When AI compute gets cheaper, who pays for yesterday’s GPUs?

Reporting cut-off: 31 October 2026. CoreWeave financial statements: quarter ended 30 June 2026. Public pricing pages: checked on 31 October 2026.

A computer chip can work perfectly, attract paying customers and still turn into a poor investment. Imagine a company that has just installed servers for artificial intelligence. It paid for the hardware at today’s prices. Its lenders expect repayment over several years. In the meantime, customers buy hours of compute. If customers begin paying less for those hours, the financing bill does not fall with the rental rate.

This resembles a hotel built with borrowed money, then forced to discount its rooms. AI infrastructure adds a particular complication: successive processors do more work, the software consuming that work changes rapidly, and customers can reserve capacity years ahead. The price of a GPU-hour, the revenue promised under a contract and the cost of completing an AI task are three different things.

That distinction determines who absorbs a price shock. A company selling primarily under binding capacity agreements may be well insulated from this month’s spot market. A smaller operator trying to fill unreserved machines may feel every reduction. Lenders, in turn, care about the cash produced while principal and interest are still due. The important fault line is between the capacity already sold and the capacity that must eventually find its next customer.

Method and scope

Two advertised prices for the same eight-H100 instance type, with different availability terms. Public North American rate card consulted on 31 October 2026. This is a cross-sectional comparison, not evidence of a historical fall or of revenues actually collected.

The market has more than one price

CoreWeave’s North American catalogue lists an eight-GPU HGX H100 instance at $49.24 per hour on demand, against $19.71 per hour on spot. The spot quote is approximately 60% lower. It would be tempting to call that a collapse in the price of eight identical processors. Yet those prices coexist on the same page. Spot capacity is interruptible; a firm reservation or a conventional on-demand service has different delivery characteristics. An informed comparison has to consider interruption risk, commitment length, configuration and the services bundled into the offer. S05

A second distinction matters even more. Model developers may buy raw GPU time, while API users pay for tokens, requests or completed tasks. Software can become dramatically more efficient and drive down the price of completing a task without forcing a particular GPU rental price to move in parallel. The OECD estimates that its quality-adjusted price index for text-generating models fell by nearly 80% between January 2024 and April 2026. The index measures AI-model services at comparable levels of capability. It says nothing directly about the rent charged for a particular H100 cluster. S08

Contemporary readings underline the importance of method. At 09:29 UTC on 31 October, 06.ai’s tracker of posted offers showed $3.72 per H100 GPU-hour on demand and $2.33 spot, rounded from trimmed averages covering 25 and 9 providers respectively. Silicon Data’s dated neocloud H100 index showed $2.92 on 31 October, against $2.77 in its 2 October observation. These dated tables have different samples and normalisation rules. They describe offers and commercial indices; establishing a provider’s realised revenue requires its own contracts and billed hours. S06 S07

So the useful question is narrower and harder: what will the next customer actually pay for a specific machine, for how many billable hours, and how much cash will remain after operating expenses? Public evidence available by late October 2026 does not establish a universal, simultaneous collapse in realised rental revenues across the AI compute sector. The sharp decline in quality-adjusted model-service prices is real; translating that observation into GPU owner cash flows requires another layer of evidence.

CoreWeave itself supplied a counterpoint on 17 September: management said it continued to contract new compute capacity at higher prices, discussing contracts lasting three to six months. This interested-party statement describes the group’s own capacity and agreements, with a different scope from a standardised H100 rental index. It highlights the resilience of some segments. The risk examined here concerns less profitable renewals, with exposure varying by product and contract. S13

A contract buys time

CoreWeave makes this point particularly well. In its second-quarter Form 10-Q, the company says 98% of revenue came from customer commitments, including some capacity delivered ahead of the formal commitment start. Revenue for the three months ended 30 June 2026 was $2.575 billion; net interest expense was $640 million. The two line items are not a cash debt-coverage calculation. But they show the scale at which contract protection and financing costs must be examined together. S01 S12

Under a binding capacity commitment, an AI lab may have agreed to pay for access regardless of whether a newer generation of chips later offers a better deal. The operator still has to deliver an agreed standard of availability and performance. If the market price falls, its existing receipts may remain intact while the customer’s opportunity cost rises. The first party to bear some of the technology-price risk is therefore often the buyer that signed the reservation.

