Capital Structure

This rung asks who is paying for the AI buildout and how that money is actually arranged: the loans behind the neoclouds, the multi-year compute contracts the model labs sign, the backlog figures companies report, and the depreciation schedules that decide how fast a chip is written off. It sits between two neighbours. What gets built belongs to datacenters, and what it costs to run belongs to inference-economics. This rung is the part in between: what the money costs, and who is on the hook if the demand does not show up.

What this rung prices

Every token is served on a chip somebody borrowed money to buy. The interest on that loan, and the number of years the chip is written off over, are both baked into the hourly price of the chip, and so into the token price and the task price above it. Three of the rung's four primary sources are about that first cost, and the cleanest thing they show is how much it moved in about four and a half months (31 March to 10 August 2026). Across three CoreWeave loan facilities backed by the same kind of collateral, the interest margin over the benchmark rate went from SOFR + 2.25% to + 4.50% to + 5.50%, while the Moody's rating fell from A3 to Ba2 and the facility size shrank from $8.5bn to $3.1bn to $2.6bn (DDTL 4.0; DDTL 5.0; DDTL 5.5). The cheapest of the three, DDTL 4.0, also had a fixed-rate slice at about 5.9%, the only all-in interest rate any source here gives. None of these sources turns that margin into a cost per GPU-hour or per token. That conversion is the thing the rung does not yet hold.

The load-bearing findings

• Debt backed by GPUs got about 3.25 points more expensive in about four and a half months, and lenders kept lending. The spread curve above only exists as a comparison: no single release contains it. Corrected 2026-09-24: this page and the DDTL 5.5 close read had dated DDTL 4.0 to 2025 and called the span "roughly a year". The DDTL 4.0 release, now read, closed on 2026-03-31. The honest complication is that each facility was "meaningfully oversubscribed", and DDTL 5.0 priced 0.50 points tighter than where it launched. So the curve does not say the market is closing on GPU debt. It says the market repriced this debt to a high-yield level and then bought it eagerly at that level (DDTL 5.0). The three facilities also differ in structure, tenor and customer mix, so the direction is solid and the exact gaps are not a clean series (DDTL 5.5).
• The first step on the curve is mostly about who rents the chips. DDTL 4.0 finances a contract with a "leading AI enterprise" and is rated investment grade; DDTL 5.0 finances two customers its release calls below investment grade. A loan like this is rated largely on the customer paying for the chips, so most of the jump from + 2.25% to + 4.50% is probably a change of customer credit rather than the market turning. Neither release puts a number on the split (DDTL 4.0; DDTL 5.0).
• The rung's first special-purpose vehicle. DDTL 4.0 is non-recourse: the lenders' claim is on one ring-fenced subsidiary, CoreWeave Compute Acquisition Co. VIII, LLC, and its assets, not on CoreWeave as a whole (DDTL 4.0).
• The newest loan outlasts the contracts it finances. DDTL 5.5 runs about five years against customer contracts that average about three. CoreWeave's own words: "Lenders are now comfortable financing shorter-dated contracts." The facility allows for renewing those contracts or leasing the capacity to someone else when they end, which means part of the risk that a used GPU fleet is worth less than planned has moved onto the lender (DDTL 5.5).
• The chips are the collateral, so the write-off schedule is part of the loan. DDTL 5.0 is secured on the GPU and HPC assets, with the draw schedule matched to both deployment and the assets' useful life. That makes the depreciation assumption a direct input to the credit, and it is disclosed that both named customers behind that facility are below investment grade (DDTL 5.0).
• On the other side of the ledger, a frontier lab is committing gigawatt-scale money to one vendor's custom chip while keeping two others. Anthropic committed to up to one million Google TPUs and well over a gigawatt of capacity online in 2026, a deal worth tens of billions of dollars, while still running Amazon Trainium and NVIDIA GPUs. "Up to" is a ceiling, not a floor (Anthropic–Google TPU deal). A later expansion of up to 5 GW from 2027 is carried in the same file at moderate evidence, from a separate announcement that was not fetched directly.
• Backlog numbers are large, and they can count the same demand twice. The contracted backlog page holds Oracle's $523B of remaining performance obligations, CoreWeave's $99.4B backlog, and a named $21B Meta commitment sitting inside CoreWeave's backlog while Meta is itself tracked as a hyperscaler. That is one sourced case of the same end-demand booking at two layers; how much overlap there is across the whole stack is not quantified by any source.
• Whether early buyers keep a cost edge is modelled, not measured. The depreciation conveyor page carries a single-author preprint's model in which fully written-off fleets keep an incumbent ahead of a new entrant (3.2x in 2026, 1.9x in 2027, 3–4x by 2029–30). Every figure there is a model output, and the page records a competing source that attributes almost none of the fall in token prices to hardware at all.

Concept map

Concepts

Concept	Sources	Evidence	Last Updated
Contracted Backlog & RPO	4 (0 papers)	Weak to moderate	2026-07-23
The Depreciation Conveyor & Vintage Economics	2 (2 preprints)	Weak to moderate (modelled)	2026-07-23

Entities

Entity	Type	Sources
CoreWeave	Company	5
Oracle	Company	3

The four pages above were carried onto this rung from datacenters and inference-economics, and cite sources held in those topics (their links were repointed to those rungs on 2026-09-24). The CoreWeave page now carries all three DDTL facilities as a table; the two concept pages and the Oracle page do not yet fold in this rung's own sources.

What we do not know yet

• What any of this costs per GPU-hour or per token. The spread curve is a price of money, not a price of compute. No source here converts it.
• The terms that carry the real risk. All three primary sources are press releases, not credit agreements: no advance rates, covenants, amortisation, draw conditions, collateral haircuts or residual-value assumptions (DDTL 5.5).
• How much of the widening is customer credit and how much is the market. One borrower, three customers, no split disclosed.
• Who the customers are. DDTL 5.0's two customers and DDTL 4.0's "leading AI enterprise" are unnamed, so concentration cannot be assessed (DDTL 5.0; DDTL 4.0).
• What useful life the GPUs are written off over. The collateral is tied to it, and no source on this rung discloses it.
• The rest of the scope. Vendor financing and circular or related-party deals are named in scope and have no source. SPVs now have one example (DDTL 4.0's ring-fenced subsidiary) and no analysis.
• Standing: read it. Every figure here is read from company releases or other rungs' sources. The CoreWeave entity page records that we hold a position; that is exposure, not having run anything.

Timeline

See the timeline for chronological developments, and the open questions for the open questions.

Sources

#	Title	Type	Date	Status
1	Expanding Our Use of Google Cloud TPUs and Services	technical-report	2025-10-23	compiled
2	CoreWeave Closes Landmark $8.5 Billion Financing Facility (DDTL 4.0)	technical-report	2026-03-31	compiled
3	CoreWeave Closes $3.1 Billion Loan Facility (DDTL 5.0)	technical-report	2026-05-18	compiled
4	CoreWeave Closes $2.6 Billion Loan Facility (DDTL 5.5)	technical-report	2026-08-10	compiled

Read next

• CoreWeave Closes Landmark $8.5 Billion Financing Facility, Achieving First Investment-Grade Rated GPU-backed Financing (DDTL 4.0) — the investment-grade first point, and why the first step is about the customer.
• CoreWeave Closes $2.6 Billion Loan Facility (DDTL 5.5) — the spread curve in one table, and the five-year loan against three-year contracts.
• CoreWeave Closes $3.1 Billion Loan Facility (DDTL 5.0) — why the widening spread is not a closing market.
• Expanding Our Use of Google Cloud TPUs and Services — the demand side: a lab signing for a gigawatt of one vendor's chip.