A cluster owner who is not a cloud has four exposures: the next month’s surplus or shortfall, the price of the block when it sells, the chip’s value in three years, and the credit of whoever is on the other side. No venue covers more than two. This piece maps the eleven that cover one or two, shows which gaps are left, and asks what it would take to join them.
The map is the first half. The second half is about money. Sixteen AI-infrastructure financings have been rated since May 2025, about $68 billion, and seven of them, about $22 billion, are secured on chips plus a customer contract. Those seven, and the unrated chip loans around them, already sort into tiers as cleanly as any mortgage pool. But the tiers are not senior, mezzanine and junior slices of one pool. They are separate loans, sorted by the credit of the customer whose contract sits behind the chip. A loan on a Meta contract is rated A3. The same borrower’s loan on shorter contracts is rated Ba2. A loan on chips alone pays nine to ten percent. That is the tranche structure compute has already built, and it tells you what the next one will look like.
The conclusion, stated up front: compute will tranche, but not the way mortgages did. Mortgage tranches sorted one pool of similar loans by order of loss. Compute tranches will sort dissimilar chips by three things at once: what the chip is and where it sits, how long the contract behind it runs, and who signed it. Basis, tenor, counterparty. A pool that ignores any of the three will be mispriced, and the rating agencies are telling you in public that they do not yet know how to price the first.
Eleven venues on two axes: how long the thing traded lasts, and what you get at settlement.
Read the map from the top. The top row is cash settled on an index: Kalshi’s fully collateralised contracts on Ornn, trading since early 2026, and the three exchange futures still waiting on the Commodity Futures Trading Commission. None of them delivers a chip or an hour. They pay the difference between a strike and a print. The second row is the one yesterday’s piece said was missing, and it is less empty than it looked: three venues now deliver reserved time at tenors between a day and a few months. The third row is hardware. The fourth is credit. The fifth is insurance.
The map has one empty region that matters. Nothing in the one-to-three-year column delivers reserved time through a venue with a visible price. That column is where the neocloud contracts live, priced bilaterally, and where Compute Exchange and Andromeda arrange capacity by request. The one-to-three-year tenor has a futures contract pending and a loan market functioning, but no secondary market for the contract itself.
| Venue | What it does | What it solves for the owner | What it leaves open |
|---|---|---|---|
| SF Compute | A central limit order book for H100 node-hours. Bids and asks in dollars per node-hour; fills in whole minutes from one node for one minute; partial and non-contiguous fills by default; first order on the book sets the price. Time bought on a firm reservation resells under the same rules. Fee is a fixed amount per node-hour plus a percentage of sale value; sellers receive resale credits, not cash. | The surplus problem, for H100s, inside SF Compute’s own pools. Standard Intelligence buys to peak and resells the trough; its chief executive says the cost of holding excess is “the small delta from price fluctuations.” | One chip family at scale (H200 showed no market price as of March 2026; B200 on request). Credits, not cash, so the surplus does not leave the venue. Only time in SF Compute’s pools can be sold there. An owner can bring a cluster onto the platform under SF Compute’s “own a cluster” programme, operated in full by SF Compute; it cannot list a cluster it runs itself. |
| Computable | GPU hours sold by calendar week at wholesale; future weeks allocated by sealed-bid auction; unused capacity sold back immediately. Backed by founders from Jump Trading and Coinbase; Y Combinator-backed; its own profile calls it a derivatives exchange for compute. | The week is the right unit for a training run or a pilot. Sell-back is the surplus leg. | No published volumes, prices or provider list. Commenters on its launch asked the question the Commission will: whether a venue that lets you buy future weeks and sell them back is a futures exchange. |
| Hyperbolic | Aggregates supply from neoclouds, enterprise idle capacity and customer resale into one control plane; sells on-demand, reserved and private cloud; states it will buy long-dated capacity on contract and sell it in short duration, hedged with a put on the final year. $80 million run-rate exiting September 2026. | The block holder yesterday’s piece asked for: a party that will take an owner’s idle capacity onto its own book and resell it. | It is a principal, not a venue, so its price is a quote, not a print. It takes the position the Commission’s comment file wants kept away from whoever publishes the benchmark. |
