GPU and neocloud compute financing to 2028
Who lends against the chips, what it costs, and why the lease is replacing the loan.
Two years ago a rack of GPUs was collateral almost nobody would lend against without an equity cushion and a rate in the teens. On 24 October 2026 a single borrower signed a binding commitment for a master finance lease carrying roughly $40 billion of principal payments — more than its entire existing debt stack — against graphics processing unit pods it does not own and will not own at the end.
This is a forecast rather than a market summary: seven structural shifts with the evidence under each, the loop that drives them, five scenarios with odds at December 2027 and December 2028, what commoditises and what keeps its scarcity, and seventeen dated calls a named public source will settle. Every reading carries a source grade. Every judgment carries a confidence. The odds are judgments too, and the page says so.
A lease the borrower never owns, against chips the market already discounts
One GPU lease, as committed
Aggregate principal payments expected under the lease schedules of a single master finance lease committed to on 24 Oct 2026.
Effective rate, oldest to newest recourse facility
One operator's own debt table, recourse against recourse: 15% on its March 2028 facility against 9% on the newest, entered 15 May 2026. The 7% facility in the same table is non-recourse.
H100 spot as a share of on-demand list
FD estimate: $1.00 ÷ $4.50, from one provider's public price feed and published list price, read 5 Nov 2026.
Data-center securitisation outstanding
YTD 2026 against a year-end 2028 premise, both attributed to Barclays research by a trade body.
By the end of 2028 the chips are financed on a lease the borrower never owns, and the building is financed on a bond
The thing that changes in between is not the appetite. It is which legal form the money arrives in, and that form decides who carries the residual.
We put the lease-led market at 38% by December 2028 against 24% a year earlier, and today's secured-loan market down from 40% to 17%. The second panel is how a new large GPU purchase is funded. Each column sums to 100, and both are Frontier Desk judgments rather than model output.
The financing market
Which shape compute financing is in at each date. Frontier Desk judgments, not model output
| Scenario | Dec 2027 | Dec 2028 |
|---|---|---|
| Secured-loan marketToday's shape continues: delayed-draw term loans sized off the depreciable cost of the equipment | 40% | 17% |
| Lease-led marketA third-party lessor owns the chips; the operator pays rent and never takes the residual | 24% | 38% |
| Vendor-and-offtaker fundedThe chipmaker and the capacity buyer carry the financing between them | 15% | 16% |
| RepricingResidual values disappoint, advance rates fall, and the cost of capital on chips widens again | 12% | 20% |
| Credit eventA neocloud or an offtaker fails large enough to shut the market to everyone below investment grade | 9% | 9% |
How a new large GPU purchase is funded
The dominant structure for financings signed in each year. Today the secured facility still leads
| Structure | Dec 2027 | Dec 2028 |
|---|---|---|
| Secured loan or delayed-draw facilityA term loan against the equipment, sized off its depreciable cost | 44% | 25% |
| Finance lease from a third-party lessorThe lessor buys the chips and leases them; the operator never owns them | 22% | 36% |
| Unsecured corporate debt or convertsThe balance sheet carries the chips, and the bond market carries the balance sheet | 20% | 22% |
| Vendor, offtaker or equity-linkedThe chipmaker, the capacity buyer or a joint venture funds the purchase | 14% | 17% |
Odds are Frontier Desk judgments at the two horizons, set against the triggers and early markers that decide each one in Scenarios, and the method is in Sources and method.
Bottom line
Eight forward-looking judgments, each with the evidence under it and the confidence we hold it at.
- 01The lease is winning because it is the only structure that scales past the borrower's own lien covenants.Space Exploration Technologies' June 2026 senior notes restrict additional liens to 7.5% of consolidated total assets. Total assets were $254,045 million at 30 September 2026 — FD estimate: 7.5% × $254,045M = $19,053M of lien capacity. The GPU financing it committed to on 24 October 2026 is about $40 billion of aggregate principal payments, or FD estimate: $40,000M ÷ $19,053M = 2.1 times the whole basket. A secured loan of that size could not be granted without consent. A master finance lease in which the lessor acquires the equipment and leases it to the operator may not need the basket at all, because the borrower never owns the asset and grants no lien over it. That is the mechanism as we read it, and on the same reading it would generalise to other large operators with investment-grade-style negative covenants.Confidence: high. Both halves are quoted from the same registrant's own filings with the SEC three weeks apart: the covenant and the balance sheet from the Form 10-Q for the quarter ended 30 September 2026, and the lease terms from the Form 8-K filed 26 October 2026 under Item 8.01. The inference that the covenant is the binding reason is ours, and the filings do not state a motive. The mechanism is ours too: the indenture is not quoted in the 10-Q excerpt, and lien baskets of this kind sometimes reach finance-lease obligations, sale-leasebacks or attributable debt, so whether this one leaves the structure outside it is not something the filings settle.
- 02The cost of capital on a rack of GPUs fell by two-fifths on recourse terms in three years, and the fall is nearly over.CoreWeave's own debt table splits into recourse and non-recourse, and the rate path runs differently down each side. On the recourse side it falls from 15% on the DDTL 1.0 facility of July 2023, through 11% on the DDTL 2.0 facility of May 2024, to 9% on the DDTL 2.1, 3.0 and 5.0 facilities — and DDTL 5.0, the newest facility in the table, was entered on 15 May 2026 and still prices at 9%. FD estimate: (15% − 9%) ÷ 15% = a fall of 40% on comparable terms in three years. The 7% rate sits under the table's own Non-recourse debt heading. It belongs to the March 2026 DDTL 4.0 facility, entered by an indirect subsidiary and priced at daily compounded SOFR plus 2.25% or Treasury plus 2.00%, and recourse on it is limited to that subsidiary's assets rather than running to the parent. The filing groups it with the company's other facilities at bankruptcy-remote, special-purpose consolidated subsidiaries formed to hold the financed infrastructure and the related customer contracts. FD estimate: (15% − 7%) ÷ 15% = a fall of 53% in three years, of which the last two points are bought with structure rather than with a cheaper view of the chips. Cipher Digital priced $850.0 million of 0.00% convertible senior notes due 2032 on 14 October 2026. Against that, the five-year Treasury went from 3.74% on 2 January 2026 to 4.69% on 4 November 2026 — the base rate is rising as fast as the spread is falling. We put it at 29% (judgment) that a new US GPU equipment facility prices inside Treasury plus 150 basis points by 31 December 2027. The spread compression has further to run than the all-in rate does.Confidence: high on the spread, moderate on the all-in rate. The facility rates and the recourse split are both quoted from a registrant's own quarterly filing, which files the two sets of facilities under separate headings, and the Treasury series is the Treasury's own published daily curve. Which of the two points between 9% and 7% the structure buys and which the market buys is not something the filing apportions, so the 40% is the comparison to lean on and the 53% is the one to read with the structure attached. The forward path of the base rate is not something this report forecasts, and the judgment is deliberately held below a half.
- 03The chips are financed over five years and depreciated over six, and the public spot market says both are optimistic for the oldest generation.The contemplated master finance lease amortises principal at one-twentieth of the financed cost of each tranche per quarter — FD estimate: 1/20 per quarter × 4 quarters = 20% a year, so the lessor is fully repaid in five years and takes no residual risk. CoreWeave depreciates technology equipment over six years, having moved it from five effective 1 January 2023. Nebius publishes live spot prices through an API its documentation states requires no authentication: read on 5 November 2026, preemptible H100 cleared at $1.00 a GPU-hour against a $4.50 on-demand list price, while B300 cleared between $8.40 and $10.70 against a $9.50 list. FD estimate: $1.00 ÷ $4.50 = 22% of list for the older chip; $10.70 ÷ $9.50 = 113% for the newest, taking the top of that read's B300 range. The H100 ratio is the one to lean on: it has been stable across reads, while the newest generation's range moves within the day, so the 113% describes one timestamped maximum rather than a level a reader will see at any hour. Nvidia took a $0.4 billion charge for excess H200 inventory and purchase obligations in the first half of fiscal 2027 because, in its own words, demand for H200 products diminished.Confidence: high. Four independent primary readings point the same way — a lease that refuses residual risk, an accounting life, a live public price feed and a chipmaker's own inventory charge. What they do not settle is whether the spot discount is obsolescence or oversupply, and that distinction decides the repricing scenario.
- 04The building and the chips inside it are now financed on two different clocks, and the gap is the whole risk.Applied Digital signed an approximately 15-year lease with three five-year renewal options to a high investment-grade hyperscaler on 5 June 2026, for the full 210 MW of critical IT load at one campus, representing about $5.2 billion of contracted revenue over the base term with delivery expected in the first half of calendar 2028 — FD estimate: $5,200M ÷ 210 MW ÷ 15 years = $1.65M per MW a year. The same company depreciates electrical infrastructure over 15 years and the building over 39, against 3 to 5 years for networking and electrical equipment. A data center is a 39-year asset wrapped around a 5-year one. Every financing question in this market reduces to who is holding which of those two clocks when the music changes.Confidence: high. Both the lease and the depreciation schedule are quoted from the same registrant's Form 10-Q for the quarter ended 31 August 2026, and the arithmetic is shown. The reading that the clock mismatch is the central risk is ours.
- 05Securitisation is the exit the whole stack is building toward, and equipment is the part it has not reached yet.The Structured Finance Association, writing on 23 July 2026, reports Barclays research putting outstanding data-center securitisation at $61 billion year-to-date 2026 against $4 billion in 2020, and a Barclays conditional in which, if outstanding data-center securitisations reach $180 billion by year-end 2028, they could comprise roughly 20% of the esoteric ABS market and 15% of the single-asset/single-borrower CMBS market. The note frames the $180 billion as the premise of that conditional rather than as a forecast. The same note says GPUs, servers and other computing equipment are increasingly financed through specialised lending facilities, and that equipment-backed financings may emerge as a complementary ABS sector. Four months later the largest GPU lease on the public record was signed with an entity managed by a firm whose business is originating assets for exactly that market. 34% (judgment) that a public GPU- or AI-equipment-backed securitisation of $500 million or more prices by 31 December 2028.Confidence: moderate. The issuance figures are a trade body quoting a bank's research rather than a dataset we can recompute, which is a C-grade reading carried with its attribution. The direction is corroborated by the lease structure itself, which is built to be sold.
