Vault
research

Compute as Collateral — The Residual-Value Wrap and What It Does to the Depreciation Question

Created

Compute as Collateral — The Residual-Value Wrap and What It Does to the Depreciation Question

Builds-on: the-shadow-balance-sheet-nikkei-1-65t-and-the-spv-layer, ai-circular-financing-and-banking-exposure-audit, cyclical-20-and-the-ai-capex-mask Related: anthropic-subsidy-stress-test, the-data-center-convergence, ai-infrastructure-endgame-indicators, why-the-market-refuses-to-crash, mechanism-vs-narrative-method Led-to: ai-capex-watchlist-check-august-20-2026

In the second week of August 2026, three things happened that are usually described separately and are actually one event.

Nvidia signed MOUs with Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs and KKR to stand up financing platforms targeting more than $500B of third-party capital for AI compute, and disclosed that it may provide residual-value support for up to 25% of an individual financing. Bloomberg reported that roughly $70B of such guarantees now sit off the balance sheets of AI chip vendors, and that S&P has begun adding Broadcom's version to adjusted debt. And CME Group announced that on October 5 it will list H100 and B200 Rental Index futures — the first regulated contracts on the price of compute.

Read together: the industry just built a price, a wrap, and an exchange for GPU depreciation, in that order, inside about ninety days. This doc is about what that does to the cyclical-20-and-the-ai-capex-mask thesis, which turned on depreciation eventually catching up to a masking line item. The answer is that depreciation stops being an accounting argument and becomes a quoted number — and that the entity now standing between the quoted number and the lenders is the same entity whose revenue depends on the quote staying high.

What the vault already had

the-shadow-balance-sheet-nikkei-1-65t-and-the-spv-layer got to the residual-value guarantee first, from the tenant side. Its August 1 update caught the Nikkei disclosure that Meta has guaranteed to cover investors' losses on the Hyperion JV if the facility becomes unnecessary, and drew the right conclusion — the risk is retained, only the accounting is transferred — along with the Beignet structure in more detail than is repeated here ($27B of A+ 144A bonds maturing 2049 against GPUs that depreciate in two to six years, anchored by PIMCO at $18B and BlackRock at $3B, at ~225bp over Treasuries).

Two things are new since. The guarantee has crossed from tenant to vendor, at a scale that makes it market structure rather than a one-off. And compute is about to get a listed forward curve, which prices the guarantee for the first time. Those are the subjects here.

First, what was actually announced

The popular framing (Nicholas Crown's "Nvidia is America's second central bank" is a representative version) gets the direction right and the mechanics wrong in three specific ways worth correcting before building on it.

"Six institutions are raising $500B for Nvidia." No. The $500B is an aggregate target for third-party capital the platforms are designed to mobilize over time, and it funds Nvidia's customers — neoclouds and labs buying compute — not Nvidia. It is not committed capital, not a fund, and not Nvidia revenue. The agreements are memoranda of understanding subject to final documentation. The direction of the money is the interesting part: it flows from insurance and retirement capital through asset-manager-structured debt into GPUs. Goldman puts AI financing at roughly a quarter of all gross US investment-grade issuance this year against ~$600B of AI investment. That is the actual news.

"Nvidia guarantees 25% of residual value, so $125B of exposure." Two errors compounded. The 25% is per-opportunity and case-by-case, and it covers a share of the shortfall between liquidation proceeds and the loan's assumed basis — not 25% of the asset. Multiplying a per-deal shortfall cap by an aggregate platform target assumes full deployment and a total wipe of residual value simultaneously. The number that is actually sized and disclosed today is $70B across Nvidia and Broadcom combined, and Broadcom's own filing puts maximum theoretical loss on its first transaction ($35B, over 1GW for Anthropic, via the AI XPV platform with Apollo and Blackstone) at $29B under complete default with zero recovery.

"This might be the safest loan on Wall Street." Partly tested, and the answer splits by writer. Bank of America downgraded Broadcom's issuer and bond ratings specifically over the AI XPV guarantee; the stock broke below $400; S&P called Broadcom's residual value support a "contingent debt-like obligation" and is adding it to adjusted debt; Moody's flagged it as an overhang.

(Corrected 2026-08-20: I originally generalized that to "the guarantor got repriced because of the guarantee." That is true of Broadcom and false of Nvidia. On August 18, against a newly disclosed $105B guarantee, Moody's affirmed Nvidia at Aa1 with a positive outlook — citing $106B of cash and marketable securities, Moody's-adjusted debt/EBITDA of 0.2x pro forma for a $25B issuance, and $424B of expected cumulative free cash flow over FY27–28, with the guarantees expected to peak in 2031. Balance-sheet capacity is doing exactly what it should, and Nvidia has vastly more of it than Broadcom. The correct claim is narrower: the wrap is priced when the writer's balance sheet is thin enough to notice, and ignored when it isn't — which is a statement about the agencies' current thresholds, not about whether the risk exists. The monoline comparison below survives on the correlation argument, not on a ratings argument.)