That protection expires. When customers renegotiate, they can compare new chips, alternative cloud providers, switching costs and performance per dollar. Contracts may also contain service-level provisions or adjustments whose effect depends on the exact language. Long-dated contracted revenue is contingent on actual delivery, and customers’ creditworthiness still matters. CoreWeave reported approximately $104 billion of revenue backlog at 30 June 2026. Its own definition includes remaining performance obligations and other expected revenue under committed contracts, subject to delivery and availability conditions. Backlog is neither cash already collected nor a guarantee of future margin. S12

The customer base adds a separate concentration risk. In the same quarter, three customers identified only as A, B and C represented 36%, 26% and 10% of revenue respectively: 72% in total. Public filings do not allow every letter in that table to be assigned unequivocally to a named technology firm. A single large customer’s weaker appetite for new capacity, tougher renewal terms or financial distress could nevertheless matter far more than a move in the average GPU quote. S01

Debt repayments follow a fixed schedule

At 30 June, CoreWeave disclosed $35.551 billion of future debt principal payments. Its maturity schedule covers the remaining six months of 2026, each of 2027–30 and later years. These are principal obligations, not a forecast of all future interest expense, data-centre lease payments or investments. Nor is the schedule a prediction that no refinancing will occur. What it makes unambiguous is that debt must be settled in dollars even if selling the next generation of GPU-hours turns out to be less profitable. S01

Method and scope

CoreWeave principal debt maturities at 30 June 2026, USD billions. “Beyond 2030” aggregates several years rather than a single payment. Future interest and lease obligations are excluded. Source: 10-Q, Note 10.

Accounting creates another calendar. CoreWeave’s 2025 annual report gives technology equipment a six-year estimated useful life, having extended its estimate from five to six years in 2023. Depreciation spreads a historical cost through reported earnings. It does not specify what a used processor would fetch in an orderly sale or how much money that machine will earn next year. At 30 June 2026, the company held $33.823 billion of gross technology equipment, a category that includes more than GPU chips. Calling that figure the market value of its H100 fleet would misstate the filing. S02 S01

Many AI infrastructure financings try to align loan repayment with committed customer cash flows. Funds may be drawn as facilities are delivered. Assets and receivables can be pledged to lenders; the parent company may or may not provide a full repayment guarantee. Those details determine how a shortfall propagates when the next commercial contract produces less cash than the first.

Two CoreWeave loans explain why the creditor matters

On 30 March 2026, a CoreWeave subsidiary agreed an $8.5 billion delayed-draw financing facility, known as DDTL 4.0, with a maturity in 2032. The parent’s guarantee covers specified bad acts and certain customary exceptions rather than ordinary, full recourse to the group. The filing specifies a 1.15x minimum debt-service coverage ratio subject to the contract’s definitions and timing. In broad terms, lenders rely primarily on project assets and cash flows, while the precise exceptions still matter. S04

On 7 August 2026, another subsidiary signed a $2.6 billion facility, DDTL 5.5, to finance infrastructure linked to customer commitments, including GPU servers. This time, CoreWeave, Inc. unconditionally guaranteed the obligations, and the agreement specifies a 1.35x minimum debt-service coverage ratio once the test begins to apply. The final maturity is 1 September 2031. The $8.5 billion and $2.6 billion figures are committed facility sizes on signing, not amounts necessarily drawn that day. S03

That difference determines the lender’s recovery options. Under a genuine non-recourse financing structure, the lender’s ordinary recovery is tied to the collateral and the financed entity, subject to contractual carve-outs. A full parent guarantee gives the lender another source of repayment. Shareholders, bondholders, secured creditors, customers and suppliers may ultimately experience different losses depending on the order of claims and the liquidity each party retains. Repayment duties rest with the borrower and the guarantors named in the agreements. S03 S04

Method and scope

Simplified movement of revenue and credit risk, illustrating different parent-company recourse structures. Conceptual mechanics, not an allocation of actual cash flows to a named financing facility.

