The Chips Are Not the Collateral, the Tenant Is: GPU Debt Is Term-Matched, Amortising and Domiciled in Delaware — and the Four-Year Number Is a Meta Lease, Not a Chip Life
The filings invert the premise on two of its three legs — every GPU-secured instrument on the record is deliberately SHORT and amortising against its asset (CoreWeave depreciates technology equipment over 6 years, DDTL 4.0 matures March 2032, DDTL 5.0 runs approximately 5.5 years in quarterly instalments, and only $4,338m of $21,615m of principal falls due after 2030), and not one financing vehicle is offshore ("Cayman" appears zero times in CoreWeave's 10-K, Meta's 10-K and Nebius's 424B5) — leaving the one leg that holds decisively as the story: the collateral does no rating work, because the same sponsor, the same GPU collateral and the same arrangers produce A3/A (low) paper at SOFR+2.25% when the customer is investment grade and Ba2/BB+ paper at SOFR+4.50% when it is not, a 225bp gap of pure counterparty credit, while the only genuinely long bond in the chain — Beignet Investor LLC's $27.3bn 6.581% notes due 2049-05-30, against Meta's "initial four-year lease term" and a residual value guarantee with a threshold of approximately $28bn — is a real-estate deal in which the four years is how long Meta has agreed to rent a building, not how long a GPU lasts.
Executive summary
The hypothesis was compound: that short-lived AI compute assets are being financed with long-dated bonds, through offshore special-purpose vehicles, and rated on the tenant's covenant rather than on the chips. Two of the three legs are inverted by the primary filings. Every GPU-secured instrument on the record is deliberately short and amortising against its asset — CoreWeave depreciates technology equipment over six years and has raised that estimate, not cut it; DDTL 4.0 matures March 2032, DDTL 5.0 runs about 5.5 years, and only $4,338m of $21,615m of CoreWeave's principal falls due after 2030. And not one financing vehicle is offshore: "Cayman" appears zero times in CoreWeave's 10-K, Meta's 10-K and Nebius's 424B5. What remains is the leg that holds decisively, and it is the better story. The same sponsor, the same GPU collateral and the same arranging banks produced A3 / A (low) paper at SOFR + 2.25% when the customer was investment grade, and Ba2 / BB+ paper at SOFR + 4.50% seven weeks later when it was not — 225 basis points of pure counterparty credit. Meanwhile the only genuinely long bond in the chain, Beignet Investor LLC's $27.3bn 6.581% notes due 2049-05-30, stands on Meta's "initial four-year lease term" and a residual value guarantee with a threshold of approximately $28 billion. The four years is real. It is how long Meta has agreed to rent a building.
The Four-Year Number Is a Lease Term, and It Belongs to Meta
Start with the number the whole hypothesis rests on. Four years. It is real, it is precise, and it is printed in a primary filing — but not the one anybody expected. It is in Meta Platforms' FY2025 Form 10-K, in the note describing the Hyperion data-centre venture in Richland Parish, Louisiana: the leases commence in 2029, the aggregate initial lease commitment is approximately $12.31 billion, with "each property having an initial four-year lease term and options to renew for a total lease period of up to 20 years." [4]
That is how long Meta has agreed to rent a building. It is not how long a GPU lasts.
The commissioned hypothesis was a compound one: that short-lived AI compute assets are being financed with long-dated bonds, issued through offshore special-purpose vehicles, and rated on the tenant's covenant rather than on the chips. Three legs. On the evidence in the filings, the first is backwards, the second counts to zero, and only the third stands. And the third, standing alone, turns out to be the better story — sharper, more falsifiable, and considerably more uncomfortable for anyone holding the paper.
Three legs, adjudicated against the filings
| Leg 1 — Short assets, long bonds | Leg 2 — Rated on the tenant's covenant | Leg 3 — Issued through offshore (Cayman) SPVs | |
|---|---|---|---|
| What the hypothesis claims | GPU-secured debt runs long against hardware with a ~4-year life | The chips are collateral in name; the customer's credit is the credit | The vehicles sit offshore, where the covenant stack is invisible |
| What the primary filings show[1][2][3][5][16] | CoreWeave depreciates technology equipment over 6 years (raised from five in 2023); DDTL 4.0 matures March 2032, DDTL 5.0 runs ~5.5 years, and the DDTL facilities amortise in monthly/quarterly instalments | Same sponsor, same GPU collateral, same arrangers: A3 / A (low) at SOFR + 2.25% for an investment-grade customer; Ba2 / BB+ at SOFR + 4.50% for two non-investment-grade customers | CoreWeave Financing DDTL V, LLC and CW Financing DDTL V Holdco, LLC are 'a Delaware limited liability company'; Beignet Investor LLC carries ISIN US076912AA21, invCountry US; Nebius is Dutch |
| The count that settles it[1][2][3][4][13] | Only $4,338m of $21,615m of CoreWeave principal falls due after 2030 | 225 basis points of spread, on hardware that did not change | 'Cayman' appears 0 times in CoreWeave's 10-K, 0 times in Meta's 10-K, 0 times in Nebius's 424B5 |
| Verdict | INVERTED — the paper is shorter than the chips, and it amortises | HOLDS — and it is the whole story | REFUTED to zero — the paper is in Delaware, in plain sight |
Here is the spine of it, up front. The same sponsor, the same class of GPU collateral, the same two arranging banks, the same Delaware special-purpose wrapper produced two loans seven weeks apart. One is rated A3 by Moody's and A (low) by DBRS and prices at SOFR + 2.25%. [3] The other is rated Ba2 by Moody's and BB+ by Fitch and prices at SOFR + 4.50%. [16] [2] The only stated difference between them is the credit quality of the customer whose contract the infrastructure serves. The hardware did not change. The customer did. The gap is 225 basis points of pure counterparty spread, and it is the single most instructive fact in the entire AI credit complex.
One thing must be said at the top rather than buried in a footnote: we could not read the rating agencies. Moody's rating action for the CoreWeave borrower returned HTTP 403 Forbidden to every method attempted, and S&P's publication on securitising data-centre equipment — a paper whose title is precisely the question at issue here — returned a security block page. [23] [24] No figure in this report is attributed to a rating agency's own words. Everything below is built from 8-Ks, credit agreements, 10-Ks, NPORT-P holdings files and the Bank of England. It turns out that is enough.
Leg One, Inverted: The Paper Is Shorter Than the Chips, and It Amortises
A maturity mismatch needs two numbers: the life of the asset and the tenor of the debt. Take the asset first, because the hypothesis quietly assumes a denominator that no issuer in the public record actually uses.
CoreWeave depreciates technology equipment over 6 years. [1] More to the point, it raised that estimate — from five years to six, effective January 1, 2023, "reflecting continuous advancements in hardware performance, software optimization, and data center design improvements." [1] Data center equipment and leasehold improvements run to the shorter of the remaining lease term or a useful life of up to 12 years. [1] Meta moved in the same direction: it increased the estimated useful lives of most servers and network assets to 5.5 years effective January 1, 2025, a change that reduced depreciation expense by $2.92 billion and increased net income by $2.59 billion, or $1.00 per diluted share, for the year ended December 31, 2025. [4] Those are the only two documented moves in the dossier, and both went up.