The mechanism, precisely

There are two distinct Nvidia backstops in circulation and the discourse keeps merging them.

The offtake backstop. Per SemiAnalysis, Nvidia writes a take-or-pay minimum revenue guarantee on the underlying GPU capacity, typically six years, at pre-agreed annual price levels. Their illustrative case averages ~$2.36/hr; announced deals imply similar (SharonAI at $4.88B total value, ~$2.33/hr; Firmus at $25–30B expected customer revenue over six years). In exchange Nvidia takes a negotiated cut of revenue above the backstop — 40% in their example, working out to an ~18% effective take rate in the short-term rental scenario. Nvidia is selling a floor and buying a call.

The residual-value guarantee. Separately, if the borrower defaults and the lender liquidates the chips below the basis the loan assumed, Nvidia covers up to 25% of that gap.

The load-bearing detail is what these do to underwriting. Lenders are sizing at 70–80% LTV to a 1.3x minimum debt service coverage ratio calculated on the assumption that the backstop is fully activated. That is not a cushion. That means the credit being underwritten is Nvidia's, not the neocloud's — the operating business is structurally irrelevant to the debt decision, and the equity holder knows it: SemiAnalysis models the neocloud's IRR at 25.4% in the short-term-rental case and zero or slightly negative if the backstop activates. The lender is fine either way. The lender is lending to Nvidia through a neocloud-shaped hole.

That is the kernel of truth in the central-bank framing. Nvidia is not a lender of last resort to the borrower; it is a lender of last resort to the collateral. It is standing behind the asset so that the asset can be financed.

Why the wrap had to exist: there was no price

Until 2026 GPU rental was bilateral, opaque and unquotable. Lenders had no reference price for the collateral and no way to hedge it, so they wanted a floor from a strong credit before they would lend at all. The wrap is what a market does when it needs a price and doesn't have one: it borrows the credit of the counterparty who does know.

That is the temporary condition, and it is ending. Silicon Data's H100 and B200 rental indices go to CME on October 5 (pending CFTC review), with Architect's perpetuals on the Bermuda-regulated AX exchange running since January. Once a forward curve on GPU rental exists, three things follow mechanically:

  1. The wrap gets priced. A residual-value guarantee is a put on a hardware price. Today it is written at a strike nobody can quote, so it is carried at roughly zero cost. With a listed curve, the option has a mark, the writer's auditor sees it, and the accounting drifts toward fair value.
  2. The depreciation debate resolves into a number. Michael Burry's estimate of ~$176B of understated depreciation across the hyperscalers 2026–2028 (Oracle's earnings overstated 26.9%, Meta's 20.8% by 2028) is currently unfalsifiable in real time — it is an argument about the right useful life, and management holds the pen. A traded forward curve on rental rates is an observable claim about the economic life of the asset. When the six-year book and the two-year forward disagree publicly, that is an impairment conversation with a citation.
  3. Someone can short it. The first genuine two-sided market in AI infrastructure risk. Every prior expression of the bear case had to route through NVDA equity or hyperscaler credit, both of which carry enormous unrelated exposure. This one is clean.

The gap the curve will be measured against is already large. H100 rental went from $8–10/hr in early 2024 to $1.80–3.50/hr by Q2 2026, a 64–75% decline in about two years. One-year contract pricing did firm ~40% off an October 2025 low of $1.70/hr to $2.35/hr by March 2026 — supply tightness, not a reversal of the secular curve. Set that against a six-year take-or-pay struck near $2.36/hr and the shape of the trade is visible: the backstop holds the contract price roughly flat across a window in which the spot price of the prior generation fell by two-thirds.

What it does to the mask

The cyclical-20-and-the-ai-capex-mask thesis held that AI capex is filling the hole in the cyclical 20% the way residential investment did in 2005–07, and that it would persist because it is funded by cash-rich balance sheets that don't need to stop. That second clause is the part that just changed.

