From 15% to 5.9%: the three years in which Wall Street decided a GPU is real estate

Companion to the shadow-debt map. Figures from company filings, lender disclosures and cited reporting; rate history from CoreWeave's disclosed facilities. Research, not investment advice.

Strip away the LLCs, the residual-value guarantees, the credit derivatives and the $1.09 trillion of uncommenced leases, and the entire AI credit machine rests on one question: what is a used GPU — and the building around it — worth in year five? If the answer is "a lot," the structures in our previous piece are conservative infrastructure finance. If the answer is "little," they are the mechanism by which that discovery gets distributed to insurance and pension portfolios.

The remarkable thing is how fast the market changed its answer.

The repricing: 15% → 5.9% in three years

In August 2023, CoreWeave — then the only company borrowing against GPUs at scale — paid roughly 15% for a $2.3 billion facility from Magnetar and Blackstone, chips as collateral. By March 2026 it borrowed $8.5 billion for seven years at roughly 5.9% fixed. Nine points of compression in thirty months is not a rate move; it is an asset class being born. Lambda's $500M GPU securitization penciled a 50% residual value at year three; secondary-market H100s trade around 45% of new at year three — the collateral assumption has, so far, held. Meanwhile H100 rental rates rose from about $1.70 to $3.35 per card-hour between late 2025 and early 2026, and the A100 — a 2020 chip — is still earning commercial rents six years in.

That is the empirical basis on which NVIDIA ($NVDA) made it official: an August 2026 financing platform with Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs and KKR to mobilize $500 billion+ of third-party capital, with NVIDIA offering residual-value support of up to 25% on select deals. "In AI, compute is revenue," in Jensen Huang's words — AI factories financed like toll roads, GPUs underwritten like buildings. The vendor guaranteeing the resale value of its own product is either the final proof of confidence or the classic late-cycle tell; it was also exactly Meta's move with its data centers, and the market accepted both at investment grade.

What the risk actually is

The bear case is not complexity. It is one sentence: the rents stop. Every layer of the machine — Beignet's bonds, Microsoft's leases, Google's backstops, the Valor GPU fund — is serviced by a hyperscaler rent payment, and every rent payment is justified by end demand for tokens. If AI usage stalls: rents pause → SPV bonds miss → residual guarantees get called (Meta's $28B + $13B, Alphabet's $43.8B notional, NVIDIA's 25% supports, Oracle's lessor guarantees) → private-credit marks fall → and the holders of last resort are annuity books and pension funds, some of which are already gating redemptions.

Two accounting choices amplify the tail. Depreciation schedules have stretched — GPU fleets from ~3 years to 6–7, Microsoft's data centers from 15 to 25 — which flatters current earnings and, per short-seller math, defers on the order of $100 billion of cost recognition over three years. And the circularity is real: NVIDIA invests in the funds that buy its chips; PIMCO bought Blue Owl's own bond while co-anchoring Blue Owl's deals; the same six names sit on every side of every table. The OFR and FSB have both flagged the interconnection.

What the market is saying — and why it isn't 2008

Here is the part the "AI subprime" headline misses: the repricing is already happening, in public, deal by deal. Oracle — the one that kept the debt on its own books — wears a BBB– rating and CDS at 18-year wides. Meta's second SPV bond cleared 62.5bp wider than its first. Amazon's order books thinned from 5.3x to 1.6x. An August CoreWeave facility needed covenants and an extra 100–125bp. Credit investors are not asleep; they are charging more every quarter, which is precisely what a functioning market does to a maturing story.

The systemic comparison also fails on two structural points. Scale: data-center debt, for all its growth, remains a sliver of the investment-grade market and nothing like housing's share of 2007 GDP. Seating: in 2008 the risk sat on the levered balance sheets of banks; today it sits mostly with unlevered long-duration holders — insurers and pensions — who can absorb marks without fire sales, however unpleasant that is for beneficiaries. The banks, more tightly regulated and newly re-armed by the Basel endgame retreat, are only now wading back in. A demand stall would be a painful repricing for retirement capital, not a payments-system seizure. That distinction matters — and it is cold comfort if the retirement capital is yours.

The six dials

This is a cash-flow machine, so watch the cash flows, not the org charts:

1. Token demand growth — Azure +43% with a $678B backlog is the single number currently holding the whole stack up.
2. GPU spot rents — the H100 $/hour series; the first sustained decline is the earliest warning anywhere in the chain.
3. Depreciation schedules — every further stretch is borrowed earnings.
4. CDS on Oracle and Meta — the market's live vote on the two most extended names.
5. Private-credit redemption gates — Blue Owl's 19–41% request rates; spreading gates mean the exit doors are shrinking.
6. RVG and backstop notionals in the filings — Meta's $41B, Alphabet's $43.8B and Oracle's guarantees, quarter over quarter. Growth there is leverage growth, whatever the debt line says.

Financing innovation is how buildouts of this scale have always happened; the railroads, the grid and the mortgage market all invented their instruments mid-boom, and two of those three stories ended with the instruments outliving the crash that tested them. The GPU credit machine has not had its test yet. Until it does, the honest position is the one the filings support: the structures are real, the disclosures are there for those who read them, the market is repricing risk in real time — and the collateral is a bet on demand that has, so far, only gone up.

Research, not investment advice. Not a price call.