- 06The largest buyer of compute is moving the residual off its own balance sheet, and disclosing the size of what it moved.Meta's maximum exposure to loss in its unconsolidated data center ventures went from $45,952 million at 31 December 2025 to $68,404 million at 30 September 2026, of which maximum residual value guarantees went from $28,091 million to $41,493 million. It holds 20% of each venture, the parties have committed about $27 billion for the Louisiana venture and about $14 billion for the Texas venture, the leases run an initial four years with renewals to a total of up to 20, and no liability has been recorded because residual value guarantee payments are not probable. Quarterly capital expenditure including finance lease principal was $34.91 billion. The residual question is not being answered. It is being allocated.Confidence: high. Every figure is quoted from one registrant's Form 10-Q for the quarter ended 30 September 2026, including the statement that no liability has been recorded. Whether the allocation is prudent is a question this report asks rather than answers.
- 07Supply is committed further ahead than demand is contracted, and that is a financing problem before it is a demand problem.Nvidia's supply and capacity commitments reached $279 billion at 26 July 2026, from $119 billion the quarter before, with $92 billion falling due in the remainder of fiscal 2027, $87 billion in fiscal 2028 and $88 billion in fiscal 2029. Its own equity investment commitments total $25 billion, $18 billion of it in the remainder of fiscal 2027. Data center revenue was $89,023 million in the quarter. Somebody has to finance the gap between a chip being built and a chip being paid for, and the three candidates on the public record are the chipmaker, the offtaker and a lessor. 56% (judgment) that a chipmaker or hyperscaler discloses a new equity, guarantee or backstop commitment of $5 billion or more to a compute operator by 31 December 2027.Confidence: moderate. The commitments and the revenue are quoted from one registrant's quarterly filing. Reading a two-and-a-third-fold rise in supply commitments as a financing signal rather than a demand signal is an inference, and the filing supports both readings.
- 08Power is the covenant nobody wrote three years ago and everybody writes now.CoreWeave's DDTL 4.0 facility requires interest rate hedges covering at least 95% of reasonably anticipated outstanding floating-rate borrowings, and the agreement includes power cost hedging requirements. That is a lender treating electricity as a priced input it will not let the borrower leave open. The reason is on the other side of the same build: PJM's 2028/2029 Base Residual Auction cleared at the $325.00 per MW-day cap on 14 July 2026, against $28.92 four delivery years earlier, and the product of cleared supply and clearing price was $16.4 billion against $2.2 billion. Our companion forecast on data-center power and interconnection carries that argument in full.Confidence: high. The covenant is quoted from the borrower's own filing and the auction figures from the market operator's own published results report. The link between them is an inference, but it is the inference the covenant exists to express.
The baseline
Where compute financing stands on 5 November 2026. This band is collection, not analysis: every row is a reading from a named public source, graded A, B or C, with no inference attached. The inferences are everywhere else on the page.
Two years ago a GPU was an asset lenders discounted heavily and priced in the teens. The readings below are the position the rest of this report forecasts from: a cost of capital that has collapsed, a lease structure large enough to double a borrower's debt in one signature, a public spot market that clears the oldest chips at about a fifth of their on-demand list price, and a securitisation exit that has not yet opened for equipment.
One GPU lease, as committed
Aggregate principal payments expected under the lease schedules of a single master finance lease committed to on 24 Oct 2026.
Effective rate, oldest to newest recourse facility
One operator's own debt table, recourse against recourse: 15% on its March 2028 facility against 9% on the newest, entered 15 May 2026. The 7% facility in the same table is non-recourse.
H100 spot as a share of on-demand list
FD estimate: $1.00 ÷ $4.50, from one provider's public price feed and published list price, read 5 Nov 2026.
Data-center securitisation outstanding
YTD 2026 against a year-end 2028 premise, both attributed to Barclays research by a trade body.
Every reading, with its source
Publicly observableGrades describe the source, not the conclusion. A — record: a filing with a securities regulator, a court docket, a public dataset or API, a published paper, or a government report. B — party: an organisation's own published page, terms, documentation or announcement about itself. C — crowd: aggregators, trade bodies summarising research they did not perform, and surveys of unnamed practitioners.
Confidence describes Frontier Desk's inference from those readings: high where two or more independent sources agree and none contradicts, moderate where the direction is corroborated but the magnitude is not, low where a single source or an analogy carries it. The odds on this page are judgments of the same kind, not the output of a model, and they are stated as judgments wherever they appear.
| Reading | Source | Figure | Grade |
|---|---|---|---|
| Aggregate principal payments expected under one master finance lease for GPU pods | Form 8-K, Item 8.01, filed 26 Oct 2026 for an event of 24 Oct 2026 | Approximately $40 billion | A · record |
| Term of each lease schedule under that lease | The same Form 8-K | Approximately five to five and one-quarter years from each commencement date | A · record |
| Quarterly principal amortisation under that lease | The same Form 8-K | One-twentieth of the financed cost of the tranche, plus a fixed rate on the outstanding balance | A · record |
| That borrower's total debt and finance leases | Form 10-Q for the quarter ended 30 Sep 2026 | $38,989 million principal | A · record |
| That borrower's lien covenant under its June 2026 senior notes | The same Form 10-Q | Additional liens restricted to 7.5% of consolidated total assets, which were $254,045 million | A · record |
| That borrower's June 2026 unsecured notes | The same Form 10-Q | $25,000 million, five tranches, 11.7-year weighted-average maturity, 5.350% to 6.650% coupons, 5.855% weighted average | A · record |
| One neocloud's total debt principal | Form 10-Q for the quarter ended 30 Jun 2026 | $35,551 million, with $6,184 million due in 2027 and $4,416 million in 2028 | A · record |
| Effective interest rates across that neocloud's recourse facilities | The same Form 10-Q | 15% on the Mar 2028 facility; 11% on the May 2024 facility; 9% on three later facilities; 10% on four senior notes; 9% on the 2032 EUR notes; 2% on two converts; 11–12% on OEM and term financing | A · record |
| Effective interest rates across that neocloud's non-recourse facilities | The same Form 10-Q | 7% on the Mar 2026 delayed-draw facility and 9% on OEM financing, $3,663 million in total net of discount and issuance costs | A · record |
| Which of that neocloud's delayed-draw facilities carry recourse to the parent | The same Form 10-Q | All of them except the Mar 2026 facility, which sits at an indirect subsidiary and on which recourse is limited to that subsidiary's assets | A · record |
| Pricing of that neocloud's March 2026 delayed-draw facility | The same Form 10-Q | $8.5 billion, at daily compounded SOFR + 2.25% or Treasury + 2.00%, with a 0.50% undrawn fee | A · record |
| What sizes that facility | The same Form 10-Q | The depreciable cost of computing equipment, projected debt service coverage and project-level conditions | A · record |
| Hedging that facility requires | The same Form 10-Q | Interest rate hedges on at least 95% of anticipated floating-rate borrowings, plus power cost hedging requirements | A · record |
| That neocloud's depreciable life for computing equipment | Form 10-K for the year ended 31 Dec 2025 | Six years, moved from five effective 1 Jan 2023 | A · record |
| A zero-coupon convertible priced by a listed compute operator | Form 8-K, Item 8.01, filed 14 Oct 2026 | $850.0 million of 0.00% convertible senior notes due 2032 | A · record |
| One operator's 15-year lease to a high investment-grade hyperscaler | Form 10-Q for the quarter ended 31 Aug 2026 | 210 MW of critical IT load, about $5.2 billion of contracted revenue over the base term, delivery H1 CY2028 | A · record |
| That operator's total minimum contracted lease payments | The same Form 10-Q | $35,731,035 thousand, of which $26,580,159 thousand falls after FY31 | A · record |
| That operator's depreciable lives | The same Form 10-Q | Networking and electrical equipment 3–5 years; electrical infrastructure 15; building 39 | A · record |
| The largest capacity buyer's maximum exposure in unconsolidated data center ventures | Form 10-Q for the quarter ended 30 Sep 2026 | $68,404 million, from $45,952 million at 31 Dec 2025 | A · record |
| Of that, maximum residual value guarantees | The same Form 10-Q | $41,493 million, from $28,091 million; no liability recorded, as payments are not probable | A · record |
| That buyer's quarterly capital expenditure including finance lease principal | The same Form 10-Q | $34.91 billion for the three months ended 30 Sep 2026 | A · record |
| The chipmaker's supply and capacity commitments | Form 10-Q for the quarter ended 26 Jul 2026 | $279 billion, from $119 billion the previous quarter; $366 billion of total future commitments | A · record |
| That chipmaker's data center revenue in the quarter | The same Form 10-Q | $89,023 million of $96,221 million total | A · record |
| That chipmaker's charge for excess H200 inventory and purchase obligations | The same Form 10-Q | $0.4 billion in the first half of fiscal 2027, as demand for H200 products diminished | A · record |
| Public preemptible GPU spot prices, one provider, all regions | The provider's unauthenticated spot-price API, read 5 Nov 2026 | H100 $1.00; H200 $0.90–$1.85; B200 $3.50–$5.60; B300 $8.40–$10.70 per GPU-hour | B · party |
| The same provider's published on-demand list prices | The provider's own pricing page | H100 $4.50; H200 $5.40; B200 $8.50; B300 $9.50 per GPU-hour | B · party |
| Five-year US Treasury yield | The Treasury's own published daily yield curve | 4.69% on 4 Nov 2026, from 3.74% on 2 Jan 2026; low 3.51% on 27 Feb, high 5.09% on 30 Sep | A · record |
| Data-center securitisation outstanding | Trade body note of 23 Jul 2026, attributing Barclays research | $61 billion YTD 2026, from $4 billion in 2020; a conditional $180 billion by year-end 2028 | C · crowd |
| Global data center capital expenditure through 2028 | The same note, attributing Morgan Stanley research | About $2.9 trillion, of which about $1.5 trillion to be financed externally | C · crowd |
| PJM capacity clearing price for delivery year 2028/29 | The market operator's own auction results report | $325.00 per MW-day at the cap, against $28.92 for 2024/25 | A · record |
One transaction carries more of this forecast than any other, and it is worth reading in the filer's own words. On 24 October 2026 Space Exploration Technologies Corp. entered into a binding commitment letter to enter into a master finance lease with an entity managed by Apollo Global Management, Inc., under which the lessor will acquire and lease to the company certain graphics processing unit pods and related equipment for use at data centers selected by the company.
