Monetary Architecture & Infrastructure · Compute Finance
Compute is becoming the economy’s largest capital-goods class while lacking every property that makes an asset financeable
The build-out is real: hundreds of billions of dollars a year of chips, halls, and power — a capital-expenditure cycle of railway scale. What is not settled is the noun. Is a deployed GPU cluster equipment, property, inventory, or a commodity in waiting? The question sounds scholastic and is not, because the financing follows the classification — and the financing has already arrived, years ahead of the answer. Chips now secure multi-billion-dollar debt facilities, sit in bankruptcy-remote vehicles, and anchor structures that look like securitisation. Compute, in other words, is already capital — financed, booked, borrowed against. What it is not yet is independently legible collateral: an asset a lender can price without pricing the promises wrapped around it.
This essay runs compute through the four questions every secured lender asks of collateral: can I control it, can I value it, does it hold value, and what do I recover in an insolvency? The answers, in order: control attaches to contracts more than to silicon; the valuation rests on a depreciation schedule that is a negotiated forecast of future rents, currently the subject of a public accounting dispute measured in hundreds of billions; the structured financing inherits the credit of a handful of offtakers rather than the utility of the machines; the estate holds negative-carry hardware whose best buyer usually sits inside the structure that failed; and where the market has noticed the gap, its fix — the residual value guarantee — converts the missing curve into one more corporate promise, sold by the parties most exposed to its failure. The conclusion is a distinction: compute is already capital; it becomes independently legible collateral only when it can be pledged on its own observable value — and that needs a settled control layer, a market curve, and a recovery market that barely exist — the first listed compute futures arrived only weeks before this essay, and a listed contract is not yet a curve.
Every prior capital-goods boom knew what it was building in balance-sheet terms. Railways were fixed infrastructure on land; ships and aircraft were registered mobile equipment with mortgage conventions older than most banks; power plants were project-financed real assets with offtake agreements. The AI build-out is the first one whose asset class is genuinely unsettled. A deployed GPU cluster is movable equipment that behaves like a factory, prices like a commodity, obsolesces like consumer electronics, and earns like a hotel — occupancy times rate. The accountants book it as property, plant and equipment. The lenders take it as chattel collateral. The operators talk about it as a fleet. The strategists, this publication included, have argued it functions as a reserve asset. It cannot be all of these, and the financing terms differ by hundreds of basis points depending on which one it is.
The market has not waited for the taxonomy. The first facility secured by H100 processors was raised in August 2023 — $2.3 billion, led by credit funds, against chips as security (Reuters, 2023). Three years later the same borrower carried more than $30 billion of debt across a series of delayed-draw facilities, each housed in its own bankruptcy-remote special-purpose vehicle established to own the GPUs and borrow against them (company filings, 2026). Whatever compute is, it is already collateral. The purpose of this essay is to ask the question the term sheets defer: collateral for what, exactly, and what happens when the structure is tested?
Strip secured lending to its skeleton and the lender asks four questions of any collateral. Can I control it — attach, perfect, and if necessary seize, ahead of everyone else? Can I value it — against a market price, not a model? Does it hold value — at what rate does it decay, and is that rate knowable? What do I recover — when the borrower fails, what does the estate actually contain and who buys it? Land answers all four so well that the mortgage built modern banking. Aircraft answer them well enough that a specialised finance industry runs on registries, standard mortgages, and a deep second-hand market, with an international convention to make the control question portable across jurisdictions (Cape Town Convention, 2001).
Compute must be run through the same four questions, and the honest answers occupy the sections that follow. The preview: the answers are fragmented — legally available, economically incomplete — and the structures financing compute are workarounds for the gap between the two: sophisticated, rational workarounds, each of which quietly substitutes a different asset for the one on the label.
Taking security over a GPU is legally unremarkable — it is movable equipment, and every mature system has machinery for charging chattels. The difficulty is that the chip is not where the value lives. A processor in a crate is worth a fraction of the same processor deployed: racked, powered, cooled, networked into a cluster, wrapped in scheduling software, and — decisively — contracted. The economic asset is capacity under offtake, and capacity under offtake is not a chattel. It is a bundle: hardware, leases, power agreements, and customer contracts, each governed by different law, each failing differently. Name the distinction, because the essay turns on it: legal collateral and economic collateral are different things. A lender can perfect an interest over the chip — the legal collateral — while the economic collateral spreads across three layers financed as one: the hardware, which provides residual recovery; the infrastructure around it — power, shell, network — which decides whether the hardware stays economically alive; and the cash-flow layer — offtake, receivables, guarantees — which actually services the debt. The market finances all three simultaneously and calls the transaction “GPU-backed”.