Eight GPUs, one price cut and a vanishing cushion

Consider a completely hypothetical eight-GPU operation. The machinery and setup cost $320,000. Its owner borrows $240,000 at a fixed 10% interest rate, repaid over four equal annual instalments, and spends $85,000 a year on cash operating costs excluding debt service. The initial rental rate is set at $4 per GPU-hour. These are teaching assumptions, not CoreWeave expenses, a supplier quote or a claim about the purchase price of a particular H100 rack.

At 75% billable utilisation, eight machines sell 52,560 GPU-hours in a 365-day year. At four dollars, annual revenue reaches $210,240. Subtracting the assumed operating expenses leaves $125,240 to service an annual debt payment of approximately $75,713. Simplified debt-service coverage is 1.65 times. It measures cash before debt divided by the contractual annual instalment, excluding taxes, working-capital movements, replacement capital expenditure and other complexities.

Now let the realised rate fall to $3 per hour, while utilisation holds steady. Revenue declines 25%; cash available for debt service drops to $72,680. Coverage becomes 0.96 times, leaving an annual $3,033 shortfall before further adjustments. If billable utilisation also falls to 65%, coverage declines to 0.68 times and the annual gap widens to about $24,057. Every server can remain operational throughout this exercise. The operator is still selling compute, yet must fund a gap, reduce its debt or seek a restructuring.

Method and scope

Hypothetical M01: eight GPUs, $320,000 investment, $240,000 debt at 10% over four equal annual payments and $85,000 annual operating cash costs. Simplified coverage excludes taxes, working capital and replacement investment.

At 75% utilisation, the hypothetical operation needs roughly $3.06 per GPU-hour to meet the entire scheduled annual payment with cash. Requiring an illustrative 1.35x coverage cushion raises the threshold to about $3.56. The 1.35 threshold is a teaching reference linked to a real facility’s headline ratio. Its actual legal definition, adjustments, testing dates and scope are specific to DDTL 5.5; this simplified model cannot reproduce the covenant. S03

l0g / COMPUTE AND REPAYMENT / LOCAL CALCULATOR

The rental rate that covers the debt payment

A fictional eight-GPU project with a four-year loan. Change the rate, billed hours and cash costs to follow the money available for repayment.

Change the three assumptions

Fictional scenario inputs

Realised rate per GPU-hour (USD)

$4.00

Share of hours actually billed (%)

75 %

Annual cash costs excluding debt (USD)

$85,000

Restore fictional example

Rate: $1 to $8 per GPU-hour. Billed hours: 20% to 100%. Annual cash costs: $30,000 to $150,000.

Simplified debt-payment coverage

1.65×

Debt payment covered

1.65×

Annual revenue

$210,240

Cash available for debt service

$125,240

Cash after the annual payment

+$49,527

Rental rate for 1.00× coverage

$3.06

Rental rate for 1.35× coverage

$3.56

Assumptions and method

Fixed inputs: 8 GPUs, $320,000 upfront investment, including $240,000 of debt at 10% fixed interest, repaid in four equal annual instalments. A year contains 8,760 hours. Cash operating costs are held constant.

Theoretical annual payment: $75,713.

The annuity consists of four equal year-end payments: debt × rate ÷ [1 − (1 + rate)⁻⁴]. Revenue = GPUs × annual hours × billed share × rate. Cash available = revenue − operating cash costs. Coverage = cash available ÷ annual debt payment. The rate threshold divides costs plus the target debt payment by billed GPU-hours.

