Capital Structure — Research Frontier

What's genuinely new and where the rung is heading. First compiled 2026-09-24 from three primary sources and four pages carried over from neighbouring rungs; extended the same day with the DDTL 4.0 release, which completes the borrowing-cost curve.

The binding constraint, named

The price of money for GPU-backed debt, and the useful life that debt assumes. CoreWeave's facilities make both explicit: the loans are secured on the chips, and the draw schedule is matched to the chips' useful life (DDTL 5.0). If the write-off period is too long, or the rate too low, the hourly price of the chip is understated, and so is every token served on it.

The number this rung owns

SOFR + 2.25% → + 4.50% → + 5.50% across CoreWeave DDTL 4.0 (2026-03-31), 5.0 (2026-05-18) and 5.5 (2026-08-10), with the Moody's rating falling from A3 to Ba2. Latest print 2026-08-10 (DDTL 4.0; DDTL 5.5). All three points are now held from their own releases. It is a comparison across differently structured facilities for different customers, so the direction is solid and the gaps are approximate.

Active Frontiers

1. The repricing of GPU-backed debt

Status: Steady Key sources: DDTL 4.0, DDTL 5.0, DDTL 5.5 Key players: CoreWeave

Spreads widened by about 3.25 points in about four and a half months while every facility was oversubscribed and one priced tighter than launch. The first step (+ 2.25% to + 4.50%) coincides with a move from an investment-grade customer to two below it, so much of the widening is probably the customer, not the market. The debt was repriced to high-yield levels and still cleared strongly there.

Open problems:

• How much of the widening is customer credit and how much is timing? No release splits it.
• What spread does a borrower other than CoreWeave pay? One borrower is not a market.
• How does the spread convert into a cost per GPU-hour?

2. Lenders taking the residual-value risk

Status: Early stage Key sources: DDTL 5.5 Key players: CoreWeave

A five-year loan against three-year contracts means the lender is betting the chips can be re-leased when the first contracts end. That is the first sourced sign of residual-value risk moving off the operator.

Open problems:

• What residual value do lenders assume for a GPU after three years? Not disclosed.
• What happens to the spread if re-leasing rates fall?

3. Backlog as proof of demand, and its double count

Status: Stalled Key sources: CoreWeave Q1 2026, Futurum Key players: CoreWeave, Oracle

Backlog is the best public answer to "who pays the rent", and one edge (Meta's $21B inside CoreWeave's $99.4B) shows the same demand booking at two layers. See Contracted Backlog & RPO.

Open problems:

• Duration, cancellability and customer concentration of every backlog held here.
• How much of the aggregate is counted more than once.

4. Vintage: which year's chips you own

Status: Early stage Key sources: Matsuoka preprint

The depreciation conveyor argues the cost gap between early and late buyers never closes. It is modelled by one author, and a competing source attributes almost none of the fall in token prices to hardware.

Open problems:

• Does it survive a check against disclosed useful-life assumptions in hyperscaler filings?

What this rung does not yet hold

• No credit agreement. Three press releases; the covenants and haircuts are where the risk sits, and none is disclosed.
• Almost nothing on SPVs, vendor financing or circular deals, all named in scope. DDTL 4.0's ring-fenced, non-recourse subsidiary is the one SPV example held.
• No disclosed GPU useful life for any company on the rung.

Standing

Read it. Nothing here has been run or modelled in-house. A position in CoreWeave is recorded on the entity page; that is exposure to the rung, not standing on it.