Now the numerator. CoreWeave's DDTL 4.0 Facility — its largest, an $8.5 billion delayed draw term loan — closed on March 31 2026 and matures in March 2032. [3] That is roughly six years of paper against a six-year book life. The DDTL 5.0 Facility, a $3.1 billion delayed draw term loan entered into on May 15 2026 with a maturity date of November 15, 2031, runs approximately 5.5 years. [2] [16] And CoreWeave does not merely tolerate this; it advertises it. The company describes DDTL 5.0 as "structured as a delayed draw term loan designed to align funding with the deployment schedule and useful life of the underlying GPU infrastructure assets." [16] The issuer's own marketing copy is a maturity-matching claim. Whatever else that is, it is not a mismatch being concealed.
Then comes the part for which the hypothesis has no room at all: these loans amortise. The principal of DDTL 1.0 "is required to be repaid in quarterly installments, with the final payment due on March 28, 2028." DDTL 2.0 repays "in quarterly installments, beginning in January 2026, with the final payment due five years after the applicable loan was funded." DDTL 3.0 repays "in monthly installments, beginning in April 2026, with the final payment due in August 2030." [1] A bullet bond sits at full face value against a collateral pool that depreciates underneath it. An amortising loan walks down alongside the asset — and in DDTL 2.0's case, began walking down in January 2026, while the GPUs it financed still had years of book life ahead of them.
The sizing works the same way. The 10-K states that the total loans available under the DDTL Facilities "are constrained by the purchase price of assets for which the loans are being used to finance with such percentage based upon the depreciable cost of graphics processing unit ('GPU') servers." [1] The advance rate is pegged to depreciable cost. As the collateral depreciates, so does the borrowing base. This is the opposite of the structure the hypothesis describes.
Zoom out to the whole stack and the picture holds. CoreWeave's total principal of debt was $21,615 million at December 31 2025, up from $8,033 million a year earlier — growth of $13,582 million in twelve months, which is a genuinely alarming number for other reasons. [1] But look at where it falls due: 2026, $6,708m; 2027, $4,298m; 2028, $2,393m; 2029, $1,769m; 2030, $2,109m; thereafter, $4,338m. [1] Only $4,338m of $21,615m — roughly a fifth, by our arithmetic — matures after 2030. The ladder is front-loaded, not long-dated. The nearest term problem CoreWeave has is a refinancing wall in 2026, not a residual hole in 2038.
The pattern repeats elsewhere. WULF Compute LLC, TeraWulf's financing subsidiary, offered $3.2 billion of senior secured notes due 2030. [9] And in the rated market, Charles River Associates plots the weighted-average life of A/A- rated data-centre ABS on a 0-12 year axis, measured as of November 17, 2025. [19] The multi-decade tenors are not where the compute collateral is.
CoreWeave's debt is front-loaded, not long-dated
Future principal payments on $21,615m of total debt principal at 31 December 2025. Only $4,338m — roughly a fifth, by our arithmetic — falls due after 2030.
Reconciling with the regulator, rather than duelling with it
The Bank of England's July 2026 Financial Stability Report is the strongest apparent support for the hypothesis, and reading it carefully is what dissolves the hypothesis. The FPC does say that "To date, debt issued to finance AI investment has tended to have longer (more than ten year) duration, in line with the appetites of investors in the deepest capital markets." But the very next sentence tells you what that money bought: "The majority of this debt has funded data centre buildings (ie the physical shell of the data centres) and facilities – rather than the servers and AI chips inside them." [15]
"issuing debt maturities aligned with the asset life of AI chips would imply shorter tenors than are typically preferred by most credit investors… The scale of this maturity mismatch risk across the debt financing of data centre facilities and AI chips, and how that risk is allocated across lenders, borrowers or tenants, will depend on the terms of the debt agreements." [15]
That is the language of a prospective structural tension, not an accomplished fact. The BoE is describing a conflict between what chips need and what credit investors want — and observing that the market has resolved it, so far, by lending long against buildings and short against chips. The two halves of the hypothesis, GPU security and long tenor, do not co-occur in any filing we could find.
Nebius makes the seam visible. The company closed convertible senior notes with a total aggregate original principal amount of $4.3375 billion — 1.250% notes due 2031 and 2.625% notes due 2033 — whose proceeds fund data-centre build-out and "the procurement of key components (including GPUs)." [11] Long-dated, chip-adjacent, and precisely the instrument the hypothesis predicts. Except that the notes "will be senior, unsecured obligations of the Company." [12] No GPU security interest. No special-purpose vehicle. The moment the paper goes long, the collateral disappears; the moment the collateral appears, the paper goes short. That is not a coincidence. It is the market pricing exactly the risk the hypothesis says it is ignoring.
The BoE is also careful about the asset-life input itself, and so should we be: "there is mixed evidence on how quickly AI chips depreciate: current shortages and demand for older chips support the argument for longer useful lives, while rapid innovation and efficiency gains could shorten them." [15] Nobody knows. What we can document is that the two issuers who changed their minds in public both revised upward.
The asset-life clock: every documented revision has gone UP
- 1 January 2023
CoreWeave raises computing-equipment life from five years to six[1]
The 10-K attributes the change to 'continuous advancements in hardware performance, software optimization, and data center design improvements.' Technology equipment is depreciated over 6 years; data centre equipment and leasehold improvements run to the shorter of the remaining lease term or a useful life of up to 12 years.
- 1 January 2025
Meta raises most servers and network assets to 5.5 years[4]
The change reduced depreciation expense by $2.92 billion and increased net income by $2.59 billion, or $1.00 per diluted share, for the year ended December 31, 2025. Servers and network assets carried a gross book value of $98,040 million at year end.
- 18 May 2026
CoreWeave markets DDTL 5.0 as a maturity-matching instrument[16]
The facility is 'structured as a delayed draw term loan designed to align funding with the deployment schedule and useful life of the underlying GPU infrastructure assets,' with 'a maturity of approximately 5.5 years.'
- July 2026
The Bank of England declines to settle it[15]
'There is mixed evidence on how quickly AI chips depreciate: current shortages and demand for older chips support the argument for longer useful lives, while rapid innovation and efficiency gains could shorten them.' The regulator treats the mismatch as a prospective tension, not an accomplished fact.
The honest residual worry here is the inverse of the one commissioned. The term risk in this chain is long buildings financed with money that must be rolled, and a refinancing market that has to stay open. JP Morgan, cited by the Bank of England, estimates that over $2 trillion of aggregate funding may be needed for the AI chips used to train and run AI models over the next five years — an estimate, and a very large one. [15] Almost none of that has been raised yet. What exists today is a small, short, amortising, heavily-guaranteed book.
The Natural Experiment: Same Chips, Same Bank, 225 Basis Points
Financial journalism rarely gets a controlled experiment. This is one.
