Analysis
The collateral ladder
Fourteen months ago AI borrowing was secured on company cash flow. The record since then shows lenders accepting sites, then unstarted leases, then chips, then a private shareholding, then silicon that had not previously secured anything. Each step is dated.
What secures the loan
The interesting question about a debt boom is rarely how much. It is what the lender gets to seize.
That question has a clean answer in the coverage record, because each new kind of collateral arrives as a dated, reported first. Set those firsts in order and they describe a ladder — one that the AI buildout has climbed down, rung by rung, in about fourteen months.
The rungs, in order
- 2025-06-03xAI sells $5B of debt at a double-digit rate. Ordinary corporate credit: the lender's claim is on the company.
- 2025-10-24Banks assemble $38B against Oracle-leased data centres in Texas and Wisconsin. The claim narrows from a company to specific buildings.
- 2026-02-04$56B+ of construction loans marketed against Oracle's future leases — rent it has not begun paying on buildings not yet finished.
- 2026-02-20AMD backstops a $300M Goldman loan so Crusoe can buy AMD chips. First known use of AMD silicon as collateral, and the vendor stands behind it.
- 2026-02-26The GPU-backed model, pioneered at CoreWeave, spreads: chips pledged, obligation housed in a vehicle off the borrower's books.
- 2026-04-04Meta's Ohio "Project Walleye" seeks $3B in a deal reported as first of its kind — lenders funding the building and the power assets together.
- 2026-04-23SoftBank seeks $10B secured by its OpenAI shares. The collateral is a stake in a private company with no public price.
- 2026-07-18General Compute borrows $400M against inference-specific chips, apparently the first deal to pledge that class of silicon.
- 2026-08-11Lambda sells a $917M leveraged loan to buy GPUs under an Nvidia contract.
Nine steps. Every one of them was reported as a sensible piece of financial engineering, and read individually every one of them is.
The direction of travel
Read together they move consistently in one direction: away from things that generate cash and toward things that hold a price.
A claim on a company is a claim on everything it earns. A claim on a building leased to Oracle is a claim on one tenant’s willingness to keep paying. A claim on rent not yet being paid is a claim on a lease surviving until the building opens. A claim on GPUs is a claim on hardware whose earning power falls every time a faster part ships. A claim on SoftBank’s OpenAI shares is a claim on a valuation set by the same funding rounds the borrowing helps finance.
Benjamin Graham’s instruction was to read the instrument rather than the issuer’s story, and the instrument is where these nine deals differ from each other far more than the companies do. Two of the borrowers on this list are investment grade. What their lenders can seize, and what that seizure would be worth on the day it happened, are not remotely the same question.
What the word “first” does not show
Four of these nine rungs were reported as a first: the first AMD chips pledged, the first deal joining a building to its power assets, the first inference-specific silicon accepted, and the GPU-backed structure the others copied. The tempting reading is that a comfortable market does not need new collateral, so counting firsts counts the reach for novelty.
That reading does not survive being tested, and the test is worth showing.
Across the coverage of collateral-secured financing in this sector, no deal was framed as a first before 2026; four distinct ones were, from February 2026 onward. On its own that looks like support. But the practice is older than the framing. In September 2025, Forbes covered CoreWeave’s use of GPUs as collateral — the strategy that had carried it to $11.2 billion of debt — and described it without any suggestion it was novel. Five months later the same practice was reported as a model pioneered by CoreWeave.
Nothing about the deals changed in between. The coverage did. “First of its kind” is a claim made by reporting, and reporting on this sector intensified sharply across exactly these months, so a rising count of firsts is at least as consistent with journalists noticing a pattern as with a market reaching for new collateral. It is also close to circular here, because the nine rungs on this page were selected for being new kinds of collateral — a set chosen for novelty will be rich in novelty.
This shape is not particular to credit. Ray Kurzweil makes the same observation about crime statistics: a violent incident is reported extensively, while reductions in crime are “literally nonincidents” and go uncovered — so public perception tracked the coverage rather than the falling rate. His point about cause is the one worth borrowing. It “need not be the result of anyone’s conscious decision”; the incentives do it. Nobody has to be wrong for a count to end up measuring the counter.
So the adjectives are not the evidence. The evidence is the collateral itself, which sits in the deal terms rather than the framing: what was pledged, and what a lender would hold if it were seized. That is what the ladder above records, and it does not depend on anyone calling anything a first.
One thing the ratings side does show independently. By February 2026, developers were seeking ratings on unbuilt facilities to unlock capital that would not otherwise arrive, and the agencies expanded coverage to meet the demand. That is a change in what buyers required before participating, not a change in how the deal was described.
Where the ladder stops
The record also shows the other side deciding it has gone far enough, and the dates are close together.
In March 2026, KKR, Blackstone and others turned down data centre debt over insufficient insurance against risks including natural disasters — the buyers with the most capital available, declining on a question about the physical asset. Six weeks later, some banks were seeking to sell Oracle-linked loans at a discount.
Those two events sit at the bottom of the ladder, not the top. They concern site debt and construction paper, the rungs where the claim had already narrowed from a company to a building. Over the same months, unsecured bonds from Meta, Nvidia and Alphabet cleared at three to five times covered. The market did not lose its appetite for AI credit. It lost its appetite for a specific rung.
What to watch
The ladder is the thing to track, and each step is dateable. Whether a lender accepts a class of collateral not on this list — power purchase agreements and cooling plant are the obvious candidates, and the Meta Ohio deal already bundled power in. Whether any of the nine rungs reprices, meaning a comparable deal that priced at par earlier this year comes back wider. Whether a syndication that failed in April succeeds on a second attempt. Note that the first of those is a change in deal terms, which is checkable; how anyone describes it is not the signal.
None of that is a forecast. It is a list of things that will show up in the coverage record before they show up in an aggregate spread, which is the argument The Credit Wall makes at greater length.
Sources
Ray Dalio, Principles for Navigating Big Debt Crises — the bubble-phase marker of financing instruments built outside conventional lending. Reading notes.
Benjamin Graham and David Dodd, Security Analysis — read the instrument, not the issuer. Reading notes.
Ray Kurzweil, The Singularity Is Nearer — the crime-statistics case, and the point that a reporting artifact needs no one to decide anything.