GPU-backed debt: pricing the fourth year

A graphics processing unit (GPU) can still run, attract customers and retain a resale value while earning too little to repay the debt used to buy it. That gap is becoming an important part of AI infrastructure finance. A lender has to look beyond the equipment’s ability to compute: what will the next customer pay, what will it cost to deliver the service, and how much cash could be recovered if operations stop?

Recent transactions make the question tangible. On August 10, 2026, CoreWeave announced a $2.6 billion credit facility with a maturity of roughly five years, supported by customer contracts lasting about three years on average. The financing therefore extends into a period when contracts will have to be renewed or capacity leased to someone else. CoreWeave presents the structure as a way to serve customers seeking shorter commitments. ⁷

Lambda’s early-October transaction offers a useful contrast. Its new $1.008 billion financing capacity relies on commercial commitments that, in Morningstar DBRS’s analysis, remove exposure to volume and re-leasing risk. Lenders can finance GPUs for a longer period when customers contractually absorb part of the economic risk. Installation, acceptance and service availability still matter. ^{12 13}

The fourth year is our vantage point. In the example developed here, it is when the first customer can leave while debt payments continue. Actual dates differ across transactions. Understanding the collateral means following that timing gap all the way to the bank account.

Four clocks attached to one machine

Nvidia makes a coherent industrial argument: installed equipment can serve different customers, improve through software and find new uses after its original contracts expire. In an August 11 post, Jensen Huang argued that the A100’s economic life could approach a decade. AWS’s catalogue, checked on October 5, still offers P4 instances based on that generation. A commercial service thus remains available well after newer models have arrived. ^{2 4}

That observation leaves the essential credit questions open. An older generation still offered today may include units installed at different times. A catalogue demonstrates availability; it does not reveal the cash earned by a particular vintage of machines. A lender needs the realised price, billed quantity and remaining margin over the years in which payments fall due.

Accounting runs on a separate clock. CoreWeave’s 2025 accounts use an estimated six-year life for technology equipment. Depreciation allocates its cost over time, affecting reported profit and carrying value. Debt principal follows the repayment schedule agreed with creditors. ⁵

The commercial clock depends on the customer, its commitment and renewal terms. The financing clock depends on the loan agreement. Between the two, the operator promises to turn computing capacity into cash receipts. Technical life provides a fourth measure: how long the equipment can keep producing. An informative credit assessment preserves all four rather than asking one useful-life assumption to answer every question.

CoreWeave puts renewal risk into the loan agreement

The August transaction, called DDTL 5.5, is a delayed-draw term loan. Its $2.6 billion headline is committed capacity available subject to conditions. In the quarterly report filed on August 12, CoreWeave said it had drawn $1.2 billion after quarter-end. That historical amount does not establish the outstanding balance on October 5. ¹⁴ The SEC filing sets final maturity at September 1, 2031. The floating-rate borrowing option carries a margin of 5.50 percentage points over Term SOFR, a dollar interest-rate benchmark. The margin is added to the benchmark; 5.50% is not the all-in interest rate. ⁶

The press release explains the commercial opportunity. The contract shows how lenders deal with the end of the first customer commitment. Section 2.09(b)(iii) requires a prepayment when expiring agreements are not renewed or replaced in a way that meets the stipulated criteria. Resizing must restore two projected tests: debt-service coverage of at least 1.35 in each projected monthly period after the deadline and a contract-value ratio of at least 2.40. ⁸

The economics are straightforward. When future receipts become smaller or less dependable, they support less debt. The financing then calls for the exposure to be reduced before lenders have to rely on selling the servers. For shareholders, a difficult customer negotiation can turn into a need for fresh cash.

For periods not yet under contract, projected coverage may use the borrower’s good-faith assumptions consistent with the base-case model. The protection therefore also depends on the quality of the commercial forecasts. ⁸

Part of the timetable remains confidential. The public definition of the Renewal Deadline places it after the contract’s term ends, but redacts the length of that interval. The five-business-day prepayment period runs after that contractual deadline, whose precise position is withheld. Some schedules, including the amortisation grid, are also unavailable. The public text therefore cannot support a reconstruction of the actual loan’s month-by-month balance. ⁸

There is another significant protection: CoreWeave, the parent company, guarantees payment and performance. Subject to applicable law, the agreement allows creditors to pursue that obligation without first exhausting the physical collateral. The parent’s resources belong in the credit analysis. The guarantor under this agreement is CoreWeave, not Nvidia. ⁹

Lenders consequently have several forms of protection: operating receipts, security interests and a claim against the group. Their economic strength can nevertheless move together. A widespread deterioration in compute economics could affect several projects while weakening the parent’s ability to support them. A guarantee improves legal recourse; its financial usefulness also depends on the guarantor’s resources when payment is required.