Old mask New mask
Funding source Hyperscaler operating cash flow Third-party institutional debt (insurance, pension)
Who bears a stop Equity holders, immediately Creditors, on a lag
Constraint Board appetite Debt service coverage
Failure mode Capex guidance cut Default and liquidation

The trailing-four-quarter numbers show the transition mid-flight. Combined hyperscaler capex 2Q25–1Q26 was $433.9B (Amazon $151.0B, Alphabet $109.9B, Microsoft $97.2B, Meta $75.7B), with 1Q26 alone at $129.8B, up 80% YoY. Depreciation over the same window was ~$149B, about a third of capex. Amazon's capex now runs at 102% of operating cash flow, over the line for the first time since 2022, and all four issued more than $97B of debt in 4Q25 alone. The 5–6 year useful lives (Microsoft 6, Alphabet 6, Meta 5.5, Amazon cut back to 5 with accelerated write-offs on early retirements) mean the $433.9B cohort adds $65–75B of annual D&A once in service — $95–110B if the true economic life is closer to four years, which is what Amazon's reversal is signalling.

So the mask does not break when hyperscalers get bored. It extends, because leverage always extends things, and it now breaks through a credit channel instead of an equity channel. This is a duration up, severity up revision. It makes the cyclical-20-and-the-ai-capex-mask timing call later and the ai-circular-financing-and-banking-exposure-audit tail fatter, which is an uncomfortable pair to hold and is probably correct.

The right analogy is the monoline, not the central bank

A central bank can honor its guarantee in all states of the world because it issues the thing it promises. Nvidia has a product cycle.

The residual-value guarantee is worth most precisely when GPU demand weakens — which is exactly when Nvidia's revenue, cash flow, and ability to fund the guarantee are also under pressure. Financiers call this wrong-way risk, and the last institution to run it at systemic scale was the monoline insurer. MBIA and Ambac wrapped structured credit with their AAA ratings, lenders underwrote to the wrap rather than the collateral, and in 2007–08 the wrap and the collateral deteriorated together because they were driven by the same variable. The wrap did not distribute the risk. It concentrated it and made it invisible until the writer was downgraded.

flowchart TB
  A[Insurance and pension capital] --> B[Financing platform SPV<br/>Apollo, BlackRock, KKR]
  B --> C[Buys GPUs from Nvidia]
  C --> D[Leases to neocloud or lab]
  D --> E[Lease payments service the debt]
  E --> A
  F[Nvidia residual-value guarantee<br/>25% of shortfall] -. wraps .-> B
  G[Nvidia offtake<br/>6yr take-or-pay] -. floors .-> D
  H[GPU rental price falls] --> I[Lease coverage fails]
  I --> J[Collateral liquidated<br/>below loan basis]
  J --> F
  H --> K[Nvidia revenue falls]
  K -. impairs .-> F

The single variable at the bottom left drives both arrows into the guarantee. That is the whole diagram.

Two things are genuinely different from 2008 and should be held honestly. Nvidia today has real cash generation and no leverage, where the monolines were thinly capitalized relative to what they wrapped. And the wrapped exposure is a fraction of the underlying, capped at 25% of shortfall rather than a full principal-and-interest guarantee. This is a worse structure than the market currently prices and a better one than 2008. The distinction that matters is that the monolines' problem was never capital adequacy in the base case either — it was correlation.

Where this leaves the Anthropic thread

Broadcom's AI XPV first transaction is over 1GW for Anthropic, and the platform targets 20GW by 2028. anthropic-subsidy-stress-test mapped the implicit Trainium/TPU subsidy at $2–4B/year and asked what happens to Anthropic's economics when it normalizes. The answer just got more complicated: a meaningful share of Anthropic's compute is now supplied through an off-balance-sheet vehicle whose debt is cheap because Broadcom's credit is wrapped around it. That is a third subsidy layer, distinct from the compute-credit and the below-cost-silicon layers, and it is the one most sensitive to a rating action on the guarantor — which BofA already took. Worth a dedicated pass.

Watchlist

The signals in ai-circular-financing-and-banking-exposure-audit were mostly structural and slow. These are faster, because a listed contract generates data daily.

What would falsify this

The wrap is a bet that inference demand absorbs depreciating silicon faster than newer silicon obsoletes it. If that is true, this is unremarkable vendor finance and the guarantees expire worthless, which is what Nvidia is underwriting. Concretely, the thesis here is wrong if: the CME curve prints in contango past two years at launch; H100-class rental stabilizes above $2/hr through 2027 as Blackwell absorbs frontier training and older silicon takes over inference; and RVG notional grows without spreads on data-center ABS widening. That combination would say the market has looked at a real price for compute and decided six years was roughly right.

The thing to notice either way is that the question has moved. For three years "how fast do GPUs actually depreciate" was a debate between managements and short sellers with no referee. In October there will be a tape.

Sources

Backlinks