Where the Risk Goes

The compute economy is a conversion chain: fab capacity becomes accelerators, accelerators become data-center capacity, capacity becomes compute-hours, compute-hours become tokens, tokens become application revenue. Each link has its own utilization rate, its own pricing mechanism or lack of one, and its own balance-sheet home. Five features of that chain generate essentially all of the risk this piece is about. The full treatment is in the primer; what follows is the compression a risk-transfer argument needs.

Serving a model is two different jobs

Reading a prompt (prefill) is a math problem; writing the answer (decode) is a data-moving problem, bounded by memory bandwidth rather than arithmetic. Decode is where output tokens are made — the ones that carry roughly 4× the input price and, with reasoning models, dominate volume — so decode efficiency is where inference margin lives. The workload mix between the two phases moves the relative value of every chip generation, because chips differ mainly in how much memory bandwidth they bring to the fight. That is why a cross-SKU spread is a workload-mix trade and not a spec trade.

Token prices deflate hard while compute spend grows

GPT-4 launched in March 2023 at $30 per million input tokens and $60 per million output; GPT-4-level output now sells for well under a dollar. The headline rate usually quoted is a16z’s roughly 10× a year for constant capability, which is the conservative end of the published estimates — Epoch AI measures between 9× and 900× a year depending on the benchmark, with GPT-4-level science questions around 40×. So far demand elasticity — agents, reasoning, video — has outrun the deflation. Every levered structure in this piece is implicitly short the scenario where it stops.

Hardware value dies in jumps, not drifts

A GPU cascades down-market as new generations arrive: training tenancy at contracted rates, then inference duty at falling spot, then retirement when spot rental hits variable cash cost — the same shutdown logic as a power plant. But the decline is not smooth. Value is lost on telegraphed release dates, when a new generation ships in volume or a new model class changes what workloads need. The right mental model is single-name credit — a known maturity wall with an unknown outcome — not a depreciating truck.

And the cascade is not monotone. H100 rentals fell from above $7 an hour in early 2024 to a $2–4 band through 2025, and then partially reversed: Nvidia raised H100 rental pricing in 2026 and the indices recorded a spike, with August 2026 quotes spanning roughly $2.19 to $11.06 depending on tier. Legacy supply stopped growing as wafers moved to the new generation, just as inference demand accelerated. This matters for the argument rather than against it — jump risk is two-sided, and a warehouse short the tail can be hurt by a squeeze in old silicon as easily as by a cascade in it.

The capital structure is mismatched to the asset

An oft-cited estimate puts GPU-collateralized debt above $20 billion; it is a press figure without a primary tally behind it, and probably stale on the low side, since CoreWeave alone reported $14.2 billion of total debt in 2026. What is precisely documented is the price. CoreWeave’s $2.6 billion first-lien delayed-draw term loan — priced in late July 2026 after flexing wider from initial talk, closed August 10 — cleared at SOFR+550 with an original issue discount of 97, a 10.44% yield to maturity, rated Ba2/BB+, arranged by JPMorgan and MUFG. It carries maintenance covenants: a 1.35× debt-service coverage ratio and a $112.5 million liquidity requirement, at a financing SPV rather than the parent. Contracted offtake plus GPU collateral was not enough to avoid them.

Meanwhile the vendor has stepped into the gap. Nvidia signed non-binding memoranda in August 2026 with Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs and KKR to establish financing platforms intended to mobilize over $500 billion of third-party capital over time — not Nvidia’s balance sheet, and not committed. It separately agreed in September 2025 to purchase up to $6.3 billion of CoreWeave’s unsold capacity through April 2032, and in August 2026 guaranteed up to $105 billion of conditional lease and power obligations to SB Energy supporting OpenAI’s twenty-year lease at the PORTS-Pike campus in Ohio. The manufacturer of the asset is now also, in substance, the residual-value insurer of the asset.

The same chip clears at three prices by commitment length

Illustratively and for a hyperscaler-adjacent posted rate: around $9 an hour on demand, mid-$5s reserved, high-$3s multi-year. The tiering is real and observable; the levels move by SKU and tier, and B200 and H100 both span a wide range depending on whether you are quoting a marketplace or a hyperscaler list. The structural cause of the discount is that operators are forced sellers of term — lenders size loans to contracted revenue — while few buyers want to commit long. As the companion piece showed, that discount is a bundle: measurement air, expected depreciation, utilization equivalence, optionality and credit, with perhaps a tenth of it representing clean, harvestable forward risk premium.

That is the machine. Now: what risks does it throw off, who naturally holds each one, and where can each one go?

The discipline for each risk is the same four questions: who is naturally long it and wants out, who is naturally short it or paid to hold it, what announced instrument connects them, and what is left over. The leftover column is the subject of section 04 — and it is the only column in this table that has no owner by design.

R1 is the row the announced instruments actually serve. Between them, the CME, ICE and Architect products give both natural sides a real instrument for near-dated rate direction: the operator sells the strip against unsold capacity, the enterprise buys it against a coming renewal. What they do not yet give anyone is size, tenor past the front months, or freedom from basis — an operator’s specific fleet, with its region, cluster size and service level, is not the index.