Lambda illustrates another way to lend for longer

Lambda’s October 1 announcement describes three deployments for two investment-grade purchasers of computing capacity. The release’s summary box lists a fixed interest rate of 6.78%, with semi-annual coupons, final maturity on May 30, 2033, and a fully amortising repayment profile. Security covers the servers, related infrastructure and contracted cash flows. ¹²

On October 2, Morningstar DBRS assigned A (low) ratings to Lambda Compute I LLC and its senior debt. The agency emphasises take-or-pay contracts, under which customers pay for reserved capacity on the agreed terms, and the absence of exposure to re-leasing risk. Its assessment still identifies delivery, equipment acceptance and service-availability risks. DBRS also flags the parent’s non-investment-grade credit profile, particularly during installation, and a debt-service reserve weaker than usual for project finance. The vehicle’s rating therefore does not constitute a rating of the group. This is a solicited credit opinion, with the entity’s participation, rather than insurance against default. ¹³

The structure provides a counterweight to the idea that a GPU necessarily becomes unfinanceable after three or four years. A lender can accept a later maturity when commercial commitments and protections sufficiently cover the intervening period. It is underwriting the customer’s ability to pay, the enforceability of its commitment and the operator’s ability to perform.

Those contracts relocate risk. A customer committed to paying for capacity that later becomes less competitive absorbs some economic obsolescence. It may accept that exposure in exchange for availability or suitable commercial terms. Customer default or a service failure can still damage the loan, even when actual usage has little influence on contracted receipts.

CoreWeave’s March financing supplies a further comparison: a maximum $8.5 billion facility, maturity in 2032 and a floating margin of 2.25 percentage points. Its presentation describes a structure without general recourse to the parent, subject to limited guarantees. The gap to August’s 5.50-point margin is 325 basis points. That compares two different transactions. Contracts, protections and market conditions vary together, preventing the difference from being treated as a clean price for renewal risk. ^{10 11}

What carrying value leaves unanswered

Now consider an entirely hypothetical project, with all figures in nominal US dollars. It buys $100 million of equipment, financed with $70 million of debt and $30 million of equity. The loan is repaid in five equal annual instalments at a fixed 10% rate. The first three years of commercial revenue are secured in the opening scenario. These are teaching assumptions, not the repayment terms of CoreWeave or Lambda.

The annual payment, including principal and interest, is $18.47 million. After the third payment, $32.05 million of principal remains. That amount depends on how quickly the loan has amortised, not on how many years an accountant expects to depreciate the equipment.

With a six-year straight-line accounting life and zero accounting residual value, the equipment still carries a $50 million book value at the end of year three. A four-year life would leave $25 million; a ten-year life would leave $70 million. The physical equipment and the debt are identical in all three cases. Only the accounting assumption changes, setting aside tax and covenant effects linked to the accounts.

The $50 million balance-sheet figure does not tell us what a buyer would pay. That requires comparable transactions, the equipment specification and condition, its operating environment and the cost of putting it back into service. Going-concern economic value requires a forecast of future receipts and expenses. Carrying value, continuation value and sale proceeds answer different questions.

A lender may finance equipment whose early receipts quickly repay its purchase cost. It may also allow substantial debt to remain beyond the first customer agreement, backed by additional recourse or an expectation of strong resale proceeds. The amount already repaid is as important to the risk as the machine’s eventual lifespan.

Better hardware changes the next rental negotiation

Technical progress cuts both ways. It can create new uses and expand the market for computing. It can also offer an existing customer more output for the same budget or power allocation. An operator running an older generation must then find a price at which its customer can remain competitive.

Nvidia has advertised substantial reductions in cost per token for Rubin relative to Blackwell in certain mixture-of-experts configurations, which activate only part of a model for a given task. Those are supplier performance claims under specified technical conditions. They do not establish an automatic change in rental prices or a depreciation curve for existing equipment. ³

The unit being compared matters. A GPU-hour can include different amounts of memory, networking, storage or support across services. For the user, cost per completed task may be more informative than the hourly hardware rate. Changes in models, software and bundled services can therefore distort a superficial comparison of advertised prices.

Installed equipment also retains advantages: migration costs, tested software, immediate availability and a familiar architecture. AWS itself describes P4 as an integrated compute and networking offering. That service layer can support customer retention. Its price, operating costs and ability to preserve demand belong in the cash forecast. ⁴

These are the inputs a financing discussion should examine: realised prices after discounts, actual billed hours, firm commitments, cash operating expenses and further capital needed to keep the service competitive. A price on a website describes an offer. A lender needs the receipts of the specific fleet being financed.