This teaching model uses fictional inputs, distinct from CoreWeave’s accounts and legal covenants. It excludes taxes, fees, equipment replacement, refinancing, resale proceeds and working-capital changes. The 1.35× threshold is a teaching benchmark.

Calculations use double precision. Cash amounts are displayed to the nearest dollar and rate thresholds to two decimal places after calculation. Independent rounding can create small differences between displayed figures.

Calculations run in this page. Inputs stay in memory while it is open.

The calculator changes the price, billable utilisation and cash costs without sending data to any server. In a real centre, power spending and some operating expenses might fall with usage while rent, personnel and contractual commitments are less flexible. The model deliberately holds cost assumptions explicit. Actual outcomes would also depend on interest-rate exposure, taxes, GPU refresh spending, liquidity restrictions, customer default provisions and the terms of individual contracts.

How a price shock travels

A price drop affects every link differently. API users may perform more tasks for the same budget. Research labs can train more, switch to better-performing hardware or reconsider how much capacity they reserve. A cloud operator must keep its installed fleet occupied while paying for electricity, buildings, network gear and creditors. Secured lenders examine both the cash pledged by a contract and the recovery value of the collateral if payments stop.

A spot-market decline first hits capacity that is not protected by committed revenue. Its influence then spreads through new deals, renewals, valuations and, potentially, financing tests. New chips can accelerate that pressure if they perform more useful work per dollar or watt. At the same time, migrating software and moving workloads between clusters carry costs, and an older processor can remain economically useful for less demanding jobs. Economic obsolescence is a negotiation, not a switch flipped on a fixed birthday.

The larger risk comes when price and paid utilisation decline together. A 25% price cut and a 20% fall in billable hours reduce variable revenue by 40%: 0.75 multiplied by 0.80 equals 0.60. With fixed expenses and scheduled debt payments unchanged, the cushion shrinks rapidly. The company can try to sell more hours, improve efficiency, inject equity, sell assets or renegotiate with lenders. Each route changes who supplies the missing money.

Method and scope

Hypothetical M01: eight GPUs, $320,000 investment, $240,000 debt at 10% over four equal annual payments and $85,000 annual operating cash costs. Simplified coverage excludes taxes, working capital and replacement investment.

The strongest objection to the bearish case

Cheaper computation can make AI a better business. It can unlock products that were previously too expensive, stimulate inference demand and keep more machines busy. Improvements in software may lower the cost per completed task even if GPU hourly prices remain firm. Reserved customers preserve near-term revenue, and established infrastructure can continue serving workloads long after a newer chip launches. The OECD also points out that agents can consume dramatically more tokens per task as applications become more capable, so falling token prices do not guarantee falling overall spending on AI. S08

In January and March 2026, the Bank for International Settlements described the rapid migration of AI infrastructure financing towards borrowing, including private credit, while stopping short of declaring that a financial crisis was underway. Projects depend on substantial expected future revenues, but a disappointing return on investment is not automatically a payment default. Equity buffers, contractual cash flows, liquidity, collateral and maturity structure decide when a weak investment turns into a creditor problem. S09 S10

CoreWeave itself had $5.524 billion in unrestricted cash and cash equivalents at 30 June and reported $15.553 billion of total liquidity after including, among other items, borrowing availability subject to facility terms. For the first six months of 2026, operating activities generated $3.663 billion in net cash, investing activities used $14.874 billion and financing activities provided $13.985 billion. Those are six-month cash flows, not annual rates or spending attributable to one GPU model. They illustrate the scale of infrastructure expansion and funding, while also demonstrating why a one-line claim that the business is currently unable to pay would go beyond the evidence. S01

The accounts remain a 30 June snapshot. CoreWeave subsequently completed a $4.2 billion convertible-note issue, documented in a 22 September filing, with maturity in 2033. Drawdowns, repayments and other intervening transactions prevent a simple addition from producing current total debt. Third-quarter results are scheduled for 10 November, so the maturity chart in this analysis explicitly retains its June scope. S15 S14

So who writes the cheque?