R2 has a natural buyer nobody likes to name. The economically natural writer of obsolescence protection is Nvidia. When old GPUs lose value faster, new-generation sales accelerate, so the chipmaker is structurally short obsolescence risk in its core business and can warehouse the long side more cheaply than anyone else. Its backstops and guarantees are exactly that trade, done in vendor-finance form. The catch is circularity — vendor financing supporting demand for the vendor’s own product is the pattern risk committees remember from telecom in 1999 and merchant energy in 2001 — and concentration: protection written by the entity whose own product launch is the trigger event has obvious wrong-way properties. That is why the insurance route matters even though it is slower and currently thin. It is also worth being precise about how thin: the two names most often cited are intermediaries rather than carriers, and one of them is still preparing to launch its underwriting vehicle.

R6 is where compute meets a market that already works. Power is the one input with a deep, mature hedging complex, and the data-center operator is a natural spread entity: short power, long compute. Energy-normalized compute contracts — quoting compute in megawatt-hour-equivalent terms — make that explicit and set up the AI-era analogue of the spark spread: the compute toll. A tolling structure, in which an offtaker pays fixed for capacity and the operator runs the plant, is precisely how merchant power solved the same who-holds-utilization-risk problem twenty-five years ago. It is the structure to watch, because it moves R1 and R3 off the operator in a single document.

Aligning instruments to the ledger, from most to least available.

Run every trade in the menu and a specific set of exposures still ends up warehoused by whoever intermediates — the dealer, in the widest sense: bank desks, non-bank liquidity providers, and the trading arms of operators themselves. Five residuals recur.

Take an illustrative book: a dealer runs $500 million notional of 36-month receive-floating compute swaps against operator clients, hedged with listed futures in the front year only. At the 30–35% annualized index volatility the rental history implies, potential future exposure on the unhedged tail peaks early — on the order of 35–50% of remaining notional at the horizon — before delivery amortization pulls it down. Those figures are a sketch to size the problem, not a model output; the point is the shape, which is front-loaded and self-liquidating.

What that consumes depends on who holds it. For a bank dealer, an uncleared bilateral swap on a new commodity index draws counterparty capital under SA-CCR, plus credit-valuation-adjustment capital against mostly sub-investment-grade counterparties, plus funding valuation adjustments on uncollateralized exposure where the client cannot post margin — and operators are cash-poor by construction, because their cash is in racks. For a non-bank dealer, the constraint is funding rather than regulatory capital: there is no repo market for GPU exposure, physical inventory is cash-funded outright, and derivative inventory consumes the firm’s own risk capital against a market with no reliable exit bid between roll dates. Physical GPUs held as trading inventory are the worst of all worlds — cash-funded, jump-exposed, and illiquid in exactly the states where you would want to sell.

What the dealer earns

The cost side is only half a P&L, and a piece that prices the warehouse without pricing the franchise is not making a real argument. Three revenue lines sit against the capital above. The bid-offer on client hedges, which in a young market with one-directional flow is wide by necessity rather than by greed — early OTC quotes will look offensive to clients, and that is a signal rather than a scandal. Structuring fees on tolls, prepays and securitizations, which is where the margin actually is while the flow business is subscale. And the carry on the tenor discount itself: a dealer receiving fixed on the back end of a forced-seller curve is being paid the risk premium the companion piece isolated, which is real, structurally sourced, and small relative to the headline discount. Whether that adds up depends almost entirely on how much of the capital cost above can be compressed — which is section 07.

The inventory above is unhedgeable outright — that is what made it inventory. But almost none of it is unpriceable. A family of spread relationships, each observable or nearly so, gives the warehouse marks for what it holds and proxy hedges for part of it. The difference between an unmarked residual and a marked one is most of the difference between merchant-energy math and a risk-managed book.

The inference crack. Token price multiplied by achievable tokens per hour, less the indexed rental cost of the hour — the refiner margin of AI. It prices the demand-side ceiling on every rental curve, since no operator pays more for the hour than the tokens are worth, and it is the mark for any utilization-embedded exposure: a toll or capacity guarantee is, economically, a position in this crack. That also answers the awkward cell in the ledger. R3 looks like it has no financial buyer, but the buyer is whoever sells this crack — utilization risk becomes financial the moment the crack is quoted. Both legs are observable today, token prices publicly and rental via the indices, which makes this the first compute spread that can be marked daily without a model. We built it from a live endpoint book.

The compute toll, or spark spread. Rental per GPU-hour, less the power burn beneath it, less facility cost — the data-center operating margin, and the direct analogue of the spread that organized merchant power. Its risk-management value is asymmetric, and this is the most immediately actionable idea in the piece: the power leg hedges in a deep, mature market today, so a dealer intermediating tolling structures can strip out the energy component immediately and warehouse only the genuinely new part. Every residual should be decomposed this way — hedge the legs that have markets, warehouse only what does not.