The fourth year of our project

Return to the $100 million project. For its first three years, it receives $50 million annually and pays $18 million in operating expenses, covering operations, energy, maintenance and service. That leaves $32 million before debt service. The model excludes taxes, working-capital changes and additional capital expenditure. It is a simplified cash-flow illustration, not a complete shareholder-return calculation.

Dividing $32 million by the $18.47 million payment produces coverage of 1.73 times. That is the idea behind the debt service coverage ratio (DSCR): cash available for debt service divided by debt payments over the relevant period. Above one, the project can make its payment under the assumptions. At one, the payment uses all available cash.

Suppose that at renewal the average price falls by 25% and billed hours decline by 20%. Revenue drops from $50 million to $30 million. The reductions multiply: the project retains 75% of its initial price and 80% of its billed volume, preserving 60% of the original revenue. The resulting revenue decline is 40%.

Keeping cash costs at $18 million in this conservative scenario leaves only $12 million for debt service. Coverage falls to 0.65 times. The scheduled payment now exceeds available cash by $6.47 million each year. Years four and five retain the payment obligations of the original contract; meeting them requires funding that shortfall. No component has to fail for this problem to appear. The fleet remains usable, customers still pay, and the financing nevertheless requires adjustment.

Fixed cash costs are an assumption worth challenging. Variable expenditure may fall with activity, while leases, staff and energy commitments can remain rigid. Every $1 million actually saved adds $1 million to available cash in this model. The published parameters and equations allow the result to be recalculated with different costs, prices and volumes. The scenario also applies to an exposed commercial period: a robust take-or-pay contract can protect receipts against part of a decline in utilisation.

Two changes that compound

The sensitivity matrix identifies the economic threshold without making a price forecast. To meet the original loan payment and pay $18 million of operating costs, the project needs at least $36.47 million of annual revenue. That boundary is simply costs plus scheduled debt service.

With 80% of billed hours retained, the break-even price is approximately 91.16% of its initial level, giving a DSCR of 1. A price decline of roughly 8.84% then consumes the cushion. More billed hours allow a lower price; a more flexible cost base shifts the boundary as well.

This explains why a reported change in hourly prices tells only part of the story. A lower price may attract more customers and preserve revenue. Moving into another use case may improve occupancy. Conversely, more available capacity could depress both realised prices and hours sold. That joint sensitivity should be tested. Assigning probabilities would require additional market evidence.

Financing rates also matter. In the model, reducing the rate from 10% to 8% lowers the annual payment, but still leaves the stressed $12 million cash flow short of the obligation. Cheaper debt can help an operator without offsetting a much larger decline in revenue. The sensitivity analysis uses several hypothetical rates; none is presented as a current financing offer.

Resizing the loan before selling the servers

At the end of year three, our project still owes $32.05 million and has two annual payments remaining. Consider a possible remedy: the lender agrees to size the debt against the lower $12 million annual cash flow. We assume an illustrative prudential buffer of 1.35 times annual coverage.

The allowable new annual debt payment becomes $8.89 million. At 10%, two such payments support approximately $15.43 million of principal. Moving from $32.05 million to $15.43 million would require a $16.62 million paydown, under these assumptions alone. That money could come from a shareholder, asset sales or another source. Whether it is available is a separate question.

Even at coverage of exactly one, consuming all future cash, two $12 million payments would support only $20.83 million. The constraint reflects both the remaining balance and the time available to repay it. Extending maturity can reduce the instalment, but asks the lender to rely on more years of uncertain prices, usage and expenses.

This calculation illustrates the economic purpose of a resizing provision. It does not apply CoreWeave’s actual covenant, which has its own definitions, monthly periods, additional contract-value test and cure rights. Our annual model deliberately keeps those features separate. The point is that a loan can become too large for the next contract’s cash flow well before its collateral loses all value. ⁸

There is an important numerical counterpoint: the first three years generated $40.60 million of after-debt cash in this model. If all of it had been retained and remained available, it could cover the $16.62 million paydown without external funding. Taxes, investment and distributions, excluded from this simplified illustration, would determine how much cash actually remained.

Earlier distributions then become important. A project that retained cash has more resources to cross the renewal period than one that distributed everything. Reserves, distribution restrictions and mandatory prepayments determine how much money remains inside the structure when the next commercial agreement must be negotiated.