Two facilities. Seven weeks apart. Same sponsor — CoreWeave. Same collateral type — GPU servers and related infrastructure. Same Delaware special-purpose wrapper. Same two names at the top of the ticket: MUFG and Morgan Stanley served as co-structuring agents and joint bookrunners on the first [3], and Morgan Stanley and Mitsubishi UFJ Financial Group served as joint lead arrangers and bookrunners on the second. [16] One variable moves.
DDTL 4.0, closed March 31 2026. An $8.5 billion delayed draw term loan, allowing CoreWeave to borrow up to approximately $7.5 billion initially with the ability to increase total borrowing capacity to $8.5 billion as underlying assets reach stabilization, secured by substantially all assets of CoreWeave Compute Acquisition Co. VIII, LLC. [3] CoreWeave's own 8-K exhibit describes it as "the first investment-grade rated financing secured by HPC infrastructure and an associated customer contract." [3] Note the conjunction. The infrastructure and the customer contract. The facility received ratings of A3 by Moody's and A (low) by DBRS, and includes a floating rate tranche financed at SOFR + 2.25% and a fixed rate tranche financed at approximately 5.9%. [3] It was anchored by Blackstone Credit & Insurance, with Goldman Sachs and JPMorgan as additional coordinating lead arrangers, and was meaningfully oversubscribed. [3]
DDTL 5.0, signed May 15 2026. A $3.1 billion delayed draw term loan through CoreWeave Financing DDTL V, LLC, a Delaware limited liability company. [2] Proceeds "support the purchase and deployment of infrastructure dedicated to customer contracts with two large, non-investment grade customers." [16] Ratings: Ba2 from Moody's, BB+ from Fitch. Pricing: daily compounded SOFR plus an applicable margin of 4.50% per annum. [2] And that is the tight number — the transaction was meaningfully oversubscribed and pricing tightened by 50 basis points from initial discussions. [16]
The natural experiment: one variable moves
| CoreWeave DDTL 4.0 | CoreWeave DDTL 5.0 | |
|---|---|---|
| Closed / signed[3][2] | 31 March 2026 | 15 May 2026 |
| Size[3][2] | $8.5 billion delayed draw term loan (approx. $7.5bn initially, rising to $8.5bn at stabilization) | $3.1 billion delayed draw term loan |
| Borrowing vehicle[3][2] | CoreWeave Compute Acquisition Co. VIII, LLC | CoreWeave Financing DDTL V, LLC — 'a Delaware limited liability company' |
| Collateral[3][2] | Substantially all assets of the borrower — HPC infrastructure and an associated customer contract | Substantially all assets of the borrower and its subsidiaries, plus a pledge of 100% of the equity in the borrower |
| Arrangers[3][16] | MUFG and Morgan Stanley (co-structuring agents, joint bookrunners); Goldman Sachs and JPMorgan; anchored by Blackstone Credit & Insurance | Morgan Stanley and Mitsubishi UFJ Financial Group (joint lead arrangers and bookrunners) |
| Customer credit quality[3][16] | Investment grade — 'the first investment-grade rated financing secured by HPC infrastructure and an associated customer contract' | 'customer contracts with two large, non-investment grade customers' |
| Ratings[3][16] | A3 (Moody's) / A (low) (DBRS) | Ba2 (Moody's) / BB+ (Fitch) |
| Floating-rate pricing[3][2] | SOFR + 2.25% (fixed tranche at approximately 5.9%) | Daily compounded SOFR + 4.50%, 0.00% floor — after tightening 50bp during syndication |
| Maturity / tenor[3][2] | March 2032 — approximately 6 years, against a stated 6-year GPU book life | 15 November 2031 — approximately 5.5 years |
| Recourse[3][2] | Marketed as a 'First of a kind non-recourse facility' — in a press release | 'All obligations under the DDTL 5.0 Facility are unconditionally guaranteed by the Parent' — in the 8-K |
| The gap[3][2] | 225 basis points of floating-rate spread, and two rating categories, on functionally identical hardware — our subtraction of two verbatim margins | 225 basis points of floating-rate spread, and two rating categories, on functionally identical hardware — our subtraction of two verbatim margins |
The arithmetic is ours and we own it: SOFR + 4.50% minus SOFR + 2.25% is 225 basis points. Both margins are stated verbatim in filings; the subtraction is the reporter's. And the subtraction is the finding. Two loans against the same class of hardware, arranged by the same banks for the same borrower inside the same quarter, separated by two full rating notches and 225 basis points of spread, with one disclosed difference between them — whether the customer paying for the compute is investment grade.
If the chips were the credit, that cannot happen. GPUs do not become more valuable when Microsoft rents them and less valuable when someone else does. The collateral is doing recovery work — it determines what a lender gets back after a default. The tenant is doing probability-of-default work — it determines whether there is a default at all. Ratings and spreads are driven overwhelmingly by the second. That is why the same GPUs can carry an A3 and a Ba2 in the same quarter.
The covenant machinery says the same thing, in black and white
Pricing is an inference. Documents are not. The DDTL 5.0 credit agreement "also contains events of default related to certain adverse events with respect to certain material contracts." [2] Read that again: the loan can be defaulted by something going wrong with the customer contract, independent of anything happening to the collateral value. The lender has written its exit around the tenant, not the hardware. The borrower must also maintain a debt service coverage ratio of at least 1.35x — a cash-flow covenant, measured on contracted revenue. [2]
And you can see which contracts matter. CoreWeave had $60.7 billion of remaining performance obligations at December 31 2025 against $15.1 billion a year earlier, with a weighted-average contract duration of approximately five years. [1] Customer A accounted for 67% of 2025 revenue and 68% of accounts receivable; the 10-K names Microsoft as the top customer. [1] OpenAI "has committed to pay us up to approximately $6.5 billion through May 31, 2031" and Meta "initially committed to pay us up to approximately $14.2 billion through December 2031." [1] Revenue was $5.1 billion in 2025 with a net loss of $1.2 billion. [1] What a lender to CoreWeave is really lending against is a short list of promises from a short list of names.
TeraWulf writes it out even more plainly. The security package for WULF Compute LLC's $3.2 billion of senior secured notes due 2030 consists of first-priority liens on (i) substantially all assets of WULF Compute and the guarantors, (ii) all equity interests of WULF Compute held by TeraWulf Brookings LLC, (iii) "a designated lockbox account of Fluidstack USA I Inc." — the tenant's cash — and (iv) "a pledge by Google LLC of warrants to purchase common stock of TeraWulf." [9] The hardware is in there somewhere, under "substantially all assets." Everything else in the package is a claim on somebody's covenant. And behind it stands Google, whose total backstop for TeraWulf project-related debt financing increased to approximately $3.2 billion — the same order of magnitude as the note issue itself — in exchange for warrants to acquire 32.5 million TeraWulf shares, lifting its pro forma equity ownership to approximately 14%. [10] The Fluidstack CB-5 lease that triggered it adds an incremental 160 MW of critical IT load, taking total contracted critical IT load at Lake Mariner to approximately 360 MW and representing "$6.7 Billion in Contracted Revenue, with Potential to Reach $16 Billion with Lease Extensions." [10]
Nobody is underwriting silicon here. They are underwriting Google.
