The AI Compute Funding Loop: How CoreWeave's $132.5 Billion Backlog, $9 Billion of Prepayments and $37-39 Billion of Capex Feed Each Other

1. First, the shape of the loop ​

Lay out CoreWeave's Q3 2026 figures in time order and a clear chain appears:

1. Win long-term contracts. Backlog rose from $104 billion at end-June to $132.5 billion at end-September (the earlier figure is a media number, not a company disclosure in this release).
2. Convert contracts into money. The quarter produced roughly $4.2 billion of convertible notes, a $2.6 billion DDTL 5.5 facility the company describes as its first designed for shorter-duration customer contracts, and an inaugural at-the-market program for up to 35 million Class A shares. The CEO separately cited roughly $9 billion of customer prepayments and deposits expected within the next three months.
3. Convert money into capacity. $9.441 billion of property, equipment and capitalized software purchased in the quarter; net property and equipment rose from $30.557 billion to $56.778 billion in nine months, and operating lease right-of-use assets from $8.231 billion to $18.922 billion.
4. Convert capacity into delivery. Active power doubled from the start of the year to above 1.7 GW, contracted power reached 4.6 GW, with a year-end target above 1.85 GW.
5. Recognize revenue after delivery. Revenue of $3.661 billion in the quarter, a sequential gain above $1 billion.
6. Revenue becomes new contracts and new financing capacity. The loop returns to step one.

This is not a metaphor; it is a closed circuit verifiable with numbers. Each step maps to a line item: backlog in the footnotes, prepayments in payables or deferred revenue, financing in debt and equity, capex in the cash flow statement, capacity in operating metrics, revenue in the income statement.

2. The arithmetic: why a 62% margin still produces a $709 million loss ​

From a cash standpoint the loop is self-consistent: adjusted EBITDA of $2.256 billion at a 62% margin means 62 cents of operating cash gross profit per dollar of revenue. Does that cover interest? Net interest was $846 million in the quarter, and adjusted EBITDA is 2.7x that. On an EBITDA basis, interest is affordable.

Yet GAAP net income is -$709 million. The gap comes from three lines: $1.889 billion of depreciation and amortization, $846 million of net interest, and $83 million of income tax. Take adjusted EBITDA of $2.256 billion, subtract depreciation of $1.889 billion, and adjusted operating income is $379 million; subtract interest of $846 million and the result is negative.

That is the accounting truth of a heavy-asset model: EBITDA is the cash-layer number that looks good, depreciation is the real wear at the asset layer, interest is the real cost at the capital layer. Stack the three and a 62% EBITDA margin still maps to a -19% net loss margin.

The follow-up question is whether depreciation lives are set correctly. If GPUs are economically usable for three years while accounting depreciates them over six, depreciation is understated and EBITDA overstated. CoreWeave does not disclose GPU depreciation lives, so this cannot be tested. It is a shared accounting grey zone across neoclouds, and this site flagged it in the CoreWeave versus Nebius comparison (full comparison).

3. The premise that keeps the loop spinning: customers paying ahead ​

The most important and least discussed link is customer prepayment.

Under ordinary commercial logic, customers do not pay billions upfront for services not yet delivered. AI compute is different: supply is constrained by power and advanced packaging capacity, and leading model companies value certainty more than price. Prepaying to lock capacity therefore becomes an acceptable structure.

CoreWeave cites roughly $9 billion of customer prepayments and deposits expected within three months. That money sits between a deposit and deferred revenue: it locks capacity for the customer and provides cheap capital to the supplier — cheaper than convertibles, less dilutive than equity.

Nebius's statements offer a second verification: deferred revenue rose from $1.578 billion at end-2025 to $9.409 billion at the end of September 2026 ($1.446 billion current plus $7.964 billion non-current), an increase of about $7.8 billion in nine months. That is 1.29x its single-quarter capex of $7.3 billion.

Both sets of numbers say the same thing: on this racetrack, customers have moved from buying a service to buying an option on capacity. That explains why neocloud revenue growth so far exceeds traditional cloud: growth does not come from gradually adding customers, but from signing large multi-year capacity contracts in single transactions.