A sale produces a different number again

A creditor considering enforcement looks at recoveries after the costs needed to obtain them. In our alternative sale scenario, immediately after the third loan payment, assume an offer of $35 million for the fleet. Selling costs equal to 10% of gross proceeds, plus $2 million for removal, transfer or preparation, leave $29.50 million.

Against $32.05 million of outstanding principal, the gap is $2.55 million before any guarantor payment. At a $45 million gross sale price, the same assumptions leave $38.50 million; at $25 million, they leave $20.50 million. These are scenarios, not estimates of secondary-market prices for a particular GPU generation.

The sale replaces continued operation. The same equipment cannot earn all future rental income after it has already been sold. A continuation valuation may include a terminal sale after the intervening receipts, but the timing has to be consistent. Our comparison simply isolates an exit at the end of year three.

Recovery also depends on what moves with the equipment. An installed, powered and operable fleet may retain customer contracts. Dismantled equipment needs transport, installation and new business. CoreWeave’s disclosures describe security interests over several project components, while the August agreement includes efforts relating to collateral access. The economic value of those rights depends on their terms and execution. ^{8 10 14}

Market liquidity becomes especially important if several operators need to sell at once. Weaker rental economics could reduce buyers’ purchasing capacity just as more used equipment arrives on the market. That is a plausible channel for correlated losses to examine in a stress scenario. The sources reviewed do not establish a quantified market depth or a common recovery rate across GPU fleets.

A guarantee needs a payment rule

Nvidia’s August discussion of residual-value support refers to up to 25% of an opportunity, assessed case by case. It describes a limited possibility rather than a standard guarantee contract applying to every GPU loan. ²

A cap is insufficient to calculate a recovery. The lender needs the reference value, payment trigger, eligible equipment, covered period and beneficiary. A repurchase commitment at a specified price differs from a capped indemnity for certain losses. The provider’s capacity to pay must also be assessed.

We therefore assign no Nvidia support payment to either transaction examined here. The reviewed documents describe their own protections. CoreWeave’s parent guarantee is identifiable and public; Lambda’s credit case rests partly on customer commitments. Combining these separate signatures would obscure the exposure being measured. ^{9 13}

The data that would make this asset class easier to assess

Reuters’ October 1 report captures differing views on equipment longevity and appropriate lending caution. The public documents examined here provide a more specific answer than choosing one universal useful life: credit risk varies with contracted receipts, the amount already repaid and the recourse retained by lenders. ^{1 16}

The most valuable next dataset would track installation vintages: commissioning date, purchase cost, receipts under the initial contract and at renewal, cash maintenance costs and net resale proceeds. Billed hours would be needed to distinguish falling rates from lost business. A blended average combining old and new hardware generations would conceal important movements.

Contracts would complete the picture: termination rights, service obligations, transferability, lien priority, signed guarantees and actual amortisation schedules. Delayed or restructured transactions are as informative as successful placements. Without that history, a model can locate thresholds, but cannot honestly assign a frequency to crossing them.

GPUs can retain economic usefulness for years, and some structures already finance that life through firm commercial commitments. The exposed point is where those commitments end before the debt does. Year four does not condemn a machine in advance. It requires an answer to a question the purchase price cannot settle: who will still pay enough to keep it working?

Reproduce the scenarios

All model amounts are in millions of nominal US dollars. The hypothetical rate is annual effective, with payments at year-end. Years 1 to 5 describe an illustrative project, without identifying a real GPU cohort.

Annual debt service is A = D × r / (1 − (1 + r)^(-n)), with D = 70, r = 0.10 and n = 5. Annual interest is opening debt multiplied by the rate; principal repaid is the payment minus interest. Carrying value follows a separate straight-line depreciation of the 100 equipment cost, with zero accounting residual value.

Revenue is 50 × price index / 100 × billed-hours index / 100. Cash available for debt is revenue minus 18 of cash operating costs. Annual DSCR divides that cash by scheduled debt service. Prices and billed hours are hypothetical realised values; the matrix colours mark deterministic thresholds, without probabilities.

For resizing after year three, permissible annual debt service is 12 / 1.35. Discounting two payments at the unchanged hypothetical 10% rate gives supported debt. Required paydown is the positive difference from outstanding principal. The 40.60 of early cash could fund that paydown if retained and still available.

The alternative sale takes place immediately after the third payment. Net proceeds are max(0, gross price × 0.90 − 2). Lender recovery is capped at outstanding principal; any shortfall and residual after debt are shown separately. No rental receipts after the sale are added.

Taxes, working-capital changes, further capital expenditure, initial financing fees, accrued interest, penalties and other creditors are excluded. The annual l0g calculations implement neither CoreWeave’s monthly contractual tests nor Lambda’s financing agreements.