The Equipment will be delivered in tranches, each subject to a separate lease schedule with a term of approximately five to five and one-quarter years. Rent is payable quarterly in advance and generally consists of one-twentieth of the financed cost of the tranche, plus an amount at a fixed rate per annum on the outstanding lease balance, plus certain ongoing costs. The fixed rate will be determined at pricing. The company expects aggregate principal payments to total approximately $40 billion. Closing is expected before year end and the schedules are expected to commence in the first half of 2027.
Four things in that paragraph decide the next two years. The lessor owns the equipment, so no lien is granted and no covenant basket is used. The amortisation is one-twentieth a quarter, so the lessor is whole in five years and underwrites no residual. The rate is fixed and set at pricing, so the borrower takes the rate risk between commitment and close. And the data centers are selected by the company rather than the lessor, so the power risk stays with the operator while the equipment risk leaves it.
Hold those readings together and the next two years are already implied. A cost of capital that fell this far cannot fall much further while the base rate rises. A structure that refuses residual risk will keep winning until somebody is paid to take it. A public spot market that prices the oldest chips at a fifth of list will eventually be priced into an advance rate. And a securitisation market that has reached the buildings but not the equipment will either reach the equipment or stop the lease market where it stands.
Structural shifts underway
Seven shifts are already running. Each separates what the sources say from what we infer from them: the evidence first with its grade, then the reading with its confidence, then how fast, then the specific thing that would reverse it. These are the inputs to the scenarios.
From the secured loan to the finance leaseAccelerating
Speed: 12 to 24 months
What the sources say. A Form 8-K filed 26 October 2026 under Item 8.01 reports a binding commitment letter for a master finance lease under which the lessor will acquire and lease GPU pods, with approximately $40 billion of aggregate principal payments expected and lease schedules expected to commence in the first half of 2027. The same registrant's Form 10-Q for the quarter ended 30 September 2026 shows total debt and finance leases of $38,989 million and a covenant restricting additional liens to 7.5% of consolidated total assets of $254,045 million.
Evidence grade: A on both readings — two filings by the same registrant three weeks apart.
What we read into it. A lease is not a cheaper loan. It is a different legal object, and the difference is what makes it scale. The lessor takes title, so the lessee grants no security interest and consumes no lien basket. The rent is an operating obligation rather than funded debt for most covenant definitions. And the lessor's credit analysis is on the rent stream and the equipment, not on the lessee's capital structure. Every operator whose unsecured debt carries a lien limit now has a reason to put the next tranche of chips through a lessor instead of a bank.
Confidence: high on the structure, moderate on the generalisation. The structure is quoted. That other operators will copy it is our inference, supported by the covenant arithmetic but by only one transaction so far.
What would reverse it. A ratings agency or a covenant package that counts master finance lease obligations as funded debt for leverage tests, which would remove the covenant advantage and leave only the pricing difference.
The spread compresses while the base rate risesRunning, and close to its limit
Speed: Already happened; the next leg is the base rate
What the sources say. One neocloud's own debt table shows a 15% effective interest rate on its facility maturing March 2028, 11% on its May 2024 facility and 9% on three facilities signed later, all of them recourse, and files its March 2026 facility — priced at daily compounded SOFR + 2.25% or Treasury + 2.00% with a 0.50% undrawn fee — under a separate non-recourse heading. A listed compute operator priced $850.0 million of 0.00% convertible senior notes due 2032 on 14 October 2026. The five-year Treasury yield published by the Treasury moved from 3.74% on 2 January 2026 to 4.69% on 4 November 2026, with a low of 3.51% on 27 February and a high of 5.09% on 30 September.
Evidence grade: A on all three — two registrant filings and the Treasury's own daily curve.
What we read into it. Spread and base rate are moving in opposite directions, and the all-in cost is the sum. A fixed-rate GPU facility at Treasury + 200 — the non-recourse pricing in that table, not the recourse pricing — would have priced near 5.74% in January and near 6.69% in early November. The borrower in the largest lease on the record has agreed a fixed rate to be determined at pricing, which means it is carrying roughly five weeks of rate risk on a $40 billion book. FD estimate: a straight-line five-year amortisation of $40 billion leaves an average outstanding balance of $40bn ÷ 2 = $20 billion, so the 158 basis points between this year's five-year low and high costs 1.58% × $20bn = $316 million a year, or × 5 years = about $1.58 billion over the term.
Confidence: high on the readings, moderate on the estimate, which assumes the whole $40 billion draws evenly and amortises straight-line as the filing describes. The filing gives the amortisation but not the draw schedule.
What would reverse it. A sustained fall in the five-year Treasury below 4%, which would let the all-in rate resume falling and keep the secured-loan market competitive with the lease on price alone.
The residual question moves from the lender to the lessee, the venture, and the offtakerAccelerating
Speed: Visible now; decisive by 2028
What the sources say. The master finance lease amortises at one-twentieth of financed cost per quarter, repaying the lessor in five years. The largest capacity buyer's maximum residual value guarantees in unconsolidated data center ventures rose from $28,091 million at 31 December 2025 to $41,493 million at 30 September 2026, with 20% membership interests, four-year initial lease terms renewable to twenty, and no liability recorded because payments are not probable. One neocloud depreciates technology equipment over six years.
Evidence grade: A on all three — three registrants' own filings.
What we read into it. Nobody in this market is being paid to take residual risk on a GPU. The lessor refuses it by amortising to zero. The accounting assumes six years but books no second-hand value. The capacity buyer writes a residual value guarantee and records no liability for it. The risk has not been priced; it has been pushed to whoever is contractually last in line, and in each of these three structures that is the operator or the guarantor rather than the financier. The first public number that tests any of it will move every advance rate in the market.
Confidence: high. Three independent filings, each disclosing its own treatment, and none of them disclosing a residual value anyone has realised.
What would reverse it. A published, arm's-length sale of a large used GPU fleet at a price above its depreciated book, which would create the residual evidence the market currently lacks and let lenders lend against it.
Public price discovery arrives for rental before it arrives for the assetRunning
Speed: Already here for rental; 2027–2028 for the asset
What the sources say. One provider publishes live spot prices per GPU platform and region through REST and gRPC APIs its documentation states require no authentication, with up to 30 days of history. Read on 5 November 2026, preemptible H100 cleared at $1.00 a GPU-hour in one region, H200 between $0.90 and $1.85 across three, B200 between $3.50 and $5.60, and B300 between $8.40 and $10.70. The same provider's published on-demand list prices are $4.50 for H100, $5.40 for H200, $8.50 for B200 and $9.50 for B300.
Evidence grade: B · party — one provider's own feed and own list. It is a price anybody can read, but it is one venue, not an index.
What we read into it. An asset that rents at a public price can be underwritten against that price. FD estimate: $1.00 ÷ $4.50 = 22% of list for H100, and $10.70 ÷ $9.50 = 113% of list for B300 at the top of that read — the newest chip clears at or above its own list price and the two-generations-old chip clears at a fifth of it. The H100 ratio is the stable half of that comparison; the B300 range is a live feed that moves within the day, so the 113% is a maximum on one read rather than a standing level. That is the depreciation curve, published, dated and free to read. It is also why the five-year fully-amortising lease exists: a lender reading that curve will not underwrite a residual at year five, and a structure that does not need one wins.
Confidence: moderate. The readings are precise and reproducible, but spot capacity at one venue is a thin market and preemptible pricing is not the same as the price of owning the chip. We carry it as the best public proxy and say so.
What would reverse it. A second and third venue publishing comparable feeds that disagree sharply with this one, which would show the readings are a venue's inventory position rather than the market's view of the chip.
Power becomes a credit term, not an operating expenseRunning
Speed: 12 to 24 months to become standard
What the sources say. One neocloud's March 2026 facility requires interest rate hedges covering at least 95% of reasonably anticipated outstanding floating-rate borrowings and includes power cost hedging requirements. The market operator's own report on the 2028/2029 Base Residual Auction shows a clearing price of $325.00 per MW-day at the cap, against $28.92 for delivery year 2024/25, a cleared reserve margin of 14.4% against a 20% target, and total new generation and uprates of 524.7 MW of unforced capacity.
Evidence grade: A on both — a registrant's own filing and the market operator's own results report.
What we read into it. A lender that makes you hedge power has concluded that power is a price risk large enough to break the loan. The capacity auction says why: the cost of procuring capacity across the footprint went from $2.2 billion for delivery year 2023/24 to $16.4 billion for 2028/29, while the market cleared below its own reliability requirement for the second year running. For a GPU financing, the practical effect is that the power contract is now diligenced before the offtake contract, because a cluster with chips and no firm power is a stranded asset and a cluster with power and no chips is a queue position worth money.
Confidence: high. The covenant and the auction result are both quoted from the parties' own publications, and the direction is corroborated by the trade body note describing power availability as shaping both where data centers are built and how they are financed.
What would reverse it. A capacity auction clearing well below its cap with a reserve margin at or above target, which would take the urgency out of the power covenant and move it back to a representation.
