The CFTC extended its review of NYMEX’s two compute futures on September 21, pushing the October 5 listing to a November 9 deadline. The comment file will spend those weeks on whether the settlement index is honest. A working paper that has been on SSRN since August asks the question that comes after that one.
Compute Futures: Who Is Left Exposed?, from HEC Montréal, builds the first formal model of who a GPU rental-rate future hedges and who it does not, and its answer is asymmetric: a broad index removes essentially all of a flexible buyer’s price risk and less than a third of a locked operator’s. The result is a model, not a measurement, and the authors say so on every page. But it turns the question the regulator asked — is the index honest? — into the question the market has to answer next: even if it is, what does the exposed side use instead?
This piece does three things. It reads the paper for a practitioner. It maps its result onto the participants who were going to use the contract. And it lays out the instrument set that would have to exist, on top of the listed price, for the side the index leaves out — with payouts, writers and data requirements, and with a view on which of them can be built in the forty-five days the delay just bought. Two of the figures are live: a stylised version of the paper’s coverage result in section 02, and an operator’s hedge ledger in section 04 that runs the instruments against a stress.
A final extension on a question about settlement integrity. The design question is still open.
NYMEX filed submission 26-370 on August 11 as a Rule 40.3(a) voluntary request for approval — the exchange chose to ask rather than self-certify — for Silicon Data H100 and B200 Rental Index Futures: 730 GPU-hours per contract, settling on the arithmetic average of the daily index over the contract month, thirty-six months listed.
On August 19 the Commission issued a request for comment that “preliminarily believes” compute “may not yet exhibit” fungibility, standardization and sufficient liquidity, that price formation “primarily occurs in opaque bilateral transactions,” and that asked how to stop a provider “manipulating a cash settlement index by adjusting a posted rate.” Comments close October 20. On September 21 the Commission took the forty-five-day extension Regulation 40.3 allows for “novel or complex issues,” to November 9. Under the rule that is the last extension available without the exchange’s consent; if the Commission has not acted by then, the contracts are deemed approved.
Read plainly, the extension is about the settlement series: whether posted rates are transactions, whether contributors can move them, whether the cash market behind them is deep enough to see. Those are the right questions for a regulator, and the comment file will argue them. They are not the only questions. A settlement index can be honest, reproducible and unmanipulable and still fail as a hedge for most of the people the press release named, because the index is one number and the exposures are many. That is the paper’s subject, and the delay is the first window in which the market can take it seriously before, rather than after, the contract trades.
One locked seller, one flexible buyer, one index. The index can only be built for one of them.
Fernández-Fuertes and Grégoire start from an observation anyone who has priced a GPU cluster will recognise. A neocloud is built around a configuration — chip generation, interconnect, region, sales channel — and earns exactly that configuration’s rental rate. It cannot diversify that in ordinary operations. A buyer of spot capacity can.
The buyer spreads spending across configurations, and its cost behaves like an expenditure-weighted basket. Put those two facts into a factor model, where every configuration’s price moves with a few common forces (the paper labels them demand, scarcity and obsolescence) plus a configuration-specific residual, and the asymmetry follows mechanically: configuration-specific shocks wash out of the buyer’s basket and stay in the seller’s revenue.
Then the design result. A settlement index only matters through its relative weights — doubling every weight just halves the hedger’s position — so a single index spans one direction in the space of common factors. The buyer, holding one basket, contributes at most one direction. Any further contract has to be justified by heterogeneity among sellers, and every further contract splits hedging interest across a thinner book. Menu size, in the paper’s phrase, is a sell-side object.
