The Support Cliff

Larix Research · Framework · In The Cement Curve we argued that GPU rental is energy-bounded commodity production. This note takes the same argument to the balance sheet: the hardware's value is governed not only by resale and rental rates, but by the software-support horizon that determines whether it can still do economically valuable work. No rating, no target. Disclosures at the end.

The appreciation engine is software

On 10 August 2026, NVIDIA announced memoranda of understanding with Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs and KKR to mobilise more than $500 billion of third-party capital for AI infrastructure. The six independent financing platforms were not a fund or a vendor loan. The market read the announcement two ways: as the moment compute acquired a term structure, and as a chip supplier organising its customers' credit. NVIDIA shares fell about 3% on the day.

Each reading leaves out the underwriting question. Every structure being built rests on a single empirical claim: that GPUs hold economic value through a cycle in a way a lender can model. The structuring debate has settled on which analogue to borrow. Rental-fleet ABS uses a manufacturer repurchase agreement to absorb depreciation on "program" vehicles. Equipment-lease ABS assumes that technology hardware already carries an explicit obsolescence premium. NVIDIA's offer of case-by-case residual-value support on up to roughly 25% of an opportunity maps neatly onto the first model.

Both analogues assume a familiar sequence: the asset declines along a curve, residual value is recoverable at auction, and a manufacturer backstop absorbs the tail.

A GPU's economic life is not a smooth curve. It has an end-of-support date, set by the same entity offering residual-value support.

Jensen Huang's framing, that NVIDIA compute improves in economic usefulness over its life, is true, and it is not a claim about silicon. The die does not change after fabrication. Everything that makes an installed H100 more productive in its third year than its first happens in software: better kernels, lower-precision numerics, improved collective communication, higher achieved utilisation on identical metal. FP8 training on Hopper is the canonical case. It was a hardware capability that became economically real only once the framework layer learned to exploit it, years after the transistors shipped.

This is the strongest fact in the bull case for GPUs as collateral, and it is routinely misattributed. Lenders hear a claim about durable hardware. What is actually being described is a software organisation continuously rewriting the economics of an installed base it does not own.

That distinction matters because the mechanism is not symmetric in the way lenders assume. An asset improved by software is an asset whose value is set by a roadmap.

The same mechanism, run backwards

Run the same process in reverse and the software roadmap becomes a source of obsolescence.

Framework and toolkit support for older compute capabilities is finite, published, and periodically withdrawn. New model architectures arrive with kernels written for the newest hardware first, and often only. The problem is larger than slower performance: the workloads with economic value increasingly will not run on older hardware at all without engineering work nobody in the secondary market is willing to fund.

This produces an asset that is physically sound and economically dead on the same day. There is no equivalent in the analogues being borrowed. A five-year-old rental car is worth less than a new one; it is not rendered undriveable by a software release. Boeing cannot deprecate the 737 from the sky by shipping a compiler update. The GPU cascade has a floor that the used-car cascade does not: the point at which the ecosystem stops compiling for the architecture.

Ampere and earlier generations are where this becomes observable first. A residual-value model that treats them as the low end of a continuous distribution, the natural move for anyone trained on fleet data, is modelling the wrong object.

Three curves, conflated

Much of the public depreciation argument is people describing different curves and disagreeing.

The first is resale value: the widely circulated estimate of roughly 73% decline in H100 hardware value over three years. This is a claim about what the metal fetches.

The second is rental rates, which have fallen materially from the 2023 peak but by less than resale, because rental prices clear against utilisation and power, not against replacement cost.

The third, and the one that actually governs collateral quality, is the task cascade. Older silicon migrates from frontier pre-training to fine-tuning to inference to batch and offline work. Fleet-ABS practitioners understand cascades instinctively. It resembles the used-car market, where a vehicle moves from retail to rental to fleet to auction. The mistake is assuming the GPU cascade terminates the same way, in a long, thin tail of declining but non-zero value.

Michael Burry's argument that hyperscalers understate depreciation by using five- and six-year schedules against an annual product cycle is directionally right and mechanically wrong for the same reason. Calendar schedules do not miss the decline because they are too long. They miss it because the trigger is not calendar time.

Tokens per kilowatt-hour

In a power-constrained facility, which is to say nearly every facility now being financed, the true carrying cost of an installed GPU is not its book value or its resale price. It is the output of whatever chip could occupy the same rack and the same power envelope.

An older accelerator retires when its tokens per kilowatt-hour lose to the replacement, or when the software stops supporting it, whichever comes first. Both triggers are step functions. Neither is legible in a depreciation schedule keyed to years, and neither is captured by an auction price for the hardware in isolation, because the buyer at that auction faces the same power constraint.

This makes power the real seniority question in these structures. A facility with contracted power and a long interconnect queue can afford to run older silicon in low-value workloads for years. A facility competing for the same megawatts against a newer deployment cannot. Two identical GPU pools, financed identically, can have materially different residual behaviour because of the grid connection underneath them.

The guarantor also sets the clock

The obvious concentration risk in these structures is a credit exposure, and it has a familiar shape: an entire asset class rated on the assumption that one guarantor absorbs losses across the market, until a downturn demonstrates the guarantee was never diversified. Bond insurers in the mid-2000s are the reference case. Aircraft leasing saw a milder version when manufacturer buyback commitments came due into soft markets.

The more important exposure sits outside credit. The manufacturer offering residual-value support also controls the software roadmap that determines when the collateral's economic life ends. The guarantor does not merely stand behind the asset's value; it sets the asset's useful life. Those two exposures are correlated by construction. The conditions under which a manufacturer would most want to accelerate adoption of a new generation are precisely the conditions under which the outstanding guarantee on the prior generation becomes most expensive to honour.