The borrower remains responsible for its debt while it is performing. Revenue from committed customers may continue even when spot prices fall. If the financed operation cannot generate enough cash, lenders can enforce covenants, seek early repayment where permitted or negotiate changes. Collateral may be sold or refinanced, but its proceeds depend on the machines, their configuration and the market. Where the parent guaranteed repayment, that parent may have to provide funds. Where recourse is limited, lenders are exposed more directly to the project, subject to the contractual exceptions.

The meaningful price is therefore the one the next agreement will actually deliver in cash, multiplied by billable usage and reduced by the expenses that cannot be avoided, with the remaining principal and interest schedule laid alongside it. Public filings reveal a great deal about debt, accounting lives, major contracts and guarantees. They reveal much less about negotiated discounts, fleet-by-fleet utilisation, secondary-market recoveries and renewal economics. Those missing observations are precisely where a credible stress analysis must remain modest.

The question surrounding the AI boom is often framed as a contest over how much tomorrow’s GPU will be worth. For lenders and infrastructure owners, there is a more useful question: how much cash will today’s equipment earn before the last instalment falls due?

Indicators for testing the risk

• Matched, dated rental prices for the same hardware and service level, with an auditable vendor panel.
• Renewal rates, contract terms, billable fleet utilisation and concessions granted to large customers.
• Principal repayments, conditional liquidity, collateral protections and the exact scope of parent guarantees.
• Changes in estimated useful lives and impairment tests, kept separate from realised secondary-market transaction values.

Further reading

The fourth year of GPU-backed debt examines renewal and residual-value risk. Nebius’s 364-day Citi facility illustrates another financing timetable. The glossary definitions of GPU, DSCR, DDTL and covenant explain the contract mechanics.

Method and limits

Research cut-off: 31 October 2026. Financial statements, signed facilities, posted offers and management statements retain their own dates and scopes. M01 is a hypothetical constant-cost scenario, separate from CoreWeave’s accounts and negotiated covenants. Realised prices, utilisation by hardware generation and resale values require additional data. The simulator runs locally in the browser.

Sources and references

Sources accessed on 31 October 2026. Commercial indices and statements are attributed to their producers, with limitations provided below.

1. S01 · CoreWeave · 2026-08-12 · Accessed: 31 October 2026Form 10-Q for quarter ended 2026-06-30

   98% of quarterly revenue related to committed customers; 2,575 USDm of Q2 2026 revenue; 640 USDm net interest expense; 35,551 USDm future debt principal (see Note 10); 33,823 USDm gross technology equipment; cash and liquidity as disclosed.

   Passages and scope

   Consolidated fiscal quarter and balance-sheet snapshot. Future interest, capital commitments and operating leases excluded from the debt principal schedule. Not H100-level cash flows. Header date of report is not filing date.

2. S02 · CoreWeave · 2026-03-02 · Accessed: 31 October 2026Form 10-K FY2025

   Six-year estimated useful life for technology equipment accounting category.

   Passages and scope

   Estimated accounting useful life is not market residual value or the commercial life of all GPUs.

3. S03 · CoreWeave · 2026-08-10 · Accessed: 31 October 20268-K : DDTL 5.5

   2.6 USDbn committed delayed-draw facility, due 2031-09-01, SOFR + 5.50pp on term SOFR tranche, unconditional CoreWeave Inc parent guarantee, minimum DSCR 1.35x after specified start trigger.

   Passages and scope

   Committed facility amount differs from drawn amount. Real covenant includes negotiated definitions and timing; M01 is not a covenant recreation.

4. S04 · CoreWeave · 2026-03-31 · Accessed: 31 October 20268-K : DDTL 4.0

   8.5 USDbn committed facility, due March 2032, limited parent-company guarantee for specified bad acts, minimum DSCR 1.15x after contractual start.

   Passages and scope

   Limited recourse exceptions mean not zero recourse. Two facilities should not be compared as if borrower, collateral, covenants and risk were identical.