The numbers are illustrative, and the paper is scrupulous about what is estimated and what is assumed. Thirty-two configurations, four generations by eight provider-region variants; index moments matched to the published Silicon Data and Ornn series; loadings, seller weights, the seller share of hedging demand (set at one half) and the buyer’s substitution elasticity all assumed.
| Design | Share of buyer’s cost variance removed | Share of locked sellers’ rental-rate variance removed |
|---|---|---|
| Broad equal-weighted index | 0.997 | 0.282 |
| Exposure-weighted index | 0.993 | 0.283 |
| Best single direction the model allows | 0.997 | 0.279 |
| One index per generation | 1.000 | 0.369 |
The mechanism is simple enough to run by hand, and the explorer below does. It is a stylised version of the paper’s setup, not its calibration: every configuration’s price variance is split into a share that is common to all configurations, a share that belongs to its chip generation, and a share that is its own. A broad index averages across all thirty-two; a per-generation index averages across eight. The locked seller keeps its whole configuration; the buyer holds a basket. Coverage is the share of each side’s variance the index can reach. With the residual set at 72% and the generation share at 3%, the model lands within rounding of the paper’s numbers; move the sliders to see which assumption carries the result.
Where a configuration’s variance comes from
The hedgers
Three further findings sit around the main one, and each has a direct bearing on the contracts as filed.
The publishers disagree, and it is not noise. Silicon Data’s and Ornn’s H100 indices correlate 0.17 weekly on their overlap and 0.24 on the balanced sample; hedging one with the other removes 3% of variance. The paper cannot say whether that is economic dispersion or methodology, because it observes published levels and not the panels beneath them. Either way, “publisher identity and index construction are explicit contract terms rather than details.” It is the number Friedman led with this week in arguing that a useful price index is not yet a usable settlement, and it is the right one.
Only spot is published. Much of what buyers spend and operators earn is committed at the reserved, take-or-pay tier, and no reserved price is public. A cash-settled contract must settle on spot and covers reserved exposure only insofar as the two tiers move together — which cannot be tested without the reserved prices nobody publishes. A buyer inside a take-or-pay commitment is, for its term, locked: seller-like.
The settlement window has a price. Settling on the last week’s average keeps 95–98% of the hedge quality of a point settlement; settling on a full month’s average keeps 53–83%. The NYMEX contract is the full-month case, which is the right choice for matching monthly revenue and the wrong one for anyone whose exposure is dated.
Two smaller results round it out. The Silicon Data token index correlates 0.03 with the H100 rental index over their seven-month overlap and is twice as volatile: a different asset for a different hedger. And equity returns of twenty-eight listed firms show no positive configuration exposure for suppliers, so hedging demand cannot be read off stock prices; it has to be observed directly, from invoices and positions the public record does not contain.
It selects no benchmark. It does not say the contracts should not list. Its one signed result on the futures premium — that obsolescence news pushes sellers and buyers the same way, so that factor has to be paid for — rests on assumed loadings the authors flag as such. Its architecture, one broad hub plus a sparse basis layer, is offered as a conditional design that needs a reproducible charter, auditable inputs, cross-publisher units and out-of-sample tracking before anyone should adopt it. The paper restricts the candidate designs. It does not rank them.
The participant map, re-read through the model
The Contract Has a Date mapped who holds a floating exposure to a GPU-hour and found that every commercial participant with a real one prices it fixed. The paper adds a second column: for those who do carry floating exposure, how much of it the index can reach.
