Same Facility, Different Marks: Compute Credit in Lender Filings

01 The Observed Marks: Compute-Backed Facilities

Five compute-collateralized facilities in Q1-2026 schedules of investments are held by two or more unaffiliated advisers. The widest is the CoreWeave Compute Acquisition Co. IV first-lien delayed-draw facility (SOFR + 6.00%): Carlyle funds file 99.0, MSD Investment Corp. 100.0, Great Elm 102.0, and PIMCO Capital Solutions BDC 103.1 — a 4.06-point spread on identical first-lien paper at the same date, with two of the four holders' pieces identical to the dollar of principal. The second-widest pair is Crusoe Energy Systems' 13% fixed tranche: Monroe Capital, which reduced its position roughly 83% during the quarter, files 98.4; LAGO Evergreen, which held, files 102.5 — equal to its amortized cost, in both quarters it has reported.

Every mark shown is Level 3 (significant unobservable inputs) as flagged in its own filing, on paper that is current and performing in every holder's schedule. The marks are each fund's own, determined by its valuation designee under its own policies.

02 The Wider Market, Relative to Compute

Is four points a lot? Sixteen cross-held tranches from the same filing season, matched the same way (same borrower, same tranche economics, same date; one mark per unaffiliated adviser; term-loan core only), give the baseline. The dispersion statistic is the mean pairwise absolute difference (pwMAD) between advisers' marks — a measure that does not grow mechanically with the number of holders, so an 11-adviser facility and a 2-adviser facility compare fairly.

The pattern is regime, not sector. Every performing conventional or hard-asset tranche — syndicated fiber, medical-equipment rental, aviation MRO, and unitranche facilities with as many as eleven unaffiliated holders — lands between 0.14 and 0.94 points, median 0.26. Dispersion past two points appears in exactly two places: credits that have been through a restructuring or bankruptcy (PetVet 2.8; Hornblower 3.9; 48Forty 4.2, where six advisers mark fresh post-restructuring paper at par and one marks it 85), and performing compute paper (CoreWeave DDTL 2.0 at 2.4; Crusoe's 13% tranche at 4.2). Compute is also the only cohort in the sample whose disagreement straddles par: holders differ on how much more than face the paper is worth, not how much less.

03 The Anatomy of a Mark

Why can honest valuations of the same loan sit four points apart? A fair-value measurement on a performing private loan is, under the accounting standard, an exit price — and in practice it is usually produced by an income approach: discount the loan's remaining contractual cash flows at the yield a market participant would require today, over the loan's expected remaining life, calibrated to the price the fund actually paid at entry. Observable transactions in the same instrument, where they exist, take precedence over any model. Three inputs, then: a market yield, an expected life, and a calibration basis. Each is a place two valuation processes can diverge while both remain defensible.

The market yield needs an anchor, and anchors come in grades. The strongest is a trade print: corporate bonds — including this sector's 144A project bonds — disseminate actual transactions, and a holder marks at the print. One grade down is a dealer quote: broadly syndicated loans never touch an exchange, but dealers make markets and pricing services publish composites daily, and the controls show what that does — Zayo's five unaffiliated holders cluster inside 35 basis points, effectively on two values, and one files the position at Level 1/2 rather than Level 3. A grade below that is comparable density: no quote, but enough similar deals priced each quarter that the clearing spread is effectively known — Integrity Marketing's eleven unaffiliated holders agree inside three-quarters of a point with no quote in sight. Compute credit's loan channel has none of the three. Its comparable set is about a dozen disclosed signings over three years, across which the disclosed spread fell by more than seven points. A yield input drawn from that record is judgment, and judgments differ — by a point or more of yield, which on multi-year paper is points of price. Dispersion in this sample tracks the anchor grade, not the accounting level.

And the anchor is set by the financing channel, not the borrower. CoreWeave's convertible notes trade in a dealer market; the sector's 144A project bonds print on TRACE. The GPU-collateralized loan channel — the same credits, the same collateral economics — has no print, no dealer market, and no pricing-service composite. It is the channel where four holders sit four points apart.