What the security package actually pledges
| Instrument | Nominal collateral | The counterparty machinery | Rating / price |
|---|---|---|---|
| CoreWeave DDTL 4.0 ($8.5bn) | Substantially all assets of CoreWeave Compute Acquisition Co. VIII, LLC | Rated on 'HPC infrastructure AND an associated customer contract' — an investment-grade customer | A3 / A (low); SOFR + 2.25% |
| CoreWeave DDTL 5.0 ($3.1bn) | Substantially all assets of the borrower plus a 100% equity pledge | Proceeds serve 'two large, non-investment grade customers'; events of default attach to 'certain material contracts'; DSCR of at least 1.35x; unconditional parent guarantee | Ba2 / BB+; SOFR + 4.50% |
| WULF Compute LLC notes ($3.2bn, due 2030) | First-priority liens on substantially all assets of WULF Compute and the guarantors | Plus the equity of WULF Compute, 'a designated lockbox account of Fluidstack USA I Inc.' (the tenant's cash), and 'a pledge by Google LLC of warrants'; Google's total backstop is approximately $3.2 billion | 7.75% coupon, due 2030-10-15 |
| Beignet Investor LLC ($27.3bn, due 2049) | A purpose-built data-centre campus in Richland Parish, Louisiana, designed for up to 5 gigawatts | Meta is the tenant on an initial four-year lease term; residual value guarantees with an aggregate threshold of approximately $28 billion; Meta's maximum exposure to loss $45.95 billion | A+ (S&P, per Fortune/Forbes); 6.581% to 2049-05-30 |
| CoreWeave DDTL 1.0 / 2.0 / 2.1 / 3.0 | Loans 'constrained by… the depreciable cost of graphics processing unit ("GPU") servers' — the advance rate falls as the chips depreciate | 'All obligations under the DDTL Facilities are unconditionally guaranteed by the Company' — full parent recourse | DDTL 1.0: 15% effective. DDTL 2.0: 10% effective |
And while we are reading the documents: "non-recourse" is marketing
CoreWeave's DDTL 4.0 press release bills it as a "First of a kind non-recourse facility." [3] The 10-K, describing DDTL 1.0, 2.0, 2.1 and 3.0, says: "all obligations under the DDTL Facilities are unconditionally guaranteed by the Company." [1] The DDTL 5.0 8-K, describing the newest facility of all, says: "All obligations under the DDTL 5.0 Facility are unconditionally guaranteed by the Parent pursuant to a parent guarantee and pledge agreement." [2] Parent recourse is the template. DDTL 4.0 — the investment-grade one, the one with the investment-grade customer — is the exception, and the claim appears only in a marketing exhibit. The idea that GPU paper has been ring-fenced away from the operating company's balance sheet does not survive contact with the filings.
The Bank of England reaches the same conclusion from thirty thousand feet, and states it as a principle: "The riskiness of this debt depends on the underwriting terms, in particular the quality of the leases and guarantees which back debt holders' claims." [15] So does the methodology work, where we can read it. KBRA's proposed data-centre ABS approach "begins with evaluating the underlying collateral's financial and operating performance, starting with estimating sustainable net cash flow (KNCF) for the properties securing the transaction," stress-tests "multiple rounds of tenant defaults, a lag period to find replacement tenants, and revenue haircuts applied to replacement tenants," and determines property residual values using "stressed terminal capitalization rates." [17] That is a real-estate technique. It is not an equipment-residual technique. There is no chip in it. (KBRA's free page discloses no useful-life assumptions or numerical thresholds; the full methodology sits behind a subscription wall, and we say so.) Charles River's survey reports that Moody's data-centre securitization methodology, released February 2025, turns on "contract renewal, technological obsolescence, the timing of cash flows to repay debt, leverage ratios and supply/demand fundamentals" — obsolescence is named, but as one input to a cash-flow rating rather than as a collateral haircut. [19]
We could not read Moody's or S&P directly. [23] [24] We did not need to. The issuer published the ratings, the prices, and the customer credit quality itself, in documents filed with the SEC.
Price It Yourself: The Chip Slider Does Nothing, the Tenant Switch Does Everything
The hypothesis and this report make different predictions, and the difference can be dialled. If the chips are the credit, then changing the assumed useful life of a GPU should change what the paper costs. If the tenant is the credit, it should not.
So build the loan. The default state below reproduces DDTL 4.0: an investment-grade customer, a 6-year asset life, a roughly six-year tenor, quarterly amortisation. [3] [1] In that state the outstanding principal runs beneath the net book value of the collateral for essentially the whole life of the facility, because the facility is sized off depreciable cost and repays on a schedule. [1] There is no residual hole. There is nothing for the hypothesis to bite on.
Then build the instrument the hypothesis describes. Drag the assumed asset life down to four years. Push the tenor out. Switch amortisation off, so the balance sits flat at face until a bullet maturity. Now a residual gap opens and widens: a loan standing above the book value of the thing securing it, for years. That instrument would be a genuine maturity mismatch, and it would deserve every alarm the hypothesis raises about it.
It does not exist. Not in any filing we could find as of July 2026. Producing one is the first and cleanest way to falsify this report.
Build the loan: the chip sliders move the collateral, the customer moves the price
And then the punchline, which is a behaviour rather than a number: the spread readout does not move when you drag the chip slider. It sits at SOFR + 2.25% and it stays there — through a four-year life, an eight-year life, a bullet, an amortiser. [3] It moves exactly once, when you flip the customer from investment grade to non-investment grade, and it snaps to SOFR + 4.50%. [2] That is not a modelling artefact. It is the only structure the public record supports, because the record contains exactly two priced points and the thing that differs between them is the customer.
The bounds of that model should be stated in prose and not only in fine print. Book depreciation is not market value; a straight-line net book value is an accounting line, not a bid. The advance rate is undisclosed — the 10-K says only that available loans are "constrained by the purchase price of assets… based upon the depreciable cost of graphics processing unit ('GPU') servers," and the reader is setting the number, not reading it. [1] The two spread anchors are two real transactions, not a fitted curve; the model interpolates nothing, it snaps between the only two points the record gives us. And no rating agency's model is being replicated here, because none of them would let us read one. [23] [24]
Leg Three, Refuted to Zero: The Paper Is in Delaware, in Plain Sight
Do the count first. A programmatic scan finds the word "Cayman" zero times in CoreWeave's FY2025 10-K. Zero times in Meta's FY2025 10-K. And zero times across all 292,346 characters of Nebius's 424B5 prospectus supplement. [1] [4] [13] Three documents, three zeroes.