The vendor and the offtaker become financiers of last resortRunning
Speed: Decisive in a downturn, invisible otherwise
What the sources say. The chipmaker's supply and capacity commitments reached $279 billion at 26 July 2026 from $119 billion the previous quarter, with $92 billion due in the remainder of fiscal 2027, and its own equity investment commitments total $25 billion with $18 billion in the remainder of fiscal 2027. Data center revenue was $89,023 million in the quarter. A capacity buyer committed to fund its pro rata share of about $27 billion and about $14 billion of estimated development costs in two ventures in which it holds 20%.
Evidence grade: A on both — two registrants' own quarterly filings.
What we read into it. When the supply chain commits further forward than the demand chain contracts, something has to bridge the gap, and the parties with the most to lose bridge it. A chipmaker with $25 billion of equity commitments is funding its own customers. A capacity buyer holding 20% of a venture and guaranteeing the residual is funding its own landlord. Both are rational and both concentrate risk in the same place: if demand disappoints, the vendor's receivable, the vendor's equity stake and the vendor's revenue all fail together.
Confidence: moderate. The commitments are quoted, but reading them as financing rather than ordinary supply-chain and joint-venture activity is an inference, and neither filing frames them that way.
What would reverse it. Supply commitments flattening or falling for two consecutive quarters while data center revenue keeps rising, which would mean the chain is being funded by cash from customers rather than by the vendor.
The securitisation exit opens for buildings and has not yet opened for equipmentPending
Speed: 2027 for the first test; 2028 to be settled
What the sources say. A trade body note of 23 July 2026 reports Barclays research putting outstanding data-center securitisation at $61 billion year-to-date 2026 from $4 billion in 2020, about 12% of the esoteric ABS market against 3% in 2020, with a conditional in which reaching $180 billion by year-end 2028 would be roughly 20% of the esoteric ABS market and 15% of single-asset/single-borrower CMBS. Average issuance is given as about $600 million for ABS and $1.2 billion for CMBS. The same note says GPUs, servers and other computing equipment represent a significant share of project costs, are increasingly financed through specialised lending facilities, and that equipment-backed financings may emerge as a complementary ABS sector.
Evidence grade: C · crowd — a trade body summarising two banks' research, carried with its attribution rather than as a count we can reproduce.
What we read into it. Securitisation is where a financing market finds out what its assets are worth to someone who did not originate them. Buildings have reached it. Equipment has not, and the reason is the residual: a master trust can add stabilised data centers because their leases run fifteen years, and cannot easily add GPU pods whose rental price halves between generations. The first equipment-backed deal to price will set the advance rate for every GPU lease written afterwards, which is why we treat it as the single most informative event available between now and 2028.
Confidence: moderate. The issuance figures are second-hand. The structural reason equipment lags is our reading, and it rests on the amortisation profile of the one large lease on the public record.
What would reverse it. An equipment-backed deal pricing wide and trading badly, or a large data-center ABS deal missing an anticipated repayment date, either of which would close the exit rather than open it.
The causal model
The loop that produced everything above, in six steps, with the figure that shows each one is running. Three arms reinforce it and three brake it; the scenarios are what happens when one arm or one brake wins.
- 1
Offtake
A creditworthy buyer signs for capacity on a long term
210 MW, ~$5.2B, ~15 years - 2
Underwrite
A lender sizes against the equipment's depreciable cost and the contract
SOFR + 2.25% - 3
Build
Chips, power and shell are procured against the committed facility
$279B of supply commitments - 4
Operate
Capacity sells, at a published rate or a contracted one
H100 spot $1.00/GPU-hr - 5
Refinance
Stabilised assets move to bonds or securitisation
$61B outstanding - 6
Recycle
Freed capital funds the next tranche, at a tighter spread
15% → 9%, then T+200 non-recourse
↻ Each tighter spread makes a larger build affordable, which makes a longer offtake necessary, which makes the next financing larger still. The loop has run four times in three years at one borrower, and the lease is what it produced on the fifth.
The three arms, and the three brakes
Read the loop and the scenarios stop being moods. A lease-led market is what happens when R1 keeps running and nothing brakes it. Repricing is B1 arriving before the securitisation exit opens. A credit event is B3 resolving in public. The odds below are our judgment about which of those gets there first.
How it evolves, 2027 and 2028
Five phases to the end of 2028, each named for what changes in it, with the Frontier Desk judgments that attach to that window. Every figure here is one of the calls in band 13, carried unchanged.
The largest lease closes, or it does not
The master finance lease committed to on 24 October 2026 is expected to close before year end, subject to specified closing conditions, with the fixed rate determined at pricing. This window is where the market learns three things it cannot learn any other way: whether a $40 billion equipment lease closes on its original terms, what rate a lessor charges for five-year fully-amortising GPU paper, and whether the rate is disclosed at all. Expect the disclosure to be thin and expect the number to leak through a subsequent filing rather than a press release.
The lease closes by 31 Mar 2027: 82% · its priced fixed rate becomes public by 30 Jun 2027: 44%
Copying begins, and the first lease schedules commence
The lease schedules under the largest transaction are expected to commence in the first half of 2027 on delivery of the GPUs. In the same window, expect the second and third large operators with lien-limited unsecured debt to announce equipment lease facilities rather than secured loans, and expect at least one arranger to market the structure as a product rather than as a one-off. The tell is language: a commitment letter that says the lessor will acquire and lease rather than that the lender will advance against.
A second master equipment lease of $5B or more is disclosed: 61% · a chipmaker or hyperscaler commits $5B or more to an operator: 56%
The first residual test, and the first renewals
The earliest large GPU facilities written in 2024 reach refinancing in this window, and the chips behind them reach year three of a six-year book life. This is where the market finds out whether a used H100 fleet has a bid, and the public spot feed is the first place the answer will show. Expect the gap between accounting life and financing life to be argued in public for the first time, and expect at least one operator to extend its depreciable life rather than take the charge.
An operator extends its GPU depreciable life: 47% · a GPU- or AI-equipment ABS of $500M or more prices: 34% by end-2028
Power, not capital, becomes the binding constraint
The 2028/2029 delivery year begins on 1 June 2028 at a capacity price that cleared at its cap with the market short of its reliability requirement. The campuses contracted for delivery in the first half of calendar 2028 arrive into that. Expect financing to stop being the scarce input and interconnection to start being it, and expect lenders to price a queue position the way they price a power purchase agreement. Our companion forecast on power and interconnection covers what the queues themselves say.
A GPU financing is publicly delayed or resized for power reasons: 52% · power hedging appears in a second named facility: 58%
Either the equipment securitises, or the lease market stops where it is
On the central path the equipment-backed securitisation opens, advance rates get set by a market rather than by a bilateral negotiation, and the lease becomes the normal way a large operator funds chips. On the alternative path the residual evidence arrives badly, advance rates fall, and the market reverts to secured loans at wider spreads with more equity underneath. Both paths run through the same event, which is why the first large equipment-backed deal is the most informative thing available between now and then.
Lease-led market: 38% · repricing: 20% · finance lease the dominant structure for a new large purchase: 36%
Scenarios
Five ways compute financing can go. Each carries Frontier Desk odds at 31 December 2027 and 31 December 2028, the trigger that puts you on it, the early markers that tell you first, and what it means on each side of a financing. Each column sums to 100, and every signpost in band 12 names the scenario it moves.
Secured-loan market
Today's shape continues. Delayed-draw term loans sized off the depreciable cost of computing equipment stay the normal way a large GPU purchase is funded, spreads keep grinding tighter, and the lease stays a special case for borrowers who have run out of lien capacity. The reason this falls from 40% to 17% is not that it stops working. It is that every operator that issues unsecured debt acquires the covenant problem the lease was built to solve, and the number of such operators only goes up.
Trigger: None. This is the no-trigger path: nothing resolves, the covenant pressure stays confined to the largest borrowers, and bank facilities keep clearing.
Early markers: New facilities announced as term loans rather than leases; advance rates holding against the depreciable cost of equipment; no second large master equipment lease disclosed.
For an operator: The cheapest capital is still a secured facility if you have the lien capacity. Check the capacity before you negotiate the pricing, because the covenant decides the structure and the structure decides the price.
For lenders: Banks and private credit keep the origination. The competitive question is advance rate against a depreciating asset, and the answer moves with the public spot feed whether or not anyone says so.
Lease-led market
The modal destination by the end of 2028. A third-party lessor acquires the chips and leases them on a fully amortising schedule, the operator never owns them and grants no lien, and the rent sits outside most leverage tests. Pricing converges on a Treasury spread rather than a credit spread, because the lessor is underwriting a rent stream and a five-year amortisation rather than a balance sheet. The operator keeps the power risk, the offtake risk and the technology risk, and sheds the funding and residual risk — which is a good trade only if it has the first three under control.
Trigger: A second and third master equipment lease of comparable structure, and the first arranger marketing it as a repeatable product rather than a bespoke commitment.
Early markers: Commitment letters phrased as the lessor will acquire and lease; lease schedules commencing on delivery; equipment lease obligations disclosed separately from funded debt.
For an operator: The best structure available to a lien-constrained operator, and the one that needs the most preparation: a clean offtake, firm power, and a rate-risk plan for the gap between commitment and pricing.
For lenders: Lessors and the private credit funds behind them take the origination. Banks keep the construction and the revolver, and lose the equipment.
Vendor-and-offtaker funded
The chipmaker and the capacity buyer carry more of the financing between them: equity commitments, prepayments, guarantees and residual value guarantees do what a lender would otherwise do. Note the shape of the odds: this is a waypoint rather than a destination. It is where the market goes when third-party capital hesitates, and on most paths it hands back to the lease or the loan once the hesitation passes, which is why its share flattens rather than compounding.
Trigger: A sharp widening in third-party spreads, or one large operator failing to place a facility and being bridged by its vendor or its offtaker instead.