| Participant | Exposure | What the model says | Share the future can remove | What is left, and where it goes |
|---|---|---|---|---|
| Neocloud, merchant capacity Locked seller | One configuration’s posted rate on unsold hours | The locked seller of the model | About a quarter to a third of rental-rate variance; up to ~0.7 in the friendliest scenario | Configuration residual (variant, fabric, tier, site); generation turnover; and utilisation, which the model does not price at all |
| Aggregator or reseller on take-or-pay | Bought fixed, sells by the hour | A buyer that the commitment has made seller-like | Same as the neocloud for the committed block | Basis between the block and the index; fill; the tier discount the index cannot see |
| Enterprise or lab buying spot Flexible buyer | Basket across providers and generations | The flexible buyer the index is built for | Nearly all — if its spending is dispersed | Concentration: reserved commitments, workload lock-in and vendor switching costs shrink the basket toward a single configuration |
| Enterprise buying tokens | Bill in dollars per token | A different asset (0.03 correlation) | Little | Everything; needs a token-denominated reference, not a rental one |
| Lender on GPU collateral | Recovery value and renewal price | Asset-value risk, not the operating cash-flow risk the future carries | Proxy only | Rent and value share a driver, not a ratio; vendor backstops address the counterparty, not the price |
| Anyone marked on one publisher and hedged on the other | Ornn-settled prediction and perpetual venues against a Silicon Data future | Two assets, 0.17 correlated | 3% | The whole gap, plus restatement risk at either publisher |
The row that matters most is the first. The operator is the natural short in every exchange press release, the participant the regulator’s fungibility question is really about, and the one the model says the index serves worst. And the model’s 28% is a share of price variance. The operator’s larger risk, as the fill arithmetic in the desk literature and the utilisation basis in our earlier work both show, is hours sold, not price per hour. A cash-settled price contract cannot touch it. The true share of a neocloud’s revenue variance that the future removes is smaller than the paper’s number, not larger.
A hedge that removes a quarter of variance is still a hedge, and the first WTI contract was a poor hedge for a Gulf Coast refiner running Maya. The market adopted an imperfect index and let differentials become what the bilateral market negotiated. That is the right precedent, and it is also the point. In oil, the differentials found a dealer who would price them. The paper’s result is that in compute the differentials are most of the risk, and no listed contract will ever carry them — which is an argument for the dealer layer and the basis products, not against the listing.
What has to exist on top of the listed price, for the side the index leaves out
The paper’s map of unhedged exposure is a product list. Each row below takes one exposure the future leaves open, names who holds it, and specifies a payout that references the listed or published price on one leg and a measured private series on the other.
The organising principle is the one the paper’s liquidity result forces: the common component goes to the exchange, and the residual is warehoused and priced bilaterally — never listed as a thin contract of its own. That is hub-and-basis, done the way oil and gas actually did it, through a dealer.
| # | Exposure the future leaves open | Who holds it | Instrument | Payout, in plain words | Who writes it | Buildable from |
|---|---|---|---|---|---|---|
| 1 | Configuration residual | Locked operator; take-or-pay buyer | Fleet basis swap | Fixed spread against (own realised rate − hedge ratio × index) | Dealer; insurer on the tail | Private Fleet invoices or metered rate; basis grid |
| 2 | Generation turnover | Older-generation operator; its lender | Regrade spread option | Pays when H100 ÷ B200 settles below a strike set off the curve | Dealer against the two listed contracts | Public Listed settlements only |
| 3 | Publisher disagreement | Anyone marked on one index, hedged on the other | Index-gap swap with methodology stop | Fixed gap against realised (Ornn − Silicon Data) | Dealer | Public Published series only |
| 4 | Settlement-window mismatch | Anyone with a dated exposure | Convergence swap | Point value on the hedger’s date against the monthly average | Dealer | Public Published series only |
| 5 | Reserved-versus-spot tier | Take-or-pay buyer; operator with a term book | Commitment regret option; tier swap | Pays when spot settles below the committed rate net of a tier discount | Dealer; needs a reserved series | Gap Contract curve (not yet public) |
| 6 | Quantity — fill, utilisation, outage | Every seller; every aggregator | Metered-hours swap; revenue floor | Price leg on the index, quantity leg on the meter | Dealer for price × quantity; insurer for outage and fill tail | Gap Attested usage record |
| 7 | Downside with no natural buyer | Operator with debt service; lender | Yield floor and collar | Put on the index at debt-service breakeven; cap sold to fund it | Dealer; lender requires it | Public Listed settlements; a lender covenant |
| 8 | Token margin | Host; refiner | Crack swap | (Token index × tokens per chip-hour) − rental index, against a fixed crack | Dealer | Private A published token index; a throughput series nobody publishes |
| 9 | A fleet that is not the index | Fund; lender; operator | Custom basket swap | Fleet-weighted basket against fixed; common part laid off in futures | Dealer | Public Grade and region sub-series, at one publisher |
Before the nine in detail, the ledger. It takes the paper’s locked seller — one configuration, one posted rate, a fleet it cannot diversify — and runs a stress through it with the index future, the fleet basis swap (1) and the metered-hours option (6) switched on one at a time, against the debt-service line a yield floor (7) would be struck at. The future covers the share of the index move it is sized for. The basis swap covers the configuration’s own drift against the index. The metered-hours swap moves the notional from expected hours to hours actually sold. Nothing on the list pays for the hours that were not sold; that is the revenue floor and the insurer, and it needs a meter nobody publishes.