The financing structures acknowledge this by their perimeter. The facilities are not secured on chips; they are secured on substantially all assets of a special-purpose vehicle that owns the chips, holds the contracts, and receives the revenues (company filings, 2026). Lenders size the debt against committed customer payments, not against resale value of silicon. Which licenses a blunt translation: a “GPU-backed” loan is predominantly a contracts-backed loan wearing hardware language — the collateral that matters is the offtaker’s promise, with the equipment as the salvage layer beneath it. Stated as a rule: the machine is the nominal collateral; the cash flow is the primary source of debt service; and the contract architecture decides how much of the machine’s economic value survives enforcement. And control of the bundle raises exactly the question the finality essay asked of securities: possession of the object, assignment of the contract, and priority in the proceeds are three different legal events, perfected in three different ways. A lender who has the chips but not the contracts controls the least valuable layer of the stack.
The second question — does it hold value — has become the loudest accounting dispute in the market, and both sides are arguing about the wrong thing. The schedules first: the industry has migrated from three-to-four-year useful lives for servers to five and six years, one large operator moved from four to six in 2023, and one hyperscaler moved a subset of its estate back down to five, citing the pace of AI hardware development (company disclosures, 2023–2026). Against the schedules stands the cadence: new architectures ship every eighteen to twenty-four months, and each generation’s arrival reprices its predecessor’s rent. Rental rates for the flagship 2023 chip fell from above $7 per GPU-hour at the peak to a broad band with a median near $3.30 by mid-2026 (independent pricing trackers, 2026). The bears have put a number on the gap — an estimate of $176 billion of understated depreciation across the industry between 2026 and 2028 (Burry, 2025); the operators answer with rebooking evidence: capacity coming off expired contracts re-let at ninety-five per cent of the original price.
Both sides are forecasting — and the two claims are not even opposites: an estimate of understated depreciation is a claim about the accounting schedule, while rebooking rates are evidence about revenue, so the real dispute is whether economic usefulness decays on the curve being booked, not whether the chips still earn. For this asset an accounting life can be nothing else: not a physical constant but management’s estimate of economic usefulness under technological uncertainty. A GPU does not wear out on the schedule; it is repriced by its successor. Its depreciation is therefore not an engineering datum but a forward curve for compute rents — held by companies whose earnings depend on the curve sloping gently. The purest expression of the circularity sits in the loan documentation itself: the facilities define the collateral’s depreciated value by contract, straight-line over a stipulated six-year life (facility agreement, SEC filing, 2026). The covenant turns an uncertain economic forecast into a contractual number — and the number then does the work of a fact.
The useful life of a GPU is not an engineering fact. It is a forecast of its rent.
The third question — can I value it against a market — is being answered by structure rather than by markets. Data-centre securitisation predates the AI cycle and works because it never touches the silicon: what is securitised are the shells, the power connections, and long leases to investment-grade tenants — real-estate credit in a server costume. The new generation of compute financings moves closer to the machines: SPVs holding GPU fleets and their offtake contracts, tranched debt against the committed revenues, equity beneath. Strictly, most of it is secured SPV lending rather than securitisation — and the distinction is itself the argument: a rated, tradeable asset-backed security demands exactly the standardised, observable collateral value that compute does not yet have. The rating question for such a structure is not “what is the equipment worth?” but “who is the offtaker and will they pay?” — because the pool’s diversification is an illusion. Behind most structures stand the same handful of AI counterparties, so the senior tranche is, in substance, a concentrated credit position on a few names, with hardware salvage as the recovery assumption.
This is the Wrapper Fallacy operating at industrial scale: the vehicle inherits the credit of the offtaker, not the utility of the machines, and no amount of structuring changes which one the investor is actually holding. It is also, at one remove, the mechanism described in The Bank and Its Shadow arriving in a new sector: credit funds financed the first, illegible phase at private-credit prices; the regulated balance sheets arrive as the assets become legible, bringing the funding subsidy with them. The spread between those two phases is the price of the missing market.