| Compute Exchange | Reserved capacity from more than 100 providers by request and auction; forward requests specifying chip, count, term, start and target price; a used and refurbished hardware market since July 17, 2026; token forwards. No custody, no principal positions. Sister company to Silicon Data. | The one-to-three-year procurement, and the exit for the chip itself when the owner is done with it. Hardware requests run from hundreds to tens of thousands of units. | No order book and no published traded price; each match is a negotiation. The forward request is an enquiry, not a contract. |
| Andromeda | Standardised contracts with service levels across more than 100 providers, an automated certification engine that benchmarks each provider’s cluster, one invoice. Claims more than 1,000 completed transactions. A Paradigm-led round in March 2026 at a $1.5 billion valuation; Paradigm’s total investment is $60 million. | The grading problem. A certified cluster on a standard contract is the beginning of a deliverable grade. | Procurement, not resale. The buyer’s own surplus has nowhere to go. |
| Kalshi on Ornn | Fully collateralised event contracts on Ornn’s transaction index, live since March 2026 in thin volume; the forward-curve dashboard was taken down in September and a Kalshi amendment was stayed by Commission staff on August 21 until November 19. | A public, tradable view of the on-demand price, with no leverage. | Cash on an on-demand print. Basis to a reserved block in a colocation cage is wide and unmeasured. |
| CME, ICE, Nodal futures | Cash-settled contracts: CME monthly on Silicon Data; ICE on Ornn and on NativX; Nodal on Compute Desk’s one-year United States rental indexes. CME’s review runs to November 9, 2026. | A hedge for the block holder and the lender, once listed. | Nothing for the owner whose exposure is quantity. Nodal’s contract on a one-year reserved index is the closest to the owner’s tenor and the furthest from a print anyone can see. |
| Neocloud delayed-draw term loans | Seven rated loans secured on chips plus a customer contract, about $22 billion, inside a register of sixteen rated AI-infrastructure financings. Investment-grade customers at SOFR + 2.25 to 3.00%; other customers at SOFR + 4.00 to 5.50%; the two facilities on chips alone at 9.00 and 9.95% fixed. | The financing leg, for borrowers with a contract to pledge. | The smallest rated deal is $810 million. Nothing for an owner with no customer contract because it is the customer. |
| Onchain GPU loans | Fifteen loans from $0.2 million to $98 million, 10 to 15% fixed, loan-to-value 57 to 80%, mostly B300 collateral, to smaller operators. | The only chip-backed credit available below the syndicated minimum. | The rate says what the lender thinks of the collateral without a contract. It is the uncontracted tranche, priced. |
| Nvidia rent-back | Since July 1, 2026 Nvidia rents back unused capacity from neocloud partners at a fixed rate, so their lenders will lend. | A utilisation floor for the borrower, which is the surplus leg solved by the vendor. | Available to Nvidia’s neocloud partners, priced into their contracts, and it moves the residual risk onto the vendor that also sells the next generation. |
| Residual value cover | A floor on the chip’s sale price at set trigger dates, up to three years, paid as the shortfall below the floor on a good-faith sale within 90 days. Placed through a surplus-lines broker; limits and exclusions are in the policy wording, so whether cover holds in a market-wide downturn is the first question to ask. The Financial Times reported in September 2026 that Nvidia is in early talks with insurers on cover for lenders against resale shortfall after a default. | The residual-value leg, for the lender and for the owner who will sell the chip. | Three-year term against a five-year loan. Whether the floor holds when every chip falls at once is in the policy wording, not on the page. No cover beyond the warranty. |
Four exposures, two covered. The owner can sell a week and insure a floor. It cannot move its own cluster, or price its basis.
Lay the owner’s four exposures over the map and the pattern is plain.
The two uncovered exposures share a cause. An owner cannot move a cluster it operates onto a venue because no venue has a standard that says what a deliverable block is: chip, count, interconnect, location, power contract, condition. And nobody prints a price for the middle tenor because there is no standard block to print it for. Andromeda’s certification engine and SF Compute’s SKU definitions are each half of that standard. Neither is public, and neither is shared.