5. S05 · CoreWeave · Undated · Accessed: 31 October 2026Public cloud pricing / HGX H100 8 GPU NA

   North America 8 x H100: $49.24 per whole instance-hour on-demand, $19.71 per whole instance-hour spot.

   Passages and scope

   Cross-sectional difference in terms and availability, not a time-series price fall; not executed contract or realised cash.

6. S06 · Silicon Data · 2026-10-31 observation · Accessed: 31 October 2026H100 Rental Price Index neo-cloud SDH100RT

   Dated commercial index: USD2.92/GPU-hour on 31 October 2026, compared with 2.77 on 2 October. Dated table used rather than undated FAQ.

   Passages and scope

   Not interchangeable with 06.ai, a CoreWeave instance quote, or CoreWeave revenue per GPU. Indexed page may also contain a generic non-dated FAQ reading.

7. S07 · 06.ai · 2026-10-31T09:29:32Z · Accessed: 31 October 2026GPU Compute Price Index

   Posted-price trimmed mean H100: on demand 3.7189 USD/GPU-hour (25 providers),spot 2.3301 (9 providers),intraday snapshot; not executed trades.

   Passages and scope

   Offer quotes not realised transactions; capacity weighting unavailable, spot interruptible, long-term negotiated prices often private.

8. S08 · OECD · 2026 · Accessed: 31 October 2026Artificial Intelligence markets : quality-adjusted text model price

   Aggregate quality-adjusted text-model price index declined nearly 80% Jan2024 to Apr2026.

   Passages and scope

   Text-model service index, not GPU rent, owner revenue, or GPU depreciation.

9. S09 · BIS · 2026-01-07 · Accessed: 31 October 2026BIS Bulletin 120, Financing the AI boom: from cash flows to debt

   Analyses rising debt share in AI infrastructure build-out and risk transmission.

   Passages and scope

   Broad macro analysis; no claim that the industry has already defaulted.

10. S10 · BIS · 2026-03 · Accessed: 31 October 2026BIS Quarterly Review, Financing the AI infrastructure boom: on- and off-balance sheet borrowing

    Documents on- and off-balance sheet financing mechanisms and risks.

    Passages and scope

    Not evidence of future crisis probability or actual GPU price collapse.

11. S11 · l0g.fr · 2026-10-05 · Accessed: 31 October 2026GPU à crédit : le prix de la quatrième année

    Earlier coverage of hardware residual value, contract life, accounting life and debt maturity.

    Passages and scope

    Existing l0g work, not independent primary proof. New article concentrates on pricing formation and transmission.

12. S12 · CoreWeave · 2026-08-11 · Accessed: 31 October 2026Q2 2026 earnings release

    Q2 revenue 2.575 USDbn and around 104 USDbn revenue backlog per company definition.

    Passages and scope

    Backlog not cash, and includes capacity subject to delivery and other conditions; SEC filing S01 more authoritative for financials.

13. S13 · CoreWeave · 2026-09-17 · Accessed: 31 October 2026CoreWeave, new capacity contracted at higher prices

    17 September 2026 company release: management says it continues to sign new capacity at higher prices, discussing three-to-six-month contracts.

    Passages and scope

    Interested-party statement about the group’s own capacity and contracts; distinct from a standardised H100 GPU-hour rental series.

14. S14 · CoreWeave · 2026-10-27 · Accessed: 31 October 2026CoreWeave, third-quarter reporting calendar

    27 October 2026 announcement: third-quarter results scheduled for 10 November.

    Passages and scope

    Explains the June balance-sheet reference date; later financing transactions retain their own dates and scopes.

15. S15 · CoreWeave · 2026-09-22 · Accessed: 31 October 2026CoreWeave, September convertible notes

    22 September 2026 filing: completed $4.2 billion convertible-note issue, maturing in 2033.

    Passages and scope

    Includes $3.7 billion plus a $0.5 billion exercised option. Current total debt also depends on intervening drawdowns and repayments, so this issue is shown separately from June principal.