The GPU/RAM crack. Compute rental against memory — the conversion margin on new supply. For the warehouse it is less a hedge than a leading indicator: the crack compressing, with memory dearer and rentals softer, means new supply slows, which is bullish the warehoused legacy tail; the crack widening invites the next capacity wave. It prices the supply response that every back-month mark depends on.

Curve-implied depreciation. Once the listed curves print, the calendar slope is the market’s cascade rate — the first observable replacement for accounting assumptions in marking the warehoused tail. A dealer holding months 13–36 no longer needs to defend a model-derived decay path; it marks against the traded slope, and can hedge the slope risk with calendar spreads even where outright back-month liquidity is thin. This is the single largest reduction in valuation uncertainty — and therefore in reserves and audit friction — that the listed market delivers, and it arrives with the first credible settlement rather than with volume.

Cross-index basis. The same chip’s rental priced by different administrators — transaction-based versus blended, one composition versus another — is itself a spread with a history and a volatility. It determines proxy-hedge quality: a client swap on one index hedged with futures on another leaves the dealer running exactly this basis, and its measured volatility sets the hedge-effectiveness haircut and the capital held against it. As benchmark governance converges this spread should structurally compress — one of the few inventory lines with a built-in tailwind.

Event ladders as wing pricing. Exchange event contracts laddered on index levels are digital options, and a strip of them is a market-implied distribution — including the jump probabilities around release dates that no diffusion-based risk model captures honestly. Thin as they are, the ladders do two things for the warehouse: they price the wing, letting capital be provisioned against a market-implied gap probability rather than a committee guess, and they are the only existing instrument that pays off specifically in the jump scenario. A dealer buying deep-out ladder protection into a telegraphed release window is hedging exactly the exposure section 04 said had no hedge — imperfectly, in small size, but in a listed, cash-settled form.

The futures-versus-lease spread. The gap between the listed strip and physical lease quotes for the same capacity and tenor is the market’s price on everything the futures do not carry — utilization guarantees, service, credit, flexibility. It is the tenor-discount waterfall from the companion piece, printed as a single observable number. For the dealer it marks the non-financializable strips of any structured deal; for the market it is the convergence gauge, and that spread grinding tighter over time is what “the physical and financial markets knitting together” looks like on a screen.

The through-line: the warehouse becomes financeable not when every exposure has a perfect hedge, but when every exposure has a price — because a priced residual can be reserved against honestly, capitalized rationally, and sold at a level both sides can defend. The spread toolkit is how a compute book graduates from marks-by-committee to marks-by-market, one relationship at a time.

The saving grace of compute inventory is its shape. A capacity swap is a strip of monthly deliveries, not a bullet: notional amortizes every month as hours deliver, so exposure is front-loaded and self-liquidating. The book, left alone, melts.

And the dominant risk is calendarized. The jump events are telegraphed release dates, which means wing risk is concentrated in identifiable windows rather than smeared uniformly. A dealer can schedule capital against known event windows the way an options desk schedules earnings — and, crucially, each survived release date seasons the remaining book. A tail that has lived through one generation change has revealed its cascade behaviour, reprices tighter, and can be held against less capital or sold down at better levels.

What reduces the liability over time

Each development converts a slice of warehoused inventory into something cheaper to hold or possible to exit, and the order tracks the maturation sequence from the companion piece. Clearing and listed tenor extension transform bilateral capital into cleared margin — netted, mutualized, far lighter per unit of risk — while every added listed month converts warehoused tail into hedgeable curve. Benchmark maturation shrinks basis directly: the closer disclosed, regulated index composition sits to real fleets, the less residual composition risk each client trade leaves behind. Assignability of physical contracts is the exit door: the day term compute contracts can be novated, a dealer can originate-to-distribute — warehouse a position while it seasons, then sell it down to insurance and asset-management balance sheets built for hold-to-maturity, the same pivot that took leveraged lending from bank balance sheets to CLOs. Securitization term-out does the same in size: warehouse, season through one release date, package the survivors. Reinsurance capacity against credible triggers takes the wings, which is the one exposure a trading book should not hold structurally. Fleet telemetry and utilization data gradually make R3 underwritable, moving tolling risk from dealer inventory toward actuarial books. And the vendor’s floor bid is today’s exit valve of last resort, worth exactly as much as its continuation is credible in the states where you would need it.

The compute economy’s risks all have natural sellers, and most now have at least one announced instrument pointed at them. What the market does not yet have is enough balance sheet willing to stand between — and the analysis above says why. Today, intermediating a single 36-month client hedge consumes bilateral capital, leaves basis and wing risk with no offset, and offers exits only through structures — assignment, securitization, reinsurance — that are themselves still being built.

The dealers will come. The exposure is too large and the premium too real for them not to. But they will come in the order the infrastructure lets them: front-month market-makers first, then swap warehouses as clearing and benchmarks mature, then originate-to-distribute once contracts can change hands.

Watch the warehouse, not the headlines. The market’s real capacity to manage compute risk is measured in dealer balance sheet per unit of client hedge — and every development that shrinks that ratio moves more of the exposure from accidentally held to deliberately priced.