Four pieces, from four different builders, would make a market where any party could manage its exposure. Nobody owns more than one.
The parts exist. They sit with parties who have no reason to join them, and one reason not to.
- A public block standard, from the certifier. Andromeda already benchmarks clusters against enterprise criteria and writes standard contracts on the result. SF Compute already defines SKUs that an order book can trade. Publish the grade, with the attributes that move price (chip, interconnect, location, condition, warranty), and any cluster, including one in an enterprise’s own cage, becomes listable. The certifier earns a fee per grade and gives up nothing it sells.
- An order book that accepts any graded block, from the venue. SF Compute’s book is the right mechanism: whole minutes, partial fills, price-time priority. Open it to graded blocks the venue did not sell, settle in cash rather than credits, and the owner’s surplus has a public price at the tenor it needs. The venue gives up its closed loop and gains every owner as a potential seller.
- A holder that stands between the book and the lender, from the principal. Hyperbolic’s stated trade is exactly this: buy long, sell short, hedge the tail. Done against a public book rather than its own quotes, it becomes the market maker the comment file wants, separate from the benchmark, and its inventory becomes something a lender can see marked daily.
- A cleared tenor in the middle, from the exchange. The futures will settle on on-demand and one-year indexes. A one-to-three-month contract on the graded block print, cleared, is the instrument that lets the holder lay off inventory and the lender hedge collateral at the tenor both actually carry. Nodal, with power beside it, is the natural home; Compute Desk’s one-year rental index is the nearest existing print.
With those four, the owner lists its own block, sells next month’s surplus at a public price, buys cover on the chip’s floor, and borrows against a position a lender can mark. The holder earns the duration spread with a hedge. The lender lends against a print instead of a customer’s name. The exchange gets a contract with a physical layer under it.
The reason it has not happened is the closed loop. Each venue’s resale market is a retention tool: time bought here is sold here, in credits. Opening the book to outside blocks turns a moat into a market. The venue that does it first gives up its moat and gets the print, and in every commodity that has made this transition the print has been worth more than the moat.
It is not a case for one firm doing all four. The comment file is explicit that the party holding positions should not publish the benchmark, and the same logic separates the certifier from the venue. The combination needs four seats and a shared standard, which is closer to how power markets formed than to how a single platform scales.
The lenders have already tranched the chip. The key is the customer’s name, and it costs 350 basis points to be without one, 600 below the syndicated minimum.
Figure 2 is the tranche structure. It was not designed; it emerged from eleven public prints, and it has three tiers with a step between each. A chip with a Meta, Microsoft or Nvidia take-or-pay contract behind it borrows at SOFR plus 2.25 to 3.00 and is rated A3 to Baa2. The same chip with a weaker customer, or a contract shorter than the loan, borrows at SOFR plus 4.00 to 5.50 and is rated Ba2. The same chip with no contract borrows at 9 to 10 percent fixed, and below the syndicated minimum, on chain, at 10 to 15.
Two things in that ladder decide the answer to the tranche question. First, the rating follows the customer, not the borrower and not the chip: CoreWeave is Ba3 as a company and A3 on its Meta loan. Second, the step from an investment-grade contract to no contract is about 350 basis points at syndicated size (the 9.725 percent uncontracted median is SOFR plus 585 at the tracker’s 3.87 percent reference, against SOFR plus 237), and about 600 to 900 below it, where onchain loans to small operators pay 12 to 15 percent. That is the market’s price for residual-value risk with nothing to offset it. Ares said on its October 29 earnings call that it does not know how to underwrite the depreciation curve and would rather lend on buildings. Fitch has opened a consultation on whether chip depreciation belongs in its infrastructure ratings. The agencies have been rating the customer because they can, and leaving the chip alone because they cannot.
Will it look like mortgages?