Now the register of vehicles. CoreWeave Financing DDTL V, LLC — "a Delaware limited liability company," in the 8-K's own words. CW Financing DDTL V Holdco, LLC, the pledgor above it — the same phrase. [2] CoreWeave Compute Acquisition Co. VIII, LLC, the DDTL 4.0 borrower. [3] WULF Compute LLC, guaranteed by La Lupa Data LLC, Akela Data Holdings LLC and Akela Data LLC — all LLCs. [9] Centersquare Issuer LLC and Centersquare Co-Issuer LLC, "each a special-purpose entity and indirect wholly-owned subsidiary of Phoenix Data Center Acquisitions LLC," which issued $885.0 million of asset-backed securitized notes on October 17 2024, an additional $940.0 million on March 20 2025 and a further $815.0 million on August 21 2025, with Wilmington Trust as trustee. [14] Beignet Investor LLC, the largest issuer in the whole sector, which two independent registered-fund holdings reports identify with ISIN US076912AA21 and invCountry US. [5] [6] And Nebius Group N.V., which describes itself in its own prospectus supplement as "a Dutch public limited company." [13]
Onshore, onshore, onshore, onshore, onshore, onshore — and Amsterdam.
The register of vehicles: every issuer with a filing
| CoreWeave Financing DDTL V, LLC | DDTL 5.0 ($3.1bn) | 'a Delaware limited liability company' | Borrower |
| CW Financing DDTL V Holdco, LLC | DDTL 5.0 | 'a Delaware limited liability company' | Holdco / pledgor of 100% of the borrower's equity |
| CoreWeave Compute Acquisition Co. VIII, LLC | DDTL 4.0 ($8.5bn) | US LLC | Borrower; substantially all its assets secure the facility |
| WULF Compute LLC (with La Lupa Data LLC, Akela Data Holdings LLC, Akela Data LLC) | $3.2bn senior secured notes due 2030 | US LLCs | Rule 144A issuer and guarantors |
| Beignet Investor LLC | $27.3bn notes due 2049 (Meta / Blue Owl 'Hyperion') | ISIN US076912AA21; invCountry US; assetCat DBT / issuerCat CORP | Issuer — the largest instrument in the sector |
| Centersquare Issuer LLC / Centersquare Co-Issuer LLC | $885.0m ABS notes (Oct 2024), plus $940.0m and $815.0m taps | 'each a special-purpose entity and indirect wholly-owned subsidiary of Phoenix Data Center Acquisitions LLC' | Data-centre ABS co-issuers; Wilmington Trust as trustee |
| Nebius Group N.V. | $4.3375bn convertible senior notes due 2031 / 2033 | 'a Dutch public limited company'; zero occurrences of 'Cayman' in 292,346 characters of its 424B5 | Issuer — senior, UNSECURED; no GPU security, no SPV |
| BIF III GP (Cayman) L.P. | Csquare, Inc. S-1 (Centersquare platform) | 'PO Box 309 Ugland House Grand Cayman KY1-1104 Cayman Islands' | The only genuine Cayman entity in the dossier — general partner of Brookfield's EQUITY stockholder. It issues no debt. |
There is exactly one genuine Cayman entity anywhere in this dossier, and its location tells you something. The Csquare, Inc. S-1 contains 7 occurrences of "Cayman," every one of them the same footnote about the same entity: "BIF III GP (Cayman) L.P. ('BIF III GP Cayman') serves as the general partner of the Brookfield Stockholder… The address of BIF III GP Cayman is PO Box 309 Ugland House Grand Cayman KY1-1104 Cayman Islands." [14] Ugland House, as advertised. But BIF III GP Cayman is the general partner of Brookfield's equity stockholder. It is on the sponsor side of the cap table. It does not issue debt. The ABS issuers in that very same document are US special-purpose entities. Cayman is where a private-equity fund's general partner sits, not where the compute paper is issued.
The negative search points the same way. An unrestricted EDGAR full-text search for "graphics processing unit" together with "Cayman" returned 161 hits, and the top hits on inspection were special-purpose acquisition companies — Willow Lane Acquisition Corp., Idea Acquisition Corp. — and offshore-listed operating companies such as Starbox Group and ECARX Holdings. [8] Cayman appears in the GPU corpus as a listing holdco domicile. Never as a collateral SPV.
'Cayman' in the Csquare S-1
7
All seven are BIF III GP (Cayman) L.P. — the GP of Brookfield's equity stockholder, on the sponsor side
[14]EDGAR documents containing 'GPU-backed'
6
Two are CoreWeave 8-K exhibits; the asset class barely exists as a named category
[8]'graphics processing unit' + 'Cayman' full-text hits
161
Inspected top hits were SPACs and offshore listing holdcos — never a collateral SPV
[8]Where did the offshore limb come from? The trail leads to a promotional article published by Cayman Finance — the Cayman Islands' own financial-services industry association — written by a law-firm practitioner, which hedges throughout ("can be used," "increasingly preferred"), names no GPU or AI-compute transaction anywhere, and is explicitly scoped to Asia: "Asia has seen a steady rise in securitisation activity, particularly in markets such as Singapore, Hong Kong, and Japan." [22] It is industry marketing. It should not be cited for a number, and we do not cite it for one. It is reported here only as the origin of a claim that the filings destroy.
And the interesting part is not the debunk. It is what the debunk implies. The risk in this chain is not being hidden by geography. It is disclosed, in Delaware limited liability companies, in 8-Ks and credit agreements you can read on a Sunday afternoon. What is genuinely opaque is the covenant stack inside those entities — which is exactly what the Bank of England flags when it warns that capital is being sought "through securitised data centre and other asset-backed structures, special purpose vehicles, and other bespoke financing arrangements," that this "can result in higher asset-level leverage," and that it "increases the complexity of identifying where risk ultimately sits." [15] The FPC names special purpose vehicles. It does not name Cayman. Neither should anyone else.
One further measure of how young this is: an EDGAR full-text search for the exact phrase "GPU-backed" returns only 6 documents across the entire index — two CoreWeave 8-K exhibits, two SharonAI documents, and two micro-cap filings. [8] As a named, disclosed asset class in SEC filings, GPU-backed debt is very nearly one company.
Beignet: A 2049 Bond Standing on a Four-Year Tenancy
Which brings us back to the four years, and to the one place in this whole chain where a real, documented, structural mismatch is sitting in the open.
The bond first, from primary sources rather than the press. Two independent registered-fund NPORT-P holdings reports identify the same instrument: issuer BEIGNET INVESTOR LLC, CUSIP 076912AA2, ISIN US076912AA21, maturity date 2049-05-30, fixed coupon, annualized rate 6.581%, assetCat DBT, issuerCat CORP, invCountry US. [5] [6] Fortune independently reports a "yield of 6.58% at issue" and an "A+ rating from S&P." [21] Forbes puts the size at $27.3 billion, with Blue Owl-managed funds owning 80% and Meta 20%, Meta as the tenant leasing the entire campus, Morgan Stanley as arranger, on a site designed for up to 5 gigawatts. [20]
The lease next, from Meta's own 10-K: the venture's leases commence in 2029, the aggregate initial lease commitment is approximately $12.31 billion, and each property has "an initial four-year lease term and options to renew for a total lease period of up to 20 years." [4]
Do the subtraction — and it is ours, not a printed figure: a bond maturing 2049-05-30, held today, against an initial four-year lease term, is roughly 23 to 24 years of paper standing on four years of contracted tenancy. That is a real mismatch. It is documented in primary filings on both sides. And it has nothing whatever to do with how long a GPU lasts. It is a lease-versus-bond mismatch, on a building.