No published rating methodology models this. Fitch has an open consultation on whether GPU depreciation should be formally built into AI-infrastructure securitisation ratings, and finalised criteria in late 2025 for single-borrower data-centre securitisations blending ABS and CMBS approaches. But that framework rates the data-centre wrapper, not the chip as discrete collateral. The gap is not an oversight. The input required does not exist in any public dataset.

The backlog and the collateral are the same bet

The structures being written rely on contracted revenue as much as on collateral value. CoreWeave's $8.5 billion facility, non-recourse and ring-fenced, is secured by GPU clusters and a contracted backlog, and it earned an investment-grade rating on the strength of both. The structure assumes that if the metal disappoints, the contracts pay; if a counterparty fails, the metal is there.

The Cement Curve argues that this pairing is not two independent supports. The obligors are merchant commodity producers: energy-heavy and capital-heavy, with revenue concentrated in a small number of counterparties and breakeven utilisation high enough that modest slippage moves them through the floor. Their contracted backlogs are written against a spot market whose rates have already fallen sharply from the 2023 peak. The operator's own accounting response to that pressure, extending assumed useful lives to four, five, six years to protect reported earnings, is a bet against the same generational transition the collateral is exposed to.

The belt and the suspenders are therefore buckled to the same event. A generational transition that strands the silicon is the same transition that pressures utilisation on the installed base, compresses renewal pricing on the backlog, and forces the depreciation catch-up the operator has been deferring. Collateral value, contract value and obligor solvency do not deteriorate independently in that scenario. They deteriorate together, on the same trigger, and the structure's diversification is notional.

The operator analysis and the collateral analysis are describing one exposure from two directions.

What this should change in the structure

Four things follow, and none require believing the asset class is unbuildable.

Advance rates should key to support horizon, not calendar age. Two pools of identical vintage can face different depreciation exposure depending on architecture and framework roadmap. Age is a proxy that will break precisely when it is needed.

The software-support commitment belongs in the documents. If a manufacturer's residual-value support is worth investment-grade treatment, a committed minimum support horizon for the financed architecture is the natural companion covenant. Its absence tells you the guarantee is a marketing instrument rather than a structural one.

Power should be underwritten as a covenant, not an operating detail. Contracted capacity and interconnect position determine whether the cascade has a floor at a given site.

Contracted backlog should be stressed on the same scenario as the collateral, not as an offset to it. Modelling them as independent mitigants overstates coverage in exactly the state of the world that matters.

Valuation is the binding constraint. A market like this eventually produces specialist appraisers, servicers and loss-curve builders; the secondary marketplaces now trading used H100 and A100 inventory are an early version of that market. But no independent appraiser can build a defensible GPU residual model without a view on support horizon and energy economics, and the people who hold that view sit inside the ecosystem being underwritten. Until that expertise moves outside, "independent underwriting" describes an ambition rather than a practice.

What would prove this wrong

Three observable markers would invalidate the argument.

Older generations sustain rental value through a full generational transition. If Ampere-class rental rates hold through the next architecture cycle rather than stepping down, the cascade has a tail and the fleet analogy survives. This was named as an invalidator of The Cement Curve and it invalidates this note for the same reason.

Framework support for deprecated architectures extends rather than contracts. Community forks, vendor long-term-support branches, or compiler backports that keep new model architectures running on old silicon would blunt the cliff into a slope. This is the most plausible falsifier, and the one worth watching most closely.

Low-end inference demand absorbs deprecated silicon faster than it is stranded. If the volume of workloads that run acceptably on older hardware grows faster than the deprecation frontier advances, the floor moves out and residual values hold.

The market is currently underwriting all three as true by default. None is currently measured.

Sources and method

This is a framework note, not an initiation of coverage, and contains no recommendation, rating, or price target. It extends the argument of The Cement Curve: AI Coding and the End of the Managed Margin (Larix Research, 11 August 2026). NVIDIA and CoreWeave appear as illustrations of a structural argument drawn from public reporting, financing disclosures and rating commentary; they are not coverage. We hold no position in the securities mentioned and received no compensation from any party in connection with this note.

NVIDIA financing: memoranda of understanding with Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs and KKR to mobilise more than $500 billion of third-party capital, announced 10 August 2026, per NVIDIA Newsroom. NVIDIA shares fell approximately 2.8% on the announcement date, representing roughly $70 billion of market value, per contemporaneous market reporting. Jensen Huang's case-by-case residual-value support of up to roughly 25% of an opportunity; terms were not publicly disclosed.

GPU depreciation and software support: H100 hardware resale value decline of approximately 73% over three years, a secondary-sourced estimate distinct from the milder reported decline in rental rates. Michael Burry's estimate that hyperscalers understate depreciation by approximately $176 billion from 2026 to 2028 through five- and six-year schedules. Fitch's open consultation on GPU depreciation in AI-infrastructure securitisation ratings and Fitch's final criteria from late 2025 combining ABS and CMBS approaches for single-borrower data-centre securitisations.

CoreWeave and secondary markets: CoreWeave's $8.5 billion non-recourse facility from March 2026, described as the first investment-grade-rated GPU-backed financing, secured by GPU clusters and contracted revenue backlog. Secondary marketplaces for used H100 and A100 inventory, together with existing IT asset-disposition channels handling GPU resale logistics and reconditioning.

All market and financial figures were checked against the public record on 29 August 2026. Where a figure rests on secondary reporting rather than a primary document, that limitation is stated above.