The market has noticed the residual-value problem, and in the space of a year it has manufactured a substitute: the residual value guarantee. The mechanics are simple — a technology company promises a minimum future value for the chips or the data centre; the SPV’s lenders are covered for any shortfall when the assets are eventually sold or re-let; the debt then prices roughly 100 to 150 basis points over the guarantor’s own bonds. The scale is not simple: up to $300 billion of such commitments were reported issued in twelve months, first by a hyperscaler to support a $27 billion data-centre financing, then by chipmakers backing SPVs that buy their chips and lease them onward to AI laboratories — within a wider tally of more than $3 trillion in off-balance-sheet commitments and credit support across the sector (FT, 2026). Bankers call the structures “balance-sheet efficient”: because the guarantor covers only the gap between sale proceeds and the guaranteed floor, little of the exposure is booked today.
Run it through this essay’s frame and the instrument is a confession. The residual value guarantee answers the third lender question — does the asset hold value? — by fiat: a corporate promise substituting for the forward curve that does not exist. And it dissolves the fourth question into the third, because the lender covered for any shortfall in a future sale is no longer underwriting the asset at all. The salvage layer, the last part of the structure that was supposed to be asset risk, has been converted back into credit. What is left is wrong-way risk of an almost architectural purity: manufacturers guaranteeing the future value of their own products, purchased by their own customers, in structures whose trigger — a collapse in compute values — is precisely the state of the world in which the guarantor’s revenues collapse with them. The monolines wrote this trade on mortgages and the lesson was expensive. There is even a refinement the monolines never managed: the party underwriting obsolescence risk is the party whose product roadmap causes obsolescence. A chipmaker guaranteeing the residual value of its current generation is, in effect, short its own release schedule — and it controls the release schedule. The cruellest feature of the design is that the trigger is success: the very technological leap that drives demand for the new generation is what strikes the guarantee on the old one, so the liability crystallises not on the guarantor’s failure but on its triumph.
The rating agencies say they adjust for the guarantees where the guaranteed value exceeds a stressed sale estimate, and report that the gap is “closer to zero” most of the time. That arithmetic assumes the stress is idiosyncratic — one campus, one fleet, sold into a functioning market. The scenario the guarantee exists for is the other one, in which every GPU fleet and every hall comes to market at once, correlation goes to one, and the stressed values were marks from a world that no longer exists. None of this means the structures fail; it means they should be read for what they are. The residual value of compute is not, today, a market observation. It is a liability of the technology sector, wearing the vocabulary of an appraisal.
The fourth question is the one no term sheet dwells on: what does the lender actually hold on the day the borrower fails? For land, the answer is land — it waits, costlessly, for its next owner. For an aircraft, a registry, a repossession convention, and a liquid second-hand market. For a failed compute operator, the estate holds hardware that is depreciating by the month, inside leased halls with live power and cooling obligations, wired into customer contracts that the insolvency itself may have given counterparties the right to terminate. Compute is negative-carry collateral: unpowered, it earns nothing and still costs money to hold; powered, it must be operated, which an administrator cannot do and a lender does not want to.
The crypto-mining insolvencies of 2022 were the dress rehearsal — fleets of specialised silicon, hosting agreements, power contracts, and the discovery that the machines were worth most to whoever already controlled the site and the electricity. The mining insolvencies offer one plausible recovery geometry for compute at scale: not repossession and sale of equipment, but a whole-business transfer to a buyer who already controls an adjacent layer of the stack — the power, the shell, or the offtake — and such buyers sit as often inside the failed structure as outside it. Either way the recovery value is bilateral, not market. A lender’s model that assumes an orderly secondary sale of a hundred thousand deployed GPUs is assuming a market that has never yet cleared at that size.