Something has to bridge a twenty-year gap, and it is not the collateral. It is Meta's balance sheet. The company has "provided residual value guarantees (RVG) with an aggregate threshold of approximately $28 billion that decreases over time," and if it decides to terminate or not renew a lease, and certain other conditions are met, its maximum RVG payment would equal any shortfall between the fair value at that time and the RVG threshold for that property. [4] Meta's "maximum exposure to loss related to the Venture was $45.95 billion as of December 31, 2025," consisting of $1.83 billion of carrying value of its equity investment, the lease commitments, its estimated future fundings, and the maximum RVG threshold. [4] As of December 31 2025, RVG payments are not probable and no liability has been recorded. [4]
And Meta does not consolidate any of it. It holds a 20% membership interest, accounts for the venture under the equity method, and states: "We do not have the power to direct the activities that most significantly impact the Venture's economic performance." [4] At formation it contributed $4.30 billion of held-for-sale assets net of liabilities and received a one-time distribution of $2.55 billion; the parties committed to fund their pro rata share of approximately $27 billion in total estimated development costs. [4]
So: a bond that runs to 2049, a lease that runs four years, and in between, a residual value guarantee from a tenant that does not consolidate the entity. This is the tenant's covenant doing the credit work again — the same finding as the CoreWeave loans, one campus over, at a scale of tens of billions.
Who ends up holding it? An insurance separate account, among others. The Brighthouse Balanced Plus Portfolio's NPORT-P shows $14,800,000 par of Beignet Investor LLC, valued at $15,216,711.02, being 0.335134740605% of the fund. [6] Fortune reports that "BlackRock bought more than $3 billion of bonds that the joint venture (dubbed Beignet) issued last week." [21] The phrase "Beignet Investor" appears in 2245 NPORT-P filings — with the important caveat that NPORT-P filings are per-fund and per-quarter, so that is a filing count and not a count of distinct funds, and should be read as an indication of breadth only. [8] The same caveat applies to "WULF Compute," which appears in 3228. [8]
These bonds are not exotic instruments locked in a private vault. CoreWeave's 9.25% senior notes due 2030-06-01 sit in the Virtus AI & Technology Opportunities Fund, a closed-end fund, marked below par: $13,990,000 of par carried at $13,771,730.26. [7] WULF Compute LLC's 7.75% notes due 2030-10-15 sit in the same fund, $5,245,000 par valued at $5,470,639.90. [7] Retail-accessible wrappers, insurance separate accounts, and the biggest bond manager on earth.
Everything in years, side by side
The GPU-secured paper clusters at 5.5-6 years, inside the issuers' own book lives. The one multi-decade instrument in the chain is a bond on a building — standing on a four-year lease.
One discipline is essential here: do not generalise Beignet to the market. It is a bespoke 144A instrument, and the funds that hold it classify it as assetCat DBT / issuerCat CORP — a plain corporate bond, not structured paper. [5] The rated data-centre ABS market is a different animal: through the first quarter of 2025, S&P had rated 42 data center ABS issuances totaling $16.2 billion in transaction value, and as of December 2024, eighty-four percent of the issuances received credit ratings of A from S&P. [19] Those deals run inside twelve years of weighted-average life. [19] The February 2026 vintage looks like real estate because it is real estate: Compass Datacenters raised $830 million against a $3.6 billion, 198.2 megawatt portfolio of 6 fully leased buildings in Phoenix and Toronto, with a $500 million senior tranche rated AAA by Moody's and the remainder at Aa3 and A2, priced at approximately 1.20 percentage points over benchmark; DataBank Holdings raised $665 million covering 36 data centers, 258 megawatts, 84% leased by floor area, approximately 1,750 customers. [20] Buildings, tenants, leases, occupancy. And at the project level, Crusoe's Abilene campus is funded through a $15 billion joint venture with funds managed by Blue Owl's Real Assets platform and Primary Digital Infrastructure, sponsoring six new buildings to bring the 1.2 gigawatt campus to eight, each designed to operate up to 50,000 NVIDIA GB200 NVL72s. [18] The chips are the tenant's problem. The debt is against the shell.
Beignet is what makes this reportable rather than merely clever, because it comes with a date. The leases commence in 2029 with four-year initial terms. [4] Whether Meta renews across the campus is a checkable event with a year attached — and it is the single observation that would convert the "tenant's covenant is the credit" framing from an argument about pricing into an argument about loss.
What Would Break This — and Why It Is Worth Watching For
The hypothesis is not foolish. It is early. That distinction matters, and it deserves to be argued at full strength rather than knocked down.
The mechanism it describes is entirely coherent: an asset whose life is short and contested, financed at enormous scale, through structures whose covenant terms are not fully public, and sold to investors who structurally prefer long paper. Every link in that chain is real. The Bank of England says as much in terms — the scale of maturity mismatch risk "will depend on the terms of the debt agreements," and "there is mixed evidence on how quickly AI chips depreciate." [15] The volumes are about to arrive: JP Morgan, cited by the BoE, estimates that over $2 trillion of aggregate funding may be needed for AI chips over the next five years. [15] The OECD, also cited by the BoE, estimated that the share of private credit financing AI investment increased from 9% in 2024 to 34% in 2025. [15] AI issuers accounted for 41% of non-refinance-related US high-yield issuance in the year so far, despite accounting for only 1% of the JP Morgan HY bond index at the end of 2025. [15] The financing of the chips has barely begun. Our finding is about what exists in July 2026 — not about what cannot exist in 2028.
And the record is thin. "GPU-backed" returns 6 documents across the whole EDGAR full-text index. [8] Our natural experiment is, in the end, one issuer's two loans. It is a very clean experiment — but a sample of one sponsor is a sample of one sponsor, and it should be said in plain language rather than smuggled past the reader.
The largest hole is in the sourcing, and it is the one that could kill the central claim outright. S&P's "ABS Frontiers: Equipping Data Centers Through Securitization" — a publication whose title is precisely the question of whether data-centre equipment can be securitised on its own collateral — could not be read; it returned a security block page. Moody's rating action for the DDTL 4.0 borrower returned HTTP 403 Forbidden to both an automated fetch and a browser user-agent. [23] [24] If those documents show the agencies applying explicit GPU residual-value assumptions or equipment-ABS haircuts, then the claim that the collateral does no rating work is wrong, and this report's central contrast collapses. That is not a decorative caveat. It is the biggest single gap in the dossier, and a reader with a terminal subscription can close it in an afternoon.
