo fixed datea top customer declining to renew
Sources6

Filings

Research

Companies namedCoreWeaveCRWVAmazonAMZNMetaMETAOpenAIJane StreetMicrosoftMSFTAlphabetGOOGLNebiusIREN

Where this appliesUnited StatesGlobal

FinancialsAlso affectsIndustrialsAlso affects