The expected life is the biggest single lever. A loan paying an above-market coupon is worth more than par only for as long as the coupon survives, and prepayable paper caps the upside at par the day the borrower refinances. The CoreWeave facility pays roughly three points a year over where new compute paper prints; at that rate, one year of assumed remaining life is approximately the entire gap between the lowest and highest filed marks. And expected life is genuinely uncertain here: documented compute facilities have resolved at a median of seven months from signing, with one issuer refinancing three consecutive facilities inside sixteen months. Whether a coupon survives to 2029 or to next spring is exactly the kind of assumption a filing does not disclose and a model must make.

The calibration basis is visible in the arithmetic. Valuation practice permits carrying a position at its calibrated entry price absent evidence of change, and requires observable transactions to be weighed when they occur. Both patterns appear in this sample's filed numbers. On the CoreWeave facility, Carlyle's fair value sits within a tenth of a point of its amortized cost, while Great Elm and PIMCO carry fair value about three points above theirs; on Crusoe, LAGO's fair value equals its amortized cost to the dollar in both quarters it has filed, while Monroe — the one holder that transacted during the period — files the only mark below par in the group. The filings do not state any fund's method, and this note does not infer one; they do state these arithmetic facts. What the coexistence of both patterns on identical paper establishes is the absence of a shared observable input that would force them together: in the quoted controls, the anchor does that work; here there is nothing to anchor to.

None of this requires a view on the collateral. Every mark in the compute sample sits between 98 and 104, every position is performing, and — as the ledger's collateral-treatment record shows — every disclosed GPU schedule amortizes its collateral toward zero, so a performing mark never has to price the hardware at all. The disagreement is about the price of money against this collateral class: the right spread, and how long the paper lives. Those are precisely the two quantities this market does not yet publish.

The panel's own structure points at the second quantity. Within the compute sample — small as it is — dispersion orders with remaining modeled life: the equipment loan weeks from maturity shows two-thirds of a point of disagreement, the older amortizing delayed-draw about one point, and the long-window facility four. Where nothing is left to model, holders agree; where an above-market coupon has years of assumed life to run, they do not.

Holders also file different maturity dates for the same facilities (on the CoreWeave facility: 5/16/2029, 5/22/2029, and 5/16/2030) — a reporting-convention artifact of delayed-draw structures, whose draws carry their own tenors; no primary document reconciles the dates, and they do not explain the mark ordering.

04 What Narrows the Range

Mark dispersion is not a permanent feature of a credit market; it is a feature of a market's age. Each of the missing inputs above has a known delivery mechanism, and each is visible in filings when it arrives.

The control panel shows two distinct roads to agreement, and they have opposite effects on this record's visibility.

The quoted road: anchors arrive with transactions. A trade in the instrument outranks any model, and one is already visible in this sample: the quarter's only seller files the group's only below-par mark. Syndication and securitization generalize that effect — Zayo's holders, with a dealer quote to anchor to, land within 14 basis points — and this road ends the public record: the paper leaves the schedules. Data-center lending, compute's nearest predecessor, took it; its cross-held Level 3 loans no longer exist (the decade of record behind that sentence is traced in section 05). GPU-backed securitization has begun; each deal that prices converts a modeled input into an observed one, and removes that paper from this panel.

The density road: inputs thicken until models converge. The controls prove a market needs no quote to agree. Eleven unaffiliated holders of one insurance-brokerage unitranche file within three-quarters of a point — paper that will never trade or be quoted — because enough similar facilities price every quarter that the clearing spread is effectively known. Compute's bilateral channel converges this way if its record thickens: every new signing adds a spread comparable — the 2026 cluster at +1.75 to +2.25 is already a tighter calibration set than the sparse 2023–2024 prints — and every documented resolution adds an expected-life observation, narrowing the assumption that currently accounts for the largest share of the gap. Unlike the quoted road, this one keeps the paper in the filings: convergence happens on the record.

Both roads narrow the range; they differ in whether anyone gets to watch. And because each mechanism surfaces in public filings, the dispersion itself is measurable quarter by quarter: the same facilities will be marked again next quarter, by the same holders, on a growing record. Whether the range narrows is not a matter of argument — it will be on the record, one filing season at a time.