Early markers: A chipmaker's equity investment commitments rising faster than its revenue; residual value guarantees growing faster than lease commitments; prepayments appearing in operators' deferred revenue.
For an operator: Vendor money is the fastest and the most expensive in rights rather than in rate. Read what it takes in exclusivity, in board consent and in change-of-control before you read the coupon.
For lenders: Third-party lenders get crowded out of the best assets and left with the rest. The adverse selection is the whole risk in this scenario.
Repricing
Residual values disappoint in public, advance rates fall, and the cost of capital on chips widens again. Spot rental prices for the two newest generations converge downward, a used-fleet sale prices below depreciated book, or an operator takes an impairment large enough to be disclosed. Lease amortisation shortens from five years to four and advance rates fall with it. This is the second-likeliest 2028 outcome and the one least discussed by the people arranging the paper.
Trigger: Either a disclosed impairment or used-fleet sale below book, or spot rental prices for the newest generation falling to where the oldest generation trades today.
Early markers: The spread between newest and oldest spot prices narrowing from below; an operator shortening rather than extending its depreciable life; an equipment ABS pricing wide or being pulled.
For an operator: Time is the risk, not price. An operator waiting for a better financing market in a repricing is waiting for the opposite of what arrives, which argues for committing capacity now rather than later.
For lenders: Whoever underwrote a residual takes the loss first, and on the public record that is the guarantor rather than the lessor. Expect advance rates to fall before spreads widen.
Credit event
A neocloud or an offtaker fails, or restructures, at a size that shuts the market to everyone below investment grade. The collateral turns out to be worth what the spot feed says rather than what the schedule says, the concentration in a handful of offtakers becomes a headline rather than a footnote, and new issuance stops for two quarters. The odds do not rise with the horizon because the same build-out that creates the exposure also creates the cash flow that services it; this is a tail, not a trend.
Trigger: A missed payment, a covenant breach made public, or an offtaker terminating a large capacity contract.
Early markers: An operator drawing a revolver to pay interest; a facility amended to defer amortisation; an offtake contract disclosed as terminated or renegotiated.
For an operator: Everything depends on work done before the event: committed power, a contract that survives the counterparty, and enough liquidity to not be a forced seller into a market with no bid.
For lenders: The secured lender with a perfected first-priority interest in equipment discovers what that equipment is worth. The lessor, having amortised to zero, discovers less.
Winners and losers
Who gains and who loses on the central path, and why. This band is analysis throughout: direction is stated against today, not against each other, and nothing here is a reading.
Operators and owners of compute
| Segment | Direction | Why |
|---|---|---|
| Large operators with investment-grade-style unsecured debt and a lien limit | Clear winner | The lease exists for exactly this balance sheet. A lien basket of 7.5% of total assets stops a secured financing long before it stops a lease, so the structure that avoids the basket is the structure that scales. These operators get the cheapest chips in the market and keep their bond covenants intact. |
| Mid-size operators with a signed, investment-grade offtake and firm power | Winner, selectively | The contract is the collateral. An operator with a long contract to a high-rated buyer and committed interconnection can borrow at a spread that reflects the buyer rather than itself, which is the single largest lever available below the top tier. |
| Operators selling on-demand or marketplace capacity with no contracted offtake | Clear loser | A public spot feed showing the oldest chips at a fifth of list removes the argument that uncontracted capacity has a floor. Without an offtake, the lender is underwriting a merchant price it can now look up, and it will size accordingly. |
| Owners of a queue position and a site, with no chips yet | Winner, quietly | Capacity that cleared at its cap in a market short of its reliability requirement makes an interconnection position a priced asset rather than a permission. The scarce input is moving from capital to power, and this is the side of that trade that has the power. |
| Operators whose entire revenue depends on one or two offtakers | Loser on the central path | Concentration is the one thing the securitisation market will not buy, so these operators stay in bilateral financings at bilateral pricing. The financing discount for concentration widens as the rest of the market gets a public bid. |
Lenders, lessors and the layers between
| Lender or layer | Direction | Why |
|---|---|---|
| Lessors backed by permanent or long-duration private capital | Clear winner | They own the structure the market is moving to, they can hold a five-year fully amortising asset to maturity, and they are the natural originator for the equipment securitisation that has not yet opened. The $40 billion commitment on the public record is this business in one transaction. |
| Banks arranging delayed-draw term loans against equipment | Loser at the top, winner in the middle | They lose the largest borrowers to the lease and keep the mid-market, where lien capacity is not binding and a relationship still prices the deal. Construction lending and revolvers stay with them in every scenario. |
| Private credit funds buying equipment paper without a residual view | Loser | Buying a loan whose advance rate was set before the spot feed existed is buying someone else's residual assumption. The repricing scenario hits this holder first and hardest, because the loan is marked and the lease is not. |
| Chipmakers extending equity and supply commitments to their own customers | Winner until it is not | Vendor financing accelerates revenue and holds price, and it concentrates receivable, equity and revenue risk in the same counterparties. It is the best trade in four of five scenarios and the worst in the fifth. |
| Ratings agencies and the securitisation arrangers behind them | Clear winner | An asset class moving from $4 billion to a conditional $180 billion in eight years needs somebody to say what the tranches are worth. Whoever writes the first equipment-backed criteria sets the advance rate for everyone. |
Frontier Desk is the advisor to the company raising the money, not a lender and never a broker. The first conversation covers how a lender or lessor will read your project, which structures fit it, and what each will ask for next.
Commoditised or scarce
What becomes a commodity in this market and what keeps its scarcity value, with the date we think each crosses. This band is a forecast, and the band after it argues against itself.
Commoditising, and by when
| What | By when | Why |
|---|---|---|
| The chips themselves, as an asset to finance | 2027 | A published spot price per platform per region, free to read, turns a bespoke underwriting question into a lookup. Once the price is public the advance rate stops being a negotiation. |
| Senior secured lending against equipment | 2027 | Recourse lending at one borrower moved from 15% to 11% to 9% in three years, and the same borrower’s non-recourse facility prices at Treasury + 200. A product whose effective rate falls by two-fifths on identical credit terms in three years is a product with no differentiation left in it. |
| The lease structure itself | 2028 | A master finance lease with tranche schedules and one-twentieth quarterly amortisation is legal engineering that copies perfectly. The second one is cheaper to write than the first, and the fifth is a form. |
| On-demand GPU capacity without a contract | Already | Preemptible H100 at $1.00 against a $4.50 list is what a commodity looks like. The newest generation is not there yet, and will be on the same schedule. |
| Construction lending on data center shells | 2028 | A 39-year building with a 15-year lease to a high investment-grade tenant is a real estate credit the market has priced for decades. The AI label does not change the underwriting. |
Keeping its scarcity value, and why
| What | How long | Why |
|---|---|---|
| Firm, dated interconnection and power | Through 2028 and beyond | A capacity market clearing at its cap, below its own reliability requirement, with 524.7 MW of new generation and uprates in a single auction, is not a constraint capital can relieve. This is the one input money cannot manufacture. |
| A long contract with a high-rated offtaker | Through 2028 | The contract is what converts a depreciating asset into a financeable cash flow. $5.2 billion over roughly fifteen years from one investment-grade tenant is worth more than the campus it sits in. |
| Willingness to underwrite a GPU residual | Until somebody publishes one | Every structure on the public record refuses this risk or allocates it without pricing it. The first party paid properly to take it will have no competition and a very large market. |
| Balance sheet capacity that is not covenant-constrained | Through 2027 | Lien baskets are finite and bond indentures are hard to amend. Unencumbered capacity is the scarce input that makes the cheapest structure available at all. |
| Operating skill at utilisation | Through 2028 | Financing cost is converging across operators. What does not converge is how much of the installed fleet is actually earning, and that gap is now the main driver of returns rather than the cost of money. |
The case against this table.
The strongest argument against this table is that we have read one spot feed as if it were a market. A single provider's preemptible pricing reflects that provider's inventory, its region mix and its commercial strategy, and preemptible capacity is the most volatile corner of any cloud's price list. If the 22%-of-list reading for H100 is a venue artefact rather than a depreciation curve, then the residual pessimism running through this whole report is wrong, the five-year fully amortising lease is over-engineered, and the secured-loan market does not shrink the way we forecast.
The second argument is that we have treated one transaction as a structural shift. There is exactly one master finance lease of this size on the public record. A sample of one supports a mechanism and does not establish a trend, and the covenant explanation — which is ours, not the filer's — would look very different if the second large operator funds its next tranche with a secured facility and the lien capacity to spare. We hold the lease-led scenario below two in five at 2028 for this reason, and the signposts in band 12 are mostly designed to catch us being wrong about it early.
What to do about it
What the forecast implies for the two readers this report is written for: the owner financing compute, and the adviser or investor reading the paper. Frontier Desk is an advisor, not a lender and not a broker, and nothing here is a recommendation to buy or sell a security.
If you are financing compute
| Decision | What the forecast says |
|---|---|
| Loan or lease? | Read your lien basket before you read either term sheet. If a secured financing of the size you need would exceed the basket in your existing notes, the structure is already decided and the only live question is the rate. If it would not, the loan is usually cheaper today and the lease is usually cheaper by 2028. |
| What should the amortisation be? | The one large lease on the public record repays one-twentieth of financed cost a quarter, which is five years to zero. Treat anything longer as the counterparty taking a residual view, and ask what it is based on, because no public source states a realised GPU residual. |
| When is the rate fixed? | A fixed rate determined at pricing leaves you carrying rate risk between commitment and close. The five-year Treasury moved 158 basis points between its low and its high this year. Ask for the pricing mechanic in writing and price the gap. |
| What does the lender want on power? | A facility that requires power cost hedging has concluded electricity can break the loan. Have the power contract and the interconnection position documented before the financing diligence starts, because on the central path power becomes the binding constraint in the first half of 2028. |
| How much does the offtake matter? | More than anything else on your balance sheet. A long contract with a high-rated buyer lets you borrow at something closer to the buyer's spread than your own, and it is the only thing that makes uncontracted capacity financeable in a repricing. |
| What if you own the site but not the chips? | You are on the scarce side of the trade. An interconnection position in a market that cleared at its capacity cap is an asset with a price, and it is worth establishing that price before you commit it to one counterparty's build. |
If you hold or underwrite the paper
Mark the residual, not the coupon. Every structure on the public record either amortises to zero or guarantees a residual without recording a liability. If you hold paper whose advance rate assumed a value at year five, you hold a view nobody has published evidence for, and the public spot feed is the cheapest available test of it.