The fleet
The stress
The instruments
The operator’s problem in the model is that the broad index removes 28% of its rental-rate variance and leaves 72% — a little of it generation mismatch a segmented contract could reach, most of it configuration. The swap covers the 72%. The operator pays its own realised rate on a defined fleet, metered and attested, and receives the index times a hedge ratio plus a fixed basis, per chip-hour on a notional hour count. The operator ends with a fixed rate on its own fleet; the dealer holds the basis, lays the index leg off in the listed contract, and warehouses the residual, priced from the exchange-for-physical grid, ComputeConnect’s adjustment tables by SKU, memory configuration and location, and whatever resale prints exist — with a cap on the monthly basis move and a configuration schedule in the confirmation.
The paper says the price of this swap is not small, because the residual is most of the seller’s variance and none of it can be laid off in any listed product. That is why it belongs with a dealer, with the tail placed to an insurer.
Under the paper’s assumptions, obsolescence news pushes sellers and buyers the same way, so the future will not carry that factor cheaply. The direct hedge for a Hopper-locked operator is a spread between the two listed contracts, long the newer generation and short the older. As an option, the strike is set at the current ratio less the depreciation the curve already implies, so the option pays only for news — a faster fall than the market expected. The paper’s own drift estimates from the published series are H100 −11.2% a year, B200 +6.2%, A100 −0.3%.
A lender can require it against Hopper collateral; an insurer can write the far strikes on the value leg; the dealer runs the near strikes against GPU1 and GPU2. This is the “obsolescence event structure” the risk ledger called the cleanest design gap in the market, expressed on the contracts that are about to list.
A 0.17 correlation between the two H100 indices means anyone whose bilateral contract references Ornn — every prediction market and the offshore perpetual venue do — and who hedges on the NYMEX contract is running an open gap. The swap pays a fixed gap against the realised gap, with three features the paper makes necessary: a per-month cap on the payout; a termination clause triggered by a published methodology change or a restatement above a stated threshold at either publisher; and a pricing input that is a kept record of the gap itself.
Silicon Data’s own history supplies the case for the stop: the December 2025 restatement moved SDH100RT −4 to −6% and SDA100RT +35–40%; the March and June 2026 provider additions moved SDH100RT −3 to −7% and SDB200RT by up to −6%. This is the cleanest way to charge for a risk the contract design created by choosing a publisher.
The NYMEX contract settles on the month’s average. A hedger whose exposure is dated — a contract that resets on the first, a delivery, a borrowing-base mark on a fixed day — keeps 17–47% of its variance under the paper’s month-average numbers. The dealer holds the intra-month path, which is small on series where a fifth to three-fifths of days print unchanged and large in the months where the series jumps. Price it off the realised distribution of point-minus-average in the published history; put the same methodology stop on it.
It is the smallest instrument on the list and the one most likely to be asked for first, because every structured product written on the future has this gap inside it.
A take-or-pay buyer at a reserved rate is locked for its term at a tier the index does not price. The option compensates the buyer when spot falls far enough below its commitment that the commitment is out of the money, net of the discount reserved capacity normally earns. The tier swap is the same exposure without the option, for an operator whose term book is marked against spot.
Both need a reserved-tier series that does not exist publicly; until it does, the option is a put on the future with a haircut for tier co-movement that nobody can yet measure, and the paper is explicit that this co-movement is the untested assumption behind hedging any reserved exposure on a spot index. The alternative is contractual: an index-linked reserved contract, which avoids the risk rather than hedging it, and which the public record does not yet contain.