05 The Data-Center Precedent

Compute credit is not the first collateral class to pass through this stage, and the previous one's record is complete. A decade ago data-center credit sat exactly here: bilateral loans on fund books, marked quarterly by judgment. Solar Capital carried a bilateral ViaWest note (13.5%, part paid-in-kind) from the end of 2010; by 2017, TierPoint's second-lien term loan (L+7.25, due 2025) was cross-held by unaffiliated complexes including FS/KKR, Benefit Street, and BlackRock TCP, whose same-tranche marks ran as much as 5.2 points apart — at year-end 2019, two complexes filed 88.8 and 94.0 on identical paper. The same borrower's broadly syndicated first lien, with a daily quote, was marked identically to the basis point by its holders. The dispersion belonged to the channel, not the borrower — the same split the CoreWeave structure shows today between its quoted converts and its unquoted loan channel.

Every pattern in the compute panel above appears in that record. One holder's mark on the TierPoint second lien sat unchanged at 99.01 for four consecutive quarters while another's moved with the market — anchoring policy and live modeling coexisting on identical paper. And a judgment mark could sit far from the eventual outcome with no default intervening: FS/KKR carried Flexential's second lien at 74 in 2019; the loan was refinanced at par two years later, weeks ahead of the company's $2.1 billion inaugural securitization.

The exit is documented to the week. Data-center securitization began in February 2018 with Vantage's inaugural ABS; DataBank and Flexential followed in 2021. TierPoint's last filed fund marks — 99.63 — are dated June 30, 2023; its $1.06 billion securitization closed July 10, 2023, and the loan left the schedules. No market price ever arrived to test the marks — the only evidence that could have — and repayment at par settles nothing: a loan can be worth 74 to any buyer in 2019 and still pay off at par in 2021. The refinancing did not answer the question of what the second lien had been worth; it retired the instrument that posed it. This quarter's filings contain no cross-held data-center loan pairs at all.

Stated as a working view rather than a fact: we anticipate compute credit splits along the two roads of section 04. The contracted, long-dated core migrates toward syndicated and securitized channels — that exit has begun — while the short, amortizing bridge vintages never pool at all: they refinance away, as they are designed to. What remains bilateral converges in place if signings densify, the way the control panel's unquoted unitranches did — and that convergence, unlike the securitized exit, happens on the record. GPU-backed securitization has begun. On the data-center clock, the compute panel above sits at roughly 2015–2017: cross-held, judgment-marked, dispersed — and, for now, completely public. The record will be maintained here for as long as the window stays open.

06 Method & Limitations

Marks are reproduced from schedules of investments in 10-Q filings for the period ended 3/31/2026 (one 10-K), read July 2026; a mark is fair value divided by principal, both as filed. Tranches match on borrower plus tranche economics (rate and maturity month); advisers are counted once per complex, with affiliated vehicles collapsed. Excluded: revolver and delayed-draw rows whose filed values net unfunded-commitment fees (not comparable as filed); positions held through joint-venture vehicles where flagged; one tower-securitization pair whose coupons indicate different classes of the same program; rows filed at millions precision where rounding exceeds the measured dispersion. The data-center precedent (section 05) applies the same matching rules to schedules of investments filed 2014–2023 — 42 cross-complex tranche-quarters across TierPoint, Flexential, ViaWest, and peers; stressed names of that era (Sungard, Internap, Cyxtera) are flagged separately and excluded from the performing comparison, and securitization dates are from issuer announcements.

Limits. Compute tranches have two to four unaffiliated holders each — the finding is the table, not a statistic, and one quarter's filings are one observation of each mark. Coverage is funds that file schedules of investments; banks and private funds, which hold most compute credit, publish no marks. One holder files identical marks on both of its CoreWeave tranches, consistent with position-level rather than tranche-level marking. Marks are each filer's own fair-value determinations; nothing here estimates, adjusts, or opines on any of them.

Underlying rows, with links to every source filing, are maintained on the credit reference series page; the instrument record is the compute credit tracker.