Read the concentration twice. The collateral, the revenue and in several cases the equity backstop in this market point at the same small set of names. A portfolio of five operators with one offtaker between them is one credit, and the securitisation market will price it as one when it gets the chance.
Watch the structure of new commitments rather than their size. A commitment letter saying the lessor will acquire and lease is a different market from one saying the lender will advance against. The first is the forecast in this report arriving; the second is it not arriving.
Frontier Desk is the advisor to the company raising the money, never a lender and never a broker. The first conversation covers how a lender or a lessor will read the project, which structures fit it, and what each will ask for next: Get your company ready.
See which lenders and lessors fund projects like yours.
The structure decides the price, and the covenant decides the structure. A short conversation covers which of the two your project is actually constrained by, and who finances projects on the other side of that constraint.
The evidence base
The present-state record the forecast rests on, as a reference band. Everything here is a reading from a named public source with its date; the inferences are in the bands above.
What the public filings show about the cost of capital
Publicly observable| Instrument | Who discloses it | Terms, as disclosed | What it tells the forecast | Source |
|---|---|---|---|---|
| Master finance lease for GPU pods | Space Exploration Technologies Corp. | ~$40bn of aggregate principal payments; tranches on ~5 to 5.25-year schedules; 1/20 of financed cost per quarter plus a fixed rate on the outstanding balance; rate set at pricing | The largest equipment financing on the public record refuses residual risk entirely and grants no lien | Form 8-K, Item 8.01, 26 Oct 2026 |
| Senior unsecured notes | Space Exploration Technologies Corp. | $25,000m in five tranches; 11.7-year weighted-average maturity; coupons 5.350%–6.650%, 5.855% weighted average; effective rate 6.072%; liens limited to 7.5% of consolidated total assets | The covenant that makes the lease necessary, and the benchmark the lease has to beat | Form 10-Q, quarter ended 30 Sep 2026 |
| DDTL 1.0 facility | CoreWeave, Inc. | Entered Jul 2023; matures Mar 2028; effective interest rate 15%; recourse; $1,300m outstanding at 30 Jun 2026 | What GPU-backed credit cost before the asset was bankable | Form 10-Q, quarter ended 30 Jun 2026 |
| DDTL 4.0 facility | CoreWeave, Inc. | $8,500m available through 30 Jun 2027; daily compounded SOFR + 2.25% or Treasury + 2.00%; 0.50% undrawn fee; matures Mar 2032; sized off the depreciable cost of computing equipment; non-recourse, at an indirect subsidiary, with $2,837m outstanding at 30 Jun 2026 | What a structured facility cost three years later, and what the lender sizes against | Form 10-Q, quarter ended 30 Jun 2026 |
| DDTL 5.0 facility | CoreWeave, Inc. | Entered 15 May 2026; effective interest rate 9%; recourse; constrained by the depreciable cost of graphics processing unit servers; secured by first-priority pledges over the borrowing subsidiaries | What recourse paper still costs at the same borrower, on the newest facility in the table | Form 10-Q, quarter ended 30 Jun 2026 |
| Total debt obligations | CoreWeave, Inc. | $35,551m of future principal payments; $4,413m in the rest of 2026, $6,184m in 2027, $4,416m in 2028, $14,896m thereafter; Q2 2026 contractual interest expense $592m against $250m a year earlier | The scale and the near-term maturity wall one neocloud is carrying | Form 10-Q, quarter ended 30 Jun 2026 |
| 0.00% convertible senior notes due 2032 | Cipher Digital Inc. | $850.0m priced 14 Oct 2026 to qualified institutional buyers under Rule 144A, with capped call transactions | Equity-linked capital at a zero coupon, for a compute operator, in the same month as the largest lease | Form 8-K, Item 8.01, 14 Oct 2026 |
| 15-year lease to a high investment-grade hyperscaler | Applied Digital Corporation | Signed 5 Jun 2026; 210 MW of critical IT load; three five-year renewal options; ~$5.2bn of contracted revenue over the base term; delivery expected H1 CY2028; aggregate contracted critical IT load 1,410 MW | What a creditworthy offtake looks like, and the clock the building is financed on | Form 10-Q, quarter ended 31 Aug 2026 |
| Unconsolidated data center ventures | Meta Platforms, Inc. | Maximum exposure to loss $68,404m, from $45,952m at 31 Dec 2025; maximum residual value guarantees $41,493m, from $28,091m; 20% interests; ~$27bn and ~$14bn of committed development costs; four-year initial lease terms renewable to up to twenty; no liability recorded | Where the residual risk went, and the fact that nobody has priced it | Form 10-Q, quarter ended 30 Sep 2026 |
| Supply, capacity and equity commitments | NVIDIA Corporation | $279bn of supply and capacity commitments, from $119bn the previous quarter; $366bn of total future commitments; $25bn of equity investment commitments; $0.4bn charge for excess H200 inventory and purchase obligations | The supply chain is committed further forward than the demand chain is contracted | Form 10-Q, quarter ended 26 Jul 2026 |
Every row is quoted from the registrant's own filing with the Securities and Exchange Commission on the date shown, and read in full on 5 November 2026. Where a figure is a total we computed from a table in the filing, the band says so. Frontier Desk holds no position in any security named here, acts for none of these companies, and names them because their filings are the public record this market has.
The cost of capital on chips, as disclosed
Publicly observableEffective or stated annual rates on named facilities, from each registrant's own filing, with recourse to the parent marked on each. The Treasury row is the base a fixed-rate facility at Treasury plus 200 starts from
Source: Each registrant's own Form 10-Q or 10-K, and the Treasury's daily yield curve, read 5 Nov 2026
What the public price feed shows, by chip generation
Publicly observablePreemptible spot against published on-demand list, per GPU-hour
Publicly observableOne provider, all regions its public API returned, read 5 November 2026. Log scale. The gap widens as the chip gets older
Source: One provider's unauthenticated spot-price API and its published price list, read 5 Nov 2026
Spot values read from the provider's unauthenticated spot-price API on 5 November 2026, at effective times between 16:51 and 18:01 UTC; list prices from the same provider's published pricing page the same day. Preemptible capacity is the most volatile part of any cloud's price list, and one venue is not an index — this is the best public proxy available, carried as a proxy. The feed and its documentation are both public.
Two clocks, as the filings state them
Publicly observable| Asset | Who discloses it | Life, as disclosed | What it implies |
|---|---|---|---|
| Technology equipment | CoreWeave, Inc. | 6 years, moved from 5 effective 1 Jan 2023 | The book life of a GPU, set by the operator, with no residual value stated |
| GPU pods under the master finance lease | Space Exploration Technologies Corp. | Schedules of approximately 5 to 5.25 years, amortising 1/20 of financed cost a quarter | The financing life of the same asset, set by the lessor, one year shorter |
| Networking and electrical equipment | Applied Digital Corporation | 3 to 5 years | The shortest clock in the building |
| Electrical infrastructure | Applied Digital Corporation | 15 years | Three GPU generations, financed once |
| Electric generation and transformers | Applied Digital Corporation | 15 to 30 years | The part that outlives every tenant |
| Building | Applied Digital Corporation | 39 years | The 39-year asset wrapped around the 5-year one |
The two clocks, drawn
Publicly observableDepreciable and financing lives in years, as three registrants state them. The building is the top bar; the chips inside it are the bottom three
Source: Three registrants' own stated depreciable and financing lives, read 5 Nov 2026
Two gaps in this evidence base are worth stating plainly, because the forecast above is weaker where they are.
First, no public source states a price at which a large used GPU fleet has actually changed hands at arm's length. The residual argument running through this report is built from a lease that refuses residual risk, an accounting life with no residual value attached, a guarantee with no liability recorded, and a rental price feed. Each is real and none of them is a sale. Second, the securitisation figures are a trade body's summary of two banks' research rather than a dataset we can recompute, and we carry them with that attribution rather than as a count.