Run the comparisons and the diagnosis converges from every direction. Real estate is financeable because land persists and location rents are observable. Aircraft are financeable because the objects are standardised, registered, long-lived, and resold into a deep market — and even that required an international convention to settle the control question. Commodities are financeable because they are fungible and carry the instrument compute has only just acquired: a forward price. In October 2026 the first listed compute futures began trading — cash-settled monthly contracts on indices of hourly H100 and B200 rental rates (CME Group / Silicon Data, 2026). It is the right development, and it changes less than it appears to, because a listed contract is not yet a curve. Two cash-settled contracts on two chips are not a term structure, and the horizons at which the collateral, the depreciation schedules and the structured financings live are four to six years out — tenors no strip yet reaches. Beneath the contract sits the index, and beneath the index sits the market: the same chip let in mid-2026 anywhere between $1.50 and $7 per hour depending on venue and counterparty, so the benchmark is, for now, an average over bilateral prices wearing one ticker.
Three defects separate the new contracts from the curve secured finance needs. Basis: the indices track neocloud on-demand rates — the most volatile sliver of the market — while the collateral inside the SPVs earns long-dated contracted offtake, so the hedge and the asset price different things. Succession: the contracts are chip-specific, which makes this the first futures market whose benchmarks themselves depreciate — an H100 future obsolesces with the H100, and rolling a hedge across chip generations is an unhedged basis no oil trader has ever faced. And history: the last attempt to commoditise a capacity market this heterogeneous was bandwidth trading, which died of non-fungibility and the collapse of its most enthusiastic market-maker. Until the strip grows term, liquidity and survivors, the root defect stands: no honest mark for the collateral, no observable basis for a depreciation schedule, no pricing reference for a structured tranche, and no way to underwrite recovery. The sequencing is worth stating plainly: the market never needed a curve to finance compute — the contracts did that. It needs one before compute can be collateral priced by markets rather than by counterparties. Every structure in the market today is a private substitute for that public curve: the offtake contract substitutes for the forward price, the negotiated useful life substitutes for the decay rate, the counterparty’s rating substitutes for the mark, and the residual value guarantee substitutes for the salvage market itself. Compute finance in 2026 is what commodity finance looked like before exchanges: real, large, profitable — and priced off relationships because it cannot yet be priced off markets.
The gap closes in a known order, because other asset classes have walked it. First, the unit: a standardised, attestable definition of delivered capacity — the thing the Edge Reserve imagined as tokenised capacity titles, and which even a conventional market needs as its contract grade. Second, the control layer: a registry of interests in high-value deployed equipment and its attached contracts, doing for compute what the Cape Town Convention did for airframes — and, if the titles are ever tokenised, requiring exactly the property-and-finality statutes examined in Settled Is Not Final, because a capacity title that is fast to transfer but never final is collateral no serious lender will take. Third, the curve: benchmark rents and a forward market with real term, from which honest depreciation and honest marks fall out mechanically — the one piece whose first instalment has now arrived, in October’s listed futures, and which is complete when the strip reaches the tenor of the loans. None of this is exotic; all of it is infrastructure, and until this year infrastructure was precisely what the financing boom had skipped.
The prediction follows the pattern of every asset class before it. The pioneer phase is financed by whoever can price illegibility — private credit, at private-credit spreads. As the unit, the registry, and the curve arrive, the asset becomes legible, the regulated balance sheets enter with their funding advantage, and the spread compresses into the banking system. The interesting question is not whether compute becomes bankable — the sums make that inevitable — but how much of the transition happens before the first large estate is tested, because insolvency is the only audit of collateral that cannot be postponed.
Answer the title honestly, then: compute has already become capital — financed, booked, borrowed against, at railway scale. What it has not yet become is independently legible collateral: an asset a lender can control as a bundle rather than a chattel, mark against a curve rather than a covenant, and recover through a market rather than a negotiation with the failed structure’s neighbours. Each condition is buildable, and in October 2026 the middle one acquired its first instrument. None is finished.
Until then, the honest description of the present arrangement is simpler than the term sheets make it look. The build-out is real, the machines are productive, and the debt is genuinely secured — on the creditworthiness of the small number of companies that have promised to rent the capacity. Even the salvage layer has now been converted: where the market lacked a residual value, a residual value guarantee was sold instead — by the makers of the machines. “GPU-backed” means what the small print says. The machines are capital goods. The collateral is still credit.
Written in a personal capacity. Analytical views only — not legal or investment advice.
Julian Gretzinger — Investor and writer on monetary history, real wealth mechanics, and financial markets. substack.com/@juliangretzinger