In form, yes, and soon. The infrastructure around the chip is already securitised, and nine of the sixteen rated deals are building and lease notes rather than chip loans: Moody’s gave its first Aaa to a data-centre securitisation in February 2026, with a senior slice at SOFR plus 120, and Nvidia’s August platform puts chips in special-purpose entities with customer fees servicing bonds sold to institutions. The chips will follow the buildings into pools with a senior, a mezzanine and a first loss.
In substance, no, and the difference is the whole point. A mortgage pool held thousands of loans that were alike in kind and differed in quality. The tranches sorted one risk, borrower default, by order of loss, and the pool diversified it. A chip pool holds assets that differ in kind: generation, interconnect, location, power contract, condition. Those differences do not diversify, because every chip in the pool loses value on the same day, when the next generation ships. Nvidia’s release cadence is a correlated shock to every loan at once, which is the thing a mortgage pool’s structure assumed away. Sorting by order of loss alone, on a pool like that, produces a senior tranche that is senior to nothing on the day it matters.
The defining attributes
So the tranches will be defined by three attributes before seniority is applied, and a buyer should price all three.
The practical consequence is that compute pools will be small and specific before they are large and general. A pool of B200s in Texas on two-year contracts with rated customers, graded and insured to a floor, is priceable today. A pool of mixed chips in mixed buildings with mixed tenants is not, and the agency that rates it will be rating the tenants. The market will build the small pools first, because those are the ones a buyer can check, and the venues in section 01 are what make them checkable: the grade from the certifier, the print from the book, the floor from the insurer, the mark from the holder.
Six signals, in the order they would arrive.
Every page opened on November 4, 2026.
Venues
- Sell unused compute time, SF Compute documentation; Standard Intelligence and Roboflow on the resale market, SF Compute, September 28, 2026; venue facts (founding, Series A, catalogue) from a competitor’s comparison, treated as secondary
- Computable, launch post and discussion
- GPUs are financed for years. AI workloads run for weeks, Hyperbolic, October 26, 2026
- Compute Exchange; hardware marketplace launch, Blockspace via Yahoo Finance, July 17, 2026
- Andromeda; the Paradigm round and valuation from SiliconANGLE, March 18, 2026
- CME Compute Futures, Ironbeam, October 2026
Financing
- Rated GPU financing tracker and GPU loan spreads and interest rates, American Compute. American Compute’s programme is placed by an affiliated surplus-lines broker, Matcha Specialty, so the publisher has a position in the market it tracks; the deal figures are public and were spot-checked against the underlying releases where found.
- GPU residual value insurance, American Compute program terms
- Nvidia courts insurers to underwrite residual value risk, Dealroom citing the Financial Times, September 29, 2026
- Nvidia signs $500B Wall Street GPU bond deal; rating standard unresolved, TechTimes, August 11, 2026
- Ares wary of AI chip financing over GPU depreciation risks, DealStreetAsia on the October 29, 2026 earnings call
- Two Markets, One Asset, Julien Simon, February 20, 2026 (Compass Aaa, CoreWeave CDS, the correlation argument)
- CoreWeave’s debt hits $35bn, Capacity, August 12, 2026 (the Nvidia rent-back)
Assumptions and limits
- The venue map and the gap lanes are Kinetic Alpha’s reading of published terms. Where a venue does not publish a term (Computable’s providers, Compute Exchange’s minimums, Hyperbolic’s settlement), the table says so.
- Spreads in Figure 2 are observed on public deals and exclude fees and original-issue discount. Fixed coupons are converted to a SOFR spread at the tracker’s 3.87% reference for comparison only.
- The uncontracted step: 9.725% fixed less SOFR 3.87% is a spread of 585 basis points, against the investment-grade floating median of 237.5 (about 350 apart) and the other-customer median of 475 (about 110 apart). The onchain median of 12% is SOFR + 813, about 575 above investment grade; the 15% loans are about 875 above.
- The four-part combination in section 03 is a proposal, not a description of anything announced. The venues named have not said they will do it.
- The tranche attributes in section 04 are an argument from the deal ledger and from the structure of the collateral, not from any rated chip pool, because none yet exists.