4. Three breakpoints in the loop ​

The loop looks elegant, but it has three explicit breakpoints.

Breakpoint one: delivery delay. Every step depends on the previous one completing on time. Grid interconnection delays, GPU delivery delays, data hall construction delays — any slip moves revenue recognition later while interest does not move. Active power of 1.7 GW against 4.6 GW contracted means 2.9 GW still needs energizing over coming quarters: a concrete execution exposure.

Breakpoint two: customer concentration. The release notes serving 9 of the world's 10 leading frontier model providers but does not disclose revenue share per customer. If a large part of the $132.5 billion backlog comes from a few customers, any one of them adjusting procurement hits revenue, backlog and financing capacity at once — because the loans are secured on those contracts.

Breakpoint three: rates and refinancing. Net interest was $846 million in Q3 and is guided to $1.05-1.15 billion in Q4. Full-year adjusted operating income guidance is $1.275-1.375 billion. Put the two together: a single quarter's interest already approaches a full year's adjusted operating income. Net income is therefore highly sensitive to rates, and if the refinancing window tightens, step two of the loop stalls.

The breakpoints couple. Delivery delays erode customer confidence; weaker confidence shrinks prepayments; smaller prepayments force heavier reliance on debt; more debt raises interest; higher interest compresses reinvestment capacity. That coupling is why valuation opinions on these companies diverge so widely.

5. Two paths: the debt loop and the equity loop ​

Place Nebius alongside and a second solution to the same loop appears.

CoreWeave spins the loop faster using debt and prepayments, at the cost of $4.0 billion of equity and $846 million of interest. Nebius spins it more conservatively using equity and convertibles, at the cost of deeper dilution and a capex-to-revenue ratio of 7.44x.

Neither path is "wrong"; they encode different risk preferences. The first bets on delivery capability plus customer creditworthiness; the second bets on financing windows plus ramp speed.

6. What the loop means for the industry ​

First, compute is being financialized. Once capacity can be sold years ahead in large contracts and used as collateral for financing, compute acquires properties closer to a commodity or infrastructure asset. That is why neocloud valuation methods increasingly resemble REITs or utilities rather than tech companies.

Second, the locus of power in the value chain is moving. For two years the scarcest inputs were advanced nodes and HBM; now power, interconnection permits and the customer relationships needed to sign long contracts are equally scarce. Thailand's newly approved data center policy framework — minimum standards on environment, water and power, and zoning — is the policy-side response to that shift.

Third, accounting and disclosure will be the next battleground. GPU depreciation lives, the credibility of backlog, and customer concentration determine how much substance a neocloud's earnings carry. None of the three main players (CoreWeave, Nebius, Oracle) discloses all of it, so cross-company comparability stays limited.

7. Boundaries and inferences not to draw ​

• All figures come from CoreWeave's November 10, 2026 press release and earnings presentation and Nebius's November 12, 2026 release; the $104 billion prior backlog figure is a media number, not disclosed in this release;
• Backlog includes company estimates and is not contracted cash; it cannot be spread linearly across future years;
• The $9 billion of customer prepayments is a forward-looking "expected" statement, not a booked amount;
• Q3 capex of $9.441 billion is a cash flow statement figure with a different basis from the $37-39 billion full-year guidance; it cannot be annualized as quarter times four;
• Adjusted EBITDA is a non-GAAP measure and cannot be mixed with GAAP net income;
• This article does not judge whether CoreWeave's or Nebius's valuation is justified and makes no forecast for either company's next quarter;
• The "three breakpoints" are an analytical framework built from public statements, not the company's own disclosed risk factors.

Sources ​

• CoreWeave press release: CoreWeave Reports Strong Third Quarter 2026 Results (November 10, 2026)
• CoreWeave Q3 2026 earnings presentation (PDF)
• Nebius press release: Nebius reports third quarter 2026 financial results (November 12, 2026)
• Related on this site: CoreWeave vs Nebius, Q3 2026, CoreWeave Q3 2026 key data, Oracle, CoreWeave and Nvidia: AI infrastructure financing compared