Download the full parameters and results in JSON, annual repayment schedule, 81 matrix points, depreciation variants, interest-rate sensitivity and sale scenarios. These files make the rounding inspectable and allow results to be recalculated from the equations above.

Documentary research closed on October 5, 2026. Financing terms come from the identified agreements and communications; committed limits, dated draws and rating opinions remain distinct. Some clauses and schedules are redacted. No current resale price or sector-wide average risk is measured. No interviews with the parties were conducted for this edition.

Further reading on AI financing

AI debt, contracts and guarantees follows the obligations behind infrastructure spending. Nvidia’s financing platforms explains the difference between a mobilisation target and capital already committed.

Sources and documents

1. S01 · Reuters · 2026-10-01Nvidia’s bet that its chips can finance the AI boom gets a Wall Street reality check. Economic longevity and residual value; views attributed to lenders and investors.
2. S02 · Nvidia / Jensen Huang · 2026-08-11NVIDIA AI Factory Compute Is Becoming an Investable Asset Class. Nvidia’s case for redeployment, software improvements and economic longevity; possible residual-value support of up to 25% per opportunity.
3. S03 · Nvidia · 2026-01-05NVIDIA Kicks Off the Next Generation of AI With Rubin. Rubin performance claims, including up to a tenfold reduction in cost per token in selected MoE cases; a supplier comparison.
4. S04 · Amazon Web Services · undatedAmazon EC2 P4 Instances. P4 instances using A100 GPUs; confirms that a commercial service remains available.
5. S05 · CoreWeave / SEC · 2026-03-02Form 10-K, year ended December 31, 2025. Note 1, Property and Equipment: straight-line depreciation and an estimated six-year life for technology equipment.
6. S06 · CoreWeave / SEC · 2026-08-10Form 8-K, event of August 7, 2026. Item 1.01: $2.6 billion DDTL 5.5 commitment, September 1, 2031 maturity, Term SOFR plus 5.50 percentage points and parent guarantee.
7. S07 · CoreWeave / SEC · 2026-08-10Exhibit 99.1 : CoreWeave closes $2.6 billion loan facility. Roughly five-year debt maturity and three-year average customer contracts; CoreWeave’s stated aim of serving a broader customer base.
8. S08 · CoreWeave / SEC · 2026-08-10Exhibit 10.1 : DDTL 5.5 Credit Agreement. August 7 agreement: renewal prepayment clause, Renewal Deadline, financial covenants, cure rights and collateral-access provisions; projected thresholds of 1.35 and 2.40.
9. S09 · CoreWeave / SEC · 2026-08-10Exhibit 10.2 : Parent Guarantee Agreement. August 7 guarantee, Article II, section 2.1, especially paragraph (g): payment and performance without first exhausting collateral.
10. S10 · CoreWeave · undatedDDTL 4.0: $8.5 Billion First-of-a-Kind Facility. March 2026 presentation, pages 4–5: $8.5 billion, roughly six years, SOFR plus 225 basis points, assets and contracts, limited parent recourse.
11. S11 · CoreWeave / SEC · 2026-03-31Exhibit 99.1 : $8.5 billion financing facility. DDTL 4.0 announcement and pricing; documentary comparison with the August facility.
12. S12 · Lambda · 2026-10-01Lambda closes $1 billion senior secured fixed rate financing. $1.008 billion delayed-draw commitment. Summary box: 6.78% fixed rate, semi-annual coupons, May 30, 2033 maturity, full amortisation, three deployments and two investment-grade offtakers.
13. S13 · Morningstar DBRS · 2026-10-02Morningstar DBRS Assigns Credit Ratings of A (low) to Lambda Compute I LLC. DBRS opinion on Lambda Compute I LLC and its senior debt: take-or-pay contracts, no volume or re-leasing exposure; installation, acceptance, service and parent risks, with a weaker debt-service reserve.
14. S14 · CoreWeave / SEC · 2026-08-12Form 10-Q, quarter ended June 30, 2026. Note 10, security interests; Note 16, DDTL 5.5 and $1.2 billion drawn after closing. Technology, utilisation and renewal risk disclosures.
15. S15 · Nvidia · 2026-08-10Financing platforms designed to mobilize over $500 billion of third-party capital. Financing partnerships and a target for mobilising third-party capital over time.
16. S16 · Reuters / WHTC · 2026-10-01Nvidia’s bet that its chips can finance the AI boom gets a Wall Street reality check. Full, bylined Reuters report republished by WHTC on October 1; differing views on economic life and lending caution.