What would break this report
| If someone produced this… | …it would overturn | Where it stands, July 2026 |
|---|---|---|
| A GPU-secured facility whose maturity runs beyond BOTH the customer contract and the asset life — say a 12-year bullet against a five-year contract | The whole inversion of Leg 1. The maturity-mismatch hypothesis would be alive again. | No such instrument in the public record. DDTL 4.0 matures March 2032 against a six-year book life; DDTL 5.0 runs ~5.5 years; DDTL 2.0 began amortising in January 2026. |
| The S&P and Moody's rating rationales themselves — in particular S&P's 'ABS Frontiers: Equipping Data Centers Through Securitization' | The central contrast. If the agencies apply explicit GPU residual-value assumptions or equipment-ABS haircuts, the claim that the collateral does no rating work is simply wrong. | Unreadable. Moody's returned HTTP 403 Forbidden to both an automated fetch and a browser user-agent; S&P returned a 'Security Controls Triggered' block page. This is the single biggest gap in the dossier. |
| Meta declining to renew the Hyperion leases when the initial four-year terms expire | Nothing in the argument — but it converts 'the tenant's covenant is the credit' from a claim about pricing into a claim about loss, and tests what a residual value guarantee on a purpose-built AI campus is really worth. | Checkable on a date. Leases commence in 2029; the RVG threshold is approximately $28 billion and decreases over time; Meta's maximum exposure to loss is $45.95 billion. |
| A Cayman-domiciled issuer of GPU or compute-backed paper, in an actual filing | The refutation of Leg 3. | Zero found. Three programmatic zero-counts, a register of Delaware/US LLCs, and one Cayman entity — a private-equity fund's general partner, on the equity side. |
| A data-centre ABS that blows through its anticipated repayment date into the legal-final tail | 'Tenor is matched' would need rewriting as 'tenor is matched only if the refinancing market stays open.' The risk becomes a refi wall — the inverse of the hypothesis, not the hypothesis. | Rated data-centre ABS weighted-average life is still plotted on a 0-12 year axis, measured as of 17 November 2025. Everyone underwrites to the refinancing. |
| A credible dataset showing used H100/GB200 prices collapsing inside three years | The premise, if not the four-year figure's provenance. It would make the amortising, depreciable-cost-sized structure of the CoreWeave loans the only thing standing between lenders and a residual hole. | Contested. The only two documented issuer revisions both went UP: CoreWeave five to six years; Meta to 5.5. The BoE calls the evidence 'mixed'. |
There is one further thread that would corrode everything here at once. The BoE notes that "Some AI-related companies' revenue forecasts may reflect 'circular financing arrangements'. One example of this is when technology companies invest in AI companies which in turn purchase those technology companies' products. This creates self-reinforcing capital loops." [15] Look again at the TeraWulf structure with that sentence in hand: Google pledges warrants as collateral, backstops the project debt to approximately $3.2 billion, and takes approximately 14% of the equity, while the tenant whose lockbox secures the notes rents capacity that ultimately serves AI demand. [9] [10] If the tenant is the credit — and this report argues that it plainly is — then the quality of the tenant's own revenue is the credit, all the way down. Underwriting the covenant is only safer than underwriting the chips if the covenant is not itself an option on the same thing.
But that is the risk that exists, and it is not the one that was commissioned. On the record as it stands in July 2026: the GPU paper is short, it amortises against a book life that issuers keep revising upward, it is unconditionally guaranteed by its parent in almost every case, and it is filed in Delaware, in public, where anyone can read it. What it lends against is not a rack of silicon. It is a customer's promise to pay — and the price of that promise, on identical hardware, arranged by the same bank, in the same quarter, differs by 225 basis points depending on whose name is on it. The four-year clock everyone is watching is real. It is ticking on a building in Louisiana, and it starts in 2029.
What would change our mind
- A GPU-SECURED FACILITY WHOSE STATED MATURITY EXTENDS BEYOND BOTH THE CUSTOMER CONTRACT AND THE ASSET LIFE. Our whole rebuttal of the maturity-mismatch limb rests on the fact that no such instrument exists in the public record as of July 2026: DDTL 4.0 matures March 2032 against a six-year book life; DDTL 5.0 runs ~5.5 years; DDTL 2.0 tranches mature five years after each funding and began amortising in January 2026 — i.e. repaid before the GPU is fully depreciated. Produce one GPU-backed loan or bond with, say, a 12-year bullet against a five-year contract and the hypothesis is alive again.
- META NOT EXERCISING THE YEAR-FOUR RENEWALS AT HYPERION. The Beignet bond matures 2049; each property carries an initial four-year lease term with renewal options to 20 years, bridged by a residual value guarantee with a threshold of approximately $28 billion that decreases over time. The checkable event a reader can actually track is whether Meta renews across the campus when the initial terms expire. Non-renewal would convert our 'the tenant's covenant is the credit' framing from an argument about pricing into an argument about loss — and would test whether a residual value guarantee on a purpose-built AI campus is worth what the rating implies.
- A DATA-CENTRE ABS THAT BLOWS THROUGH ITS ANTICIPATED REPAYMENT DATE INTO THE LEGAL-FINAL TAIL. Rated data-centre ABS has a weighted-average life plotted on a 0-12 year axis even where legal finals sit decades out, because everyone underwrites to the ARD refinancing. If 2026-2027 deals start extending past their ARDs rather than refinancing, the risk is not a long bond against a short asset — it is a refi wall, the inverse of the hypothesis, and our 'tenor is matched' conclusion would need rewriting as 'tenor is matched only if the refinancing market stays open'.
- A CAYMAN-DOMICILED ISSUER OF GPU OR COMPUTE-BACKED PAPER, IN A FILING. We drop the Cayman limb on the strength of a dozen US/Delaware issuing entities, three programmatic zero-counts for the word 'Cayman' (CoreWeave 10-K, Meta 10-K, Nebius 424B5), and NPORT-P holdings that report every instrument with a US ISIN and invCountry US. One offering circular or subsidiary schedule placing a compute-collateral SPV in Grand Cayman — as opposed to a private-equity fund's general partner, which is what we did find — would reopen it.
- THE S&P AND MOODY'S RATING RATIONALES THEMSELVES. Both were unreachable (HTTP 403 / 'Security Controls Triggered'), and S&P's 'ABS Frontiers: Equipping Data Centers Through Securitization' — which by its title addresses securitising data-centre EQUIPMENT, the exact question here — was never read. If those documents show the agencies applying explicit GPU residual-value assumptions or equipment-ABS haircuts, our claim that the collateral does no rating work is wrong, and the report's central contrast collapses.
- EVIDENCE THAT USED-GPU RESALE VALUES ARE COLLAPSING FASTER THAN THE BOOKS ASSUME. Our position is that the asset-life input is contested and that the only two moves we can document (CoreWeave five to six years; Meta to 5.5 years) went up. A credible, citable dataset showing second-hand H100/GB200 prices falling to a small fraction of cost inside three years would vindicate the hypothesis's premise even though its four-year figure came from the wrong place — and would make the amortising, depreciable-cost-sized structure of the CoreWeave loans look like the only thing standing between lenders and a residual hole.