Key assumptions
The assumptions the odds depend on, and the evidence that would break each one. Each is an analytical commitment, not a reading.
| Assumption | Breaks if |
|---|---|
| Lien covenants, rather than price, are what pushes large operators into lease structures. | A large operator with ample lien capacity choosing a master equipment lease anyway, which would mean the driver is pricing or accounting treatment and the covenant story is a coincidence. |
| Published spot rental prices are a usable proxy for the value of owning the chip. | A second and third venue publishing feeds that disagree sharply with the first, or a used-fleet sale clearing far above what the spot discount implies. |
| Nobody will be paid properly to take GPU residual risk before 2028. | A lessor writing a lease with a stated residual value, or a rating agency publishing criteria that assign one, either of which would reopen tenor past five years. |
| Demand for compute holds well enough that offtake contracts signed now are performed. | An offtaker terminating or renegotiating a large capacity contract in public, which moves the credit event scenario from a tail to a trend. |
| Power, not capital, becomes the binding constraint by 2028. | Capacity prices clearing well below cap with reserve margins at target, and interconnection queues clearing materially faster than they are now. |
| The securitisation market extends from buildings to equipment rather than stopping at buildings. | A first equipment-backed deal pricing wide or being pulled, or a rating agency declining to rate equipment collateral without a stated residual. |
| Vendor and offtaker financing stays a supplement rather than becoming the main channel. | A chipmaker's equity and guarantee commitments growing faster than third-party facilities for two consecutive quarters. |
| Rent under a master finance lease stays outside most leverage covenant definitions. | A rating agency or a widely used covenant package recharacterising it as funded debt, which removes the structural advantage and leaves price alone. |
| The concentration of offtakers does not resolve into a single credit event before 2028. | Any disclosed payment default, covenant waiver or contract termination at a top-five offtaker or a top-five operator. |
| Accounting lives and financing lives converge downward rather than upward. | Two or more operators extending GPU depreciable lives beyond six years without an impairment, which would mean the market believes the residual is better than the financing assumes. |
Signposts
A signpost is a reading that moves the odds if it crosses the threshold shown. Each one names the scenario or scenarios it moves, so no reading on this page floats free of the forecast it is supposed to update.
| Indicator | Now | Threshold | Scenario it moves | What it would mean | Checked |
|---|---|---|---|---|---|
| Master equipment leases of $5bn or more disclosed | 1, committed 24 Oct 2026 | A second one | Lease-led market | The structure is a product rather than a one-off, and the lease-led odds roughly double | SEC EDGAR, 8-K and 10-Q |
| Spread on new GPU-backed facilities | Treasury + 2.00% on one Mar 2026 facility | Inside Treasury + 1.50% | Secured-loan market · Lease-led market | Bank credit is still competing, and the loan keeps the largest borrowers longer than we forecast | SEC EDGAR |
| Preemptible H100 spot as a share of on-demand list | 22% (FD estimate: $1.00 ÷ $4.50) | Below 15%, or above 40% | Repricing · Lease-led market | Below 15% reads as repricing arriving early; above 40% reads as our residual pessimism being a venue artefact | The provider's public API |
| Spread between newest and oldest generation spot prices | B300 $8.40–$10.70 against H100 $1.00 | Newest falling below 4x the oldest | Repricing | The depreciation curve is steepening at the front, which is the single clearest early warning of a repricing | The provider's public API |
| GPU or AI-equipment-backed securitisation | None public | Any deal of $500m or more pricing | Lease-led market · Repricing | The exit the lease market is building toward has opened, and advance rates get set by a market rather than bilaterally | Rating agency presales and named press |
| Operators' stated GPU depreciable life | 6 years at one operator, from 5 | Any extension past 6, or any cut below 5 | Repricing | An extension says the market believes the residual; a cut says it does not, and advance rates follow either way | SEC EDGAR, 10-K and 10-Q |
| Disclosed impairment or sale of a used GPU fleet | None on the public record | Any disclosed figure | Repricing · Lease-led market | The first real residual datapoint this market has ever had; everything in band 03 gets rescored against it | SEC EDGAR and named press |
| Chipmaker supply and capacity commitments | $279bn, from $119bn the previous quarter | Two consecutive quarters flat or falling | Vendor-and-offtaker funded · Repricing | The chain is being funded by customers rather than by the vendor, which takes the vendor scenario off the table | SEC EDGAR, 10-Q |
| Maximum residual value guarantees at the largest capacity buyer | $41,493m, from $28,091m | A recorded liability, or a fall | Repricing | A recorded liability would be the first time anybody prices a data center residual on a balance sheet | SEC EDGAR, 10-Q |
| Power cost hedging in facility covenants | 1 named facility | 3 or more | Lease-led market · Secured-loan market | Power has become a standard credit term, and interconnection diligence moves ahead of offtake diligence | SEC EDGAR |
| PJM capacity clearing price | $325.00 per MW-day, at the cap | Clearing below the cap with reserve margin at target | Repricing | The power brake has released, and capital goes back to being the binding constraint | The market operator's own auction report |
| Five-year US Treasury yield | 4.69% on 4 Nov 2026 | Below 4.00%, or above 5.25% | Secured-loan market · Repricing | Below 4% keeps the loan competitive on price; above 5.25% widens every all-in rate regardless of spread | The Treasury's daily yield curve |
| Large vendor or offtaker commitments to operators | $25bn of equity commitments at one chipmaker | Any new commitment of $5bn or more | Vendor-and-offtaker funded | Third-party capital is hesitating and the chain is being bridged from inside | SEC EDGAR and company announcements |
| Disclosed payment default, waiver or terminated offtake | None at a top-five name | Any one | Credit event | The tail is arriving; new issuance below investment grade stops for two quarters | SEC EDGAR and named press |
Calls
Each call is dated, carries a Frontier Desk judgment as a percentage, and names the public source that settles it. No call on this page is settled by a social post, an unnamed source or a private record. We score each one when it resolves, with the Brier score: the squared gap between our odds and what happened, where 0 is perfect.
| Call | Resolves | Odds | Settled by |
|---|---|---|---|
| The $40bn master finance lease closes on substantially the disclosed terms | 31 Mar 2027 | 82% | SEC EDGAR, the lessee's 8-K and 10-K |
| The fixed rate priced under that lease becomes public in a filing | 30 Jun 2027 | 44% | SEC EDGAR, the lessee's 10-Q or 10-K |
| A second master equipment lease of $5bn or more for compute hardware is disclosed | 31 Dec 2027 | 61% | SEC EDGAR, 8-K or 10-Q |
| A chipmaker or hyperscaler discloses a new equity, guarantee or backstop commitment of $5bn or more to a compute operator | 31 Dec 2027 | 56% | SEC EDGAR and company announcements |
| A new US GPU equipment facility prices inside Treasury plus 150 basis points | 31 Dec 2027 | 29% | SEC EDGAR |
| A listed operator extends its stated GPU depreciable life beyond six years | 31 Dec 2027 | 47% | SEC EDGAR, 10-K or 10-Q |
| Power cost hedging appears as a covenant in a second named credit facility | 31 Dec 2027 | 58% | SEC EDGAR |
| A GPU financing is publicly delayed, resized or relocated for power or interconnection reasons | 30 Jun 2028 | 52% | SEC EDGAR and named press |
| Preemptible spot for the newest GPU generation falls below four times the oldest generation at the same venue | 30 Jun 2028 | 43% | The provider's public spot-price API |
| A disclosed impairment or arm's-length sale of a used GPU fleet puts a number on a residual | 31 Dec 2028 | 49% | SEC EDGAR and named press |
| A GPU- or AI-equipment-backed securitisation of $500m or more prices | 31 Dec 2028 | 34% | Rating agency presale reports and named press |
| Outstanding data-center securitisation reaches $180bn | 31 Dec 2028 | 41% | Published research from the banks the trade body cites |
| A rating agency publishes criteria for rating AI or GPU equipment collateral | 31 Dec 2028 | 46% | The agencies' own published criteria |
| A disclosed payment default, covenant waiver or terminated offtake at a top-five operator or offtaker | 31 Dec 2028 | 31% | SEC EDGAR and named press |
| The five-year US Treasury yield is below 4.00% | 31 Dec 2027 | 33% | The Treasury's published daily yield curve |
| Financing market: secured loan / lease-led / vendor-and-offtaker / repricing / credit event | Jan 2028 / Jan 2029 | 40/24/15/12/9 · 17/38/16/20/9 | Triggers and markers in Scenarios |
| Dominant structure for a new large GPU purchase: secured loan / finance lease / unsecured or converts / vendor-linked | Jan 2028 / Jan 2029 | 44/22/20/14 · 25/36/22/17 | SEC EDGAR, operators' own filings |
Questions to test this
Questions a sharp reader would use to break the argument.
- 01If no public source states what a used GPU fleet has sold for, on what evidence does anyone set an advance rate — and is the answer that nobody does, and the five-year amortisation exists precisely so nobody has to?
- 02The covenant explanation for the lease is ours, not the filer's. What would a transaction look like that settled it either way, and would any party involved have a reason to make the motive public?
- 03A 39-year building wraps a 5-year chip. Who should own which, and is the current answer — one owner for the shell, a second for the chips, a third guaranteeing the residual — a sign of a maturing market or of a risk nobody wants?
- 04Preemptible H100 at 22% of list could be obsolescence or it could be oversupply at one venue. What reading would distinguish the two before the first large refinancing lands in the second half of 2027?
- 05If a chipmaker's equity commitments are financing its own customers, at what ratio of commitments to revenue does that stop being supply-chain management and start being vendor financing — and who would say so first?
- 06Power cost hedging is now a covenant at one named facility. If capacity prices stay at their cap, does the lender end up underwriting the grid, and is that a risk any lender in this market is equipped to hold?
- 07The securitisation market reached buildings in six years and has not reached equipment. Is that a timing lag or a structural refusal, and what would the first deal have to look like for the answer to be the first?
- 08Every scenario on this page except one assumes the offtakers perform. What is the honest probability that a market with this much concentration gets to 2029 without a single public credit event, and is 9% too low?
Sources and method
Where the direction comes from
Direction comes from Frontier Desk's standing coverage of compute and the build-out behind it, which we keep as a live map of who finances capacity, on what disclosed terms, and what each of them changes. The public surface of that work is at /compute and /markets.
The standard every claim had to clear
A figure, a quotation or a market fact is carried only if an independent primary or reputable named source states it — a filing with a securities regulator, a public API or dataset, a government publication, an organisation's own page about itself, or a named trade body — and is dropped otherwise, with no intermediate framing for things we could not confirm. Every source on this page was opened and read on 5 November 2026. No figure on this page comes from a private record, from a company Frontier Desk advises, from any piece of work it is doing, or from an unnamed source.
How a reading was graded
Processing graded every reading A for a record, B for a party's own published statement about itself, and C for a crowd source, and flagged conflicts rather than resolving them silently. Two examples: the securitisation volumes are a trade body quoting two banks, so they are graded C and attributed in the sentence that carries them rather than stated flat; and the spot prices are one provider's own feed, so they are graded B and described as a proxy wherever they are used.
How the readings became a forecast
Analysis turned the readings into structural shifts, a causal model, phases, scenarios and odds, and every judgment on this page carries its own confidence and the reason for it, kept visibly separate from the grade on the evidence under it. Where a figure is derived rather than quoted it is marked FD estimate and its arithmetic is shown beside it.