The paper’s coverage numbers are shares of price variance; the operator’s larger variance is quantity. The swap’s notional floats with the meter, so revenue per hour is fixed on the hours actually sold. The revenue floor covers a price fall and a fill shortfall at once. The dealer hedges expected hours in the future and holds the covariance between price and quantity; the insurer takes outage and fill.
This is the instrument the paper cannot see, and the one an operator’s lender cares about most, because debt service is paid from revenue, not from rate. It needs an attested usage record, which no administrator publishes and which is the single largest data gap in the complex.
For an operator with a debt-service breakeven, a put on the index struck there, sold with a cap above the forward, costs little and gives a lender a floor it can lend against. The paper’s hedging-pressure result matters here: because the natural sellers of downside are few, the put will be dear, and the collar is what makes it affordable.
A lender’s covenant — hedged hours valued at the floor in the borrowing base — is what converts this from a product into demand for the future. That covenant is the sentence The Contract Has a Date said to watch for. It still has not been written.
A host earns tokens and pays for chip-hours, and the two barely co-move. Hedging either alone leaves the margin open. The rental leg is laid off in the listed future; the token leg needs a graded token price series and a throughput series to convert tokens into hours. It is the inference crack from the spread toolkit written as a contract, for the refiner and the toll host whose exposure The Refiner and the Merchant described as three bases at once. The paper’s 0.03 is the reason it exists: nobody can build it from the listed contracts alone.
The paper’s cleanest positive result is that the buyer’s own expenditure weights always form a feasible index and, when buyers are the only hedgers, the best one. The same holds fleet by fleet. Rather than list more contracts — which the paper says thins the book — a dealer writes a swap on the counterparty’s own basket, fleet-weighted by generation and region, hedges the projection of that basket onto the listed contracts, and holds the rest. The pricing problem is the projection and the residual charge, and it is a standard one.
Every instrument moves the common component to the exchange and leaves a residual with the writer: configuration basis (1, 9), regrade beyond the curve (2), publisher gap (3), intra-month path (4), tier co-movement (5), price-quantity covariance and fill (6), downside skew (7), token-rent margin (8). The paper says the first is the largest and none can be listed. Three things make holding them a business rather than a bet: measurement — a usage record, resale prints, a gap record, a contract curve; an insurer for the tails that are events rather than prices — regrade, outage, fill, restatement; and netting across counterparties whose residuals run opposite, which the paper’s obsolescence result warns will be thinner than a symmetric picture suggests. Where the Risk Goes called the warehouse the bottleneck. The paper is the proof that the warehouse is not optional.
One hub, or three? The paper answers the question the exchanges are about to settle by listing.
Three exchange complexes are in the queue — NYMEX on Silicon Data, ICE on Ornn, Nodal on Compute Desk — plus a designated-contract-market application from Architect and the prediction-market ladders on Ornn. The paper’s liquidity result reads directly onto that: every additional contract that does not span a new direction of common risk splits hedging interest without removing variance.
Two H100 contracts on two publishers that correlate 0.17 are, in the model’s terms, two different assets, so the split is not even a split of the same hedge. And the paper’s own menu count — one contract when sellers are a small share of hedging demand, one or two across the grid — is already what NYMEX filed: an H100 contract and a B200 contract, which is the generation-segmented case that lifts the model’s seller coverage from 0.28 to 0.37.
That is the strongest thing the paper says in the contract’s favour, and it should be said plainly: two generation contracts on one publisher is a defensible menu. What it does not do is reach the configuration residual, and no menu will. The implication for the three complexes is not that two of them are wrong; it is that the market can only afford one hub, and the hub will be whichever publisher’s series the physical market and the dealers mark to. The paper’s list of what a hub needs — a reproducible charter, auditable inputs, units that are stable across publishers, out-of-sample tracking, and direct evidence of who hedges — is the scorecard, and today no publisher clears all five.