Sources
- [1]T1 · Primary · filing
CoreWeave, Inc. Form 10-K for FY2025 (filed 2026-03-02) — SEC EDGAR / CoreWeave, Inc., 2026-03-02 - [2]T1 · Primary · filing
CoreWeave, Inc. Form 8-K — DDTL 5.0 Facility credit agreement (event date 2026-05-15) — SEC EDGAR / CoreWeave, Inc., 2026-05-18 - [3]T1 · Primary · filing
CoreWeave, Inc. Form 8-K Exhibit 99.1 — $8.5bn DDTL 4.0 Facility, first investment-grade GPU-backed financing (2026-03-31) — SEC EDGAR / CoreWeave, Inc., 2026-03-31 - [4]T1 · Primary · filing
Meta Platforms, Inc. Form 10-K FY2025 — Note 5 (the Hyperion 'Venture' VIE) and server useful lives — SEC EDGAR / Meta Platforms, Inc., 2026-01-29 - [5]T1 · Primary · filing
Aristotle Funds Series Trust Form NPORT-P — holdings detail showing BEIGNET INVESTOR LLC bond — SEC EDGAR / Aristotle Funds Series Trust, 2026-05-27 - [6]T1 · Primary · filing
Brighthouse Funds Trust I Form NPORT-P — Brighthouse Balanced Plus Portfolio holding of Beignet Investor LLC — SEC EDGAR / Brighthouse Funds Trust I, 2026-05-28 - [7]T1 · Primary · filing
Virtus Artificial Intelligence & Technology Opportunities Fund (AIO) Form NPORT-P — CoreWeave and WULF Compute bond holdings — SEC EDGAR / Virtus AI & Technology Opportunities Fund, 2026-03-31 - [8]T1 · Primary · filing
SEC EDGAR full-text search — count of NPORT-P fund filings disclosing holdings in each AI-credit issuer — SEC EDGAR Full-Text Search (efts.sec.gov), 2026-07-14 - [9]T1 · Primary · filing
TeraWulf Inc. Form 8-K Exhibit 99.1 — WULF Compute LLC $3.2bn senior secured notes offering — SEC EDGAR / TeraWulf Inc., 2025-10-14 - [10]T1 · Primary · filing
TeraWulf Inc. Form 8-K Exhibit 99.1 — Fluidstack CB-5 lease expansion and Google backstop increase — SEC EDGAR / TeraWulf Inc., 2025-08-18 - [11]T1 · Primary · filing
Nebius Group N.V. Form 6-K Exhibit 99.1 — closing of $4.3375bn convertible senior notes — SEC EDGAR / Nebius Group N.V., 2026-03-20 - [12]T1 · Primary · filing
Nebius Group N.V. Form 6-K Exhibit 99.1 — proposed $3.75bn convertible senior notes offering (terms and structure) — SEC EDGAR / Nebius Group N.V., 2026-03-17 - [13]T1 · Primary · filing
Nebius Group N.V. Form 424B5 prospectus supplement (September 2025) — SEC EDGAR / Nebius Group N.V., 2025-09-01 - [14]T1 · Primary · filing
Csquare, Inc. Form S-1 (Centersquare data centre platform, Brookfield-backed) — ABS co-issuer structure and Cayman GP — SEC EDGAR / Csquare, Inc., 2026-06-01 - [15]T2 · Company / regulator
Bank of England — Financial Stability Report, July 2026 (FPC) — Bank of England, 2026-07-01 - [16]T2 · Company / regulator
CoreWeave press release — closing of the $3.1bn DDTL 5.0 Facility (Business Wire text, via Morningstar syndication) — CoreWeave, Inc. / Business Wire (syndicated by Morningstar), 2026-05-18 - [17]T2 · Company / regulator
KBRA — Proposed Data Center ABS Global Rating Methodology — Kroll Bond Rating Agency (KBRA) - [18]T2 · Company / regulator
Crusoe / Blue Owl Capital / Primary Digital Infrastructure — second phase of $15bn joint venture, Abilene, Texas — Crusoe (company newsroom), 2025-05-21 - [19]T3 · Press / analyst
Charles River Associates — 'Data center ABS: Risks, yields, and ratings' (CRA Insights, December 2025) — Charles River Associates (CRA International), 2025-12-01 - [20]T3 · Press / analyst
Forbes — 'How A $27 Billion Bond Deal Quietly Funds Big Tech's AI Build-Out' — Forbes, 2026-06-17 - [21]T3 · Press / analyst
Fortune — 'Meta's $27 billion bet turns AI compute into Wall Street's hottest new investment' — Fortune, 2025-10-31 - [22]T4 · Aggregator
Cayman Finance — 'Structured lending for hyperscale data centre providers: Offshore SPVs powering securitisation-driven capital solutions' — Cayman Finance (Cayman Islands financial-services industry association), 2025-11-03 - [23]T2 · Company / regulator
Moody's Ratings — rating action assigning A3 to CoreWeave Compute Acquisition Co. VIII (DDTL 4.0) — Moody's Ratings - [24]T2 · Company / regulator
S&P Global Ratings — 'ABS Frontiers: Equipping Data Centers Through Securitization' — S&P Global Ratings
Methodology
Every figure in this report traces to a document we read. Ratings, prices, maturities and vehicle names come from CoreWeave's Form 10-K and its DDTL 4.0 and DDTL 5.0 8-Ks; the Hyperion lease, the residual value guarantee and Meta's server lives come from Meta's Form 10-K; the Beignet bond's coupon, maturity, ISIN and US domicile come from two independent registered-fund NPORT-P holdings files rather than from press coverage of the deal.
What is ours, not theirs. Four numbers in this report are arithmetic, not quotation, and we say so where they appear: the 225-basis-point spread gap (SOFR + 4.50% less SOFR + 2.25%, both margins stated verbatim in filings); the roughly 23-24 year span of the Beignet bond (stated maturity 2049-05-30 less a 2025-26 holding date); DDTL 4.0's approximately six-year tenor (March 2032 less a 31 March 2026 close); and the observation that $4,338m of $21,615m — roughly a fifth — falls due after 2030. The interactive model's advance rate is the reader's assumption, not a disclosed figure; its net book value line is straight-line accounting depreciation, which is not a market bid.
What we could not read. Moody's rating action for the DDTL 4.0 borrower returned HTTP 403 Forbidden to both an automated fetch and a browser user-agent, and S&P's "ABS Frontiers: Equipping Data Centers Through Securitization" returned a security block page. No figure here is attributed to a rating agency's own words: the A3 / A (low) ratings are sourced to CoreWeave's own 8-K exhibit, and Beignet's A+ to Forbes and Fortune. KBRA's free methodology page discloses no useful-life assumptions or numerical thresholds. If the agency documents show explicit GPU residual-value haircuts, this report's central contrast is wrong, and we would want to know.
What we refused to use. The Cayman Finance article that appears to be the origin of the offshore claim is promotional copy from the Cayman Islands' own industry association, names no GPU transaction, and is scoped to Asia. It is cited for no number. NPORT-P filing counts ("Beignet Investor" in 2,245 filings) are per-fund and per-quarter and are reported as an indication of breadth only, never as a count of distinct holders.