What the odds are, and what they are not
The odds on this page are Frontier Desk judgments at two fixed horizons, not the output of a model, and the page says so wherever they appear. They are internally disciplined rather than simulated: each column sums to 100, a call that is a scenario arriving is held at or below that scenario's odds, and a call that is a necessary marker of a scenario is held at or above it. They are intended to be scored and to be wrong in public.
What comes back, and when
Dissemination is this page. Feedback is the signposts, each of which names the scenario it moves, and the Brier-scored calls. The next update is 5 December 2026, and every reading on this page is re-read against its source before that update is published.
What we dropped, and why
Several claims that would have strengthened this report did not clear the standard above, and are not on the page in any form.
Aggregate neocloud borrowing figures
Widely circulated totals for how much the sector has borrowed in 2026 come from research notes we could not open in full, or from aggregations whose method is not published. The debt figures on this page are instead summed from named registrants' own filings, which is narrower and checkable.
Second-hand GPU prices from resale listings
Asking prices on resale venues are not arm's-length clearing prices and the listings are not dated reliably. The residual argument rests on a rental price feed, an accounting life and a lease amortisation schedule instead, and the report says plainly that none of those is a sale.
Rate and spread figures attributed to unnamed arrangers
Pricing talk on facilities that have not been filed is routine in this market and unverifiable by a reader. Every rate on this page is quoted from a filing or a published government series.
A motive for the largest transaction
The covenant explanation in band 01 and shift 01 is Frontier Desk's inference from two filings by the same registrant. Neither filing states a reason for choosing a lease, no party has commented, and the page says so each time the inference is used.
What we still cannot verify
No public source states the fixed rate under the largest lease, because it will be determined at pricing. No public source states a realised sale price for a large used GPU fleet. And the securitisation outstanding and projection figures are second-hand throughout; we could not obtain the underlying bank research.
- 01The master finance lease, in the lessee's own words: a binding commitment letter dated 24 Oct 2026 to enter into a master finance lease with an entity managed by Apollo Global Management, Inc., under which the lessor will acquire and lease “certain graphics processing unit pods and related equipment … for use at data centers selected by the Company”, in tranches on schedules of “approximately five to five and one-quarter years”, with rent “equal to one-twentieth of the financed cost of the applicable tranche” plus a fixed rate on the outstanding balance, and “aggregate principal payments … approximately $40 billion”; filed 26 Oct 2026, read 5 Nov 2026: Form 8-K, Item 8.01
- 02The same registrant's balance sheet and covenant: total assets $254,045m, total debt and finance leases $38,989m, June 2026 senior notes of $25,000m in five tranches with a 11.7-year weighted-average maturity and coupons of 5.350% to 6.650% (5.855% weighted average, 6.072% effective), and “restrictions on additional liens in excess of 7.5% of the Company's consolidated total assets”; quarter ended 30 Sep 2026, read 5 Nov 2026: Form 10-Q
- 03One neocloud's full debt table and facility terms: effective rates of 15% on the DDTL 1.0 facility, 11% on DDTL 2.0 and 9% on three later facilities, all filed as recourse debt, against a separate “Non-recourse debt” heading carrying the $8.5bn DDTL 4.0 facility at “daily compounded SOFR plus 2.25% per annum” or “2.00% per annum plus a blended rate based upon the applicable United States Treasury securities”, a 0.50% undrawn fee, debt sizing “based upon the depreciable cost of computing equipment”, interest-rate hedges on “at least 95%” of anticipated floating-rate borrowings, “certain power cost hedging requirements”, and $35,551m of future principal payments; quarter ended 30 Jun 2026, read 5 Nov 2026: Form 10-Q
- 04The same operator's depreciable lives: “Effective January 1, 2023, the Company changed its estimate of the useful life for its computing equipment utilized in data centers from five to six years”, with technology equipment listed at 6 years; year ended 31 Dec 2025, read 5 Nov 2026: Form 10-K
- 05A zero-coupon convertible for a listed compute operator: “$850.0 million aggregate principal amount of 0.00% convertible senior notes due 2032”, priced 14 Oct 2026 under Rule 144A; read 5 Nov 2026: Form 8-K, Item 8.01
- 06What a creditworthy offtake looks like, and the building's depreciation clock: “an approximately 15-year lease (with three five-year renewal options) with a high investment-grade hyperscaler” for “the full 210 MW of critical IT load, representing approximately $5.2 billion of contracted revenue over the base term”, delivery expected in the first half of calendar 2028, aggregate contracted critical IT load 1,410 MW, total minimum contracted payments $35,731,035 thousand, and depreciable lives of 3–5 years for networking and electrical equipment, 15 for electrical infrastructure and 39 for buildings; quarter ended 31 Aug 2026, read 5 Nov 2026: Form 10-Q
- 07Where the residual risk went: maximum exposure to loss in unconsolidated data center ventures of $68,404m against $45,952m at 31 Dec 2025, of which “Maximum residual value guarantees” $41,493m against $28,091m, 20% membership interests, committed development costs of about $27bn and about $14bn, initial four-year lease terms with renewals “for a total lease period of up to 20 years”, and “RVG payments are not probable, and therefore no liability has been recorded to date”; quarter ended 30 Sep 2026, read 5 Nov 2026: Form 10-Q
- 08The supply chain's forward commitments: “increasing supply commitments from $119 billion last quarter to $279 billion as of July 26, 2026”, $366bn of total future commitments including $25bn of equity investments, Data Center revenue of $89,023m of $96,221m total, and “a $0.4 billion charge associated with H200 for excess inventory and purchase obligations, as the demand for H200 products diminished”; quarter ended 26 Jul 2026, read 5 Nov 2026: Form 10-Q
- 09Live public GPU spot prices, read 5 Nov 2026 at effective times between 16:51 and 18:01 UTC across every region the feed returned, through an API the documentation states “doesn't require authentication”, with up to 30 days of history: preemptible H100 $1.00, H200 $0.90–$1.85, B200 $3.50–$5.60, B300 $8.40–$10.70 per GPU-hour: the spot-price API; its documentation
- 10The same provider's published on-demand list prices, read 5 Nov 2026: H100 $4.50, H200 $5.40, B200 $8.50, B300 $9.50 per GPU-hour: the published price list
- 11The US Treasury's own daily yield curve for 2026, read 5 Nov 2026: five-year 4.69% on 4 Nov 2026 against 3.74% on 2 Jan 2026, with a low of 3.51% on 27 Feb and a high of 5.09% on 30 Sep across 212 published business days: the daily yield curve
- 12How data-center ABS and CMBS fit a broader financing ecosystem, 23 Jul 2026, attributing Barclays Research for outstanding issuance growing “from $4 billion in 2020 to $61 billion year-to-date 2026”, the conditional that “if outstanding data-center securitizations reach $180 billion by year-end 2028, they could comprise” roughly 20% of the esoteric ABS market and 15% of single-asset/single-borrower CMBS, average issuance of “approximately $600 million for data-center ABS and $1.2 billion for CMBS”, and Morgan Stanley Research for “approximately $2.9 trillion in capital expenditures through 2028” with “an estimated $1.5 trillion to be financed externally” — and stating that “equipment-backed financings may emerge as a complementary ABS sector”; read 5 Nov 2026: Structured Finance Association
- 13The capacity market the build-out buys into: the 2028/2029 Base Residual Auction cleared 138,317.8 MW UCAP at the $325.00 price cap, 6,831.3 MW UCAP below the reliability requirement, at a 14.4% installed reserve margin against a 20% target, with total new generation and uprates of 524.7 MW UCAP and cleared supply times clearing price of $16.4bn against $2.2bn for delivery year 2023/24; read 5 Nov 2026: PJM 2028/2029 BRA results report
- 14US data center electricity consumption of 176 TWh in 2023, “about 4.4% of total U.S. electricity”, from 58 TWh in 2014, projected at “325 to 580 TWh” or “approximately 6.7 to 12% of total U.S. electricity” by 2028; read 5 Nov 2026: US Department of Energy
- 15The interconnection queue the next tranche of capacity has to pass through, computed by Frontier Desk from the market operator's own public project feed on 5 Nov 2026: 1,091 active requests totalling 254,591 MW of summer net capacity, against 128 projects totalling 20,285 MW under construction after signing an interconnection agreement: MISO's public queue feed; the interactive queue it serves
- 16Our own public work on how lenders read a compute project, which this forecast is the two-year view behind; each page checked 5 Nov 2026: Finance a GPU cluster; How lenders underwrite a GPU cluster loan; Who lends against GPUs; Buy or rent GPUs
- 17Our companion forecast on the power and interconnection side of the same build-out, published 5 Nov 2026 and read the same day: Data-center power and interconnection to 2028
Find out how a lender will read your project before you ask one.
Frontier Desk is the advisor to the company raising the money. We read the project the way a lender or a lessor will — the contract, the customer, the power, the equity in the deal and the quotes — and tell you which structures fit it and what each will ask for next. Answer what you have and the first conversation covers the rest.
If you own a site or a queue position rather than a cluster, the same review runs the other way: what your interconnection is worth to the people who need it, and what a contract for it should say.
The chips are financed over five years, the building over thirty-nine, and the power over a queue nobody controls. Knowing which of the three binds you is the whole job.
Frontier Desk outlook, 5 November 2026. Frontier Desk acts as the advisor to the company raising capital, through its relationships with lenders, lessors and investors, and a cybernetic system of its own — which is why the forecast stays current. Frontier Desk is not a broker-dealer, lender or investment adviser. Every figure is quoted from a named public source read on the date stated, or marked as a Frontier Desk estimate with its arithmetic shown. The odds are Frontier Desk judgments, not model output, and may be wrong. Nothing here is an offer, a recommendation to buy or sell any security, or advice on any particular financing. Companies are named because their public filings are the record this market has, not because they are connected to Frontier Desk in any way.