Hub-and-basis presumes a hub the physical market accepts, and the physical market prices fixed. Until a lease resets to an index or a covenant references one, the hub is a listing and the basis is an assessment. The paper is candid that its architecture is conditional on exactly that. The oil precedent says the physical market moves after the screen exists, not before; the gas precedent says the regulator unbundled first. Compute is running the film backwards, and forty-five days of review does not change which direction it is running.
Forty-five days is enough for the parts that need only published data
For the exchange and the publisher, three things cost nothing and answer the paper’s objections as far as they can be answered before trading.
State in writing which Silicon Data series settles the contract and whether a US-only cut exists, since the tier basis is undefined rather than large until then. Publish hedge effectiveness by clientele — a locked single-configuration seller against a dispersed buyer, on the publisher’s own constituent panel — not only by generation. And publish the point-versus-month-average tracking history, so the 53–83% is a known number on the contract’s own series rather than the paper’s estimate on a two-year record.
For a dealer, two of the nine instruments are buildable from published series alone and will be asked for in the first settlement month: the index-gap swap (3) and the convergence swap (4). Both need a kept record — the gap between publishers, and the intra-month path against the monthly average — that anyone with a Bloomberg terminal and a spreadsheet can start today. The regrade spread (2) and the collar (7) need only the two listed contracts and can be priced the week they trade. The custom basket (9) needs grade and region sub-series that exist at one publisher and not the other.
For a lender, the covenant. Half or more of uncontracted and renewal hours hedged on the index the deal is marked to, hedged hours valued at the hedge price, and a loan-to-value test against a published index. No chip loan on the public record carries any of the three. The first one that does is the demand the contract is waiting for, and it is a sentence in a term sheet, not a market.
For the paper’s own open question — the buyer expenditure shares and seller positions that would turn its restrictions into a selection — the data exists and is held by the wrong people. Marketplaces and aggregators see buyer shares across configurations; operators and their meters see seller positions; neither publishes. A lender working group would be the natural place for a reserved-tier series and a usage record to be pooled, because lenders are the participants with the most to gain from a mark they can defend and the least commercial reason to keep the inputs private. None exists on the public record.
The delay is about whether the index is honest. The paper is about whether it is enough. It is not, for the seller — and that is the case for building the rest.
The regulator’s forty-five days are spent on settlement integrity, and rightly. The paper’s result is that integrity is necessary and not sufficient: an honest, reproducible, unmanipulable broad index hedges the buyer’s basket nearly completely and the operator’s revenue by about a quarter, and no number of listed contracts closes that gap, because the gap is configuration and configuration cannot be listed. The exposed side is the one the exchange named first and the regulator worried about most.
None of that is an argument against the contract. It is an argument about what the contract is: a hub for the common component of compute price risk, on which everything the operator actually needs — the basis swap, the regrade option, the metered swap, the floor a lender will lend against — has to be written by someone who will hold what the exchange cannot. The instruments are not exotic. Every one of them is a commodity structure with a young index inside it. What is missing is the measurement beneath them and the balance sheet behind them, and the delay changes neither. It only makes the order of operations visible: benchmark, then basis layer, then the covenant that brings the natural hedger to the screen. November 9 is a date for the first of those. The other two do not have one.
What this piece rests on
The paper
- Rubén Fernández-Fuertes and Vincent Grégoire, Compute Futures: Who Is Left Exposed?, HEC Montréal, August 3, 2026, SSRN 7224223, with internet appendix (82 pages). Table 3 (per-side hedging effectiveness); Table IA.1 (published-index moments: drifts, volatilities, zero-change shares, correlations); Table IA.2 (input authority: seller share 0.50 assumed, elasticity 3.0 assumed); Table IA.6 (sensitivity envelope); Table IA.16 (settlement-window tracking); Table IA.17 (cross-index hedging effectiveness); Section IA.VII (token versus GPU-hour unit); Section IA.VIII (28-firm equity betas); Assumption 1(iii) (reserved capacity outside the buyer’s basket); Propositions 5–8.
The delay
- NYMEX submission 26-370 and CME SER-9785 (August 11, 2026); CFTC press release 9286-26 and 91 FR 54259 (August 19 and 21, 2026), comments due October 20; press reports of September 22–23, 2026 on the September 21 extension to November 9 under Regulation 40.3 “novel or complex issues”; Friedman, Compute Futures Need More Than a Useful Price Index (September 23, 2026), on the extension mechanics, the 0.17 correlation, and the financing-viability argument. The 40.3 deemed-approval reading follows Friedman’s; we have not seen the Commission’s extension letter itself.
Contract mechanics and index history
- 730 GPU-hours, monthly-average settlement, 36 months, from SER-9785 and Rule 1045101 as read in The Contract Has a Date. The identity of the settlement series remains an inference (Watch 1).
- Silicon Data restatements and provider additions: December 3, 2025 (SDH100RT −4 to −6%; SDA100RT +35–40%); March 25 and June 30, 2026 provider additions (SDH100RT −3 to −7%; SDB200RT up to −6%) — as recorded in The Contract Has a Date from Silicon Data’s index announcements.
Related Kinetic Alpha work
- The Contract Has a Date. It Needs a Dealer. (participant map, adoption paths, watch items) · Where the Risk Goes (risk ledger, structuring menu, five residuals, spread toolkit) · Five Indices, One Price · How Compute Index Providers Calculate Price · Open Source Is Not an Audit Trail · The Refiner and the Merchant · The Reserved Contract, Unbundled · The Compute Crack Spread · Inference Provider Hedging.
Claims register
- The 99.7% / 28.2% / 36.9% figures are model outputs under the paper’s assumed data-generating process and assumed hedger shares, not observed hedge performance; the paper says this in every table note and we repeat it. The 0.17 correlation, the 0.03 cross-index hedging effectiveness, the 0.03 token correlation, the settlement-window ranges and the drift estimates are published-index statistics through the paper’s July 2026 vintage.
- “72% left unhedged, most of it configuration” is our reading of the gap between the broad-index and generation-menu rows of Table 3; the paper attributes the unhedged share to factor-loading mismatch plus configuration residual without decomposing it further.
- The coverage explorer (Figure 2) is Kinetic Alpha’s stylised three-component variance model, not the paper’s calibration. Its defaults (72% configuration residual, 3% generation, 25% common) are chosen so that the broad-index and per-generation seller coverages reproduce Table 3; the paper’s loadings, seller weights and buyer elasticity are not represented. A buyer holding all 32 configurations equally is the index itself, so the explorer shows 1.000 where the paper’s expenditure-weighted buyer shows 0.997.
- The operator’s hedge ledger (Figure 5) is Kinetic Alpha’s arithmetic with round-number inputs, not a market quotation. The future is sized at the operator’s own rate per index unit on a share of expected sold hours and settles at the index’s stressed level; the basis swap fixes the configuration’s drift against the index on the same notional for a flat per-hour charge; the metered option sets both notionals to hours actually sold. Neither instrument compensates for hours that were not sold.
- The obsolescence sign result (Proposition 8) rests on assumed loading signs the authors flag as additional assumptions; a seller mix tilted to frontier hardware can reverse it. We describe it as conditional throughout.
- The nine instruments and their payouts are Kinetic Alpha’s design proposals, not products anyone has announced. “Buildable from published series” means the floating legs are public; it says nothing about counterparty appetite or capital.
- No public chip loan carrying a hedge requirement, hedge valuation or index-based loan-to-value test is a statement about the public record as of September 24, 2026, consistent with our earlier searches; a private facility would not appear.
- “Twice as volatile” compares the token index’s 49.9% annualised weekly volatility with the Silicon Data H100 series’ 24.3%, both from Table IA.1.
- The utilisation argument — that the operator’s dominant variance is hours sold — is ours and the desk literature’s, not the paper’s; the paper models price only and says so.