The hour is priced. The unit is not.
Three exchanges now settle futures on the price of renting a GPU for an hour. Nobody publishes what a lender gets back when it has to sell one. This report says who prices what, and why the gap sits exactly where a credit desk stands.
Compiled from public sources, 2026-10-06 to 2026-10-28. Claims link to the source. Estimates are marked.
The short version
- CME settles its compute futures on an on demand rental index built from about 10,000 daily neocloud listings across 40 or more providers; the index carries a Bloomberg ticker and its administrator is part owned by CME Ventures. CNBC.
- Two rivals build from executed contracts instead: Ornn with ICE, and Compute Desk with Nodal. ICE notice, Nodal release.
- The credit market prices the offtaker, not the hardware. CoreWeave's DDTL 5.5 credit agreement has no appraisal, loan to value or borrowing base language. Datagravity.
- Residual value is modelled everywhere, observed nowhere. The index administrators' residual products are discounted cash flows off their own rental curves with a global operating cost ratio. None publishes a realised sale.
- The open space is the recovery number: realised auction and ITAD prices, by lot, by channel, by days to sale. The firm that runs the liquidations gets the data. That is a credit and custody business, not a scraping business.
Three stacks, one hour
| Exchange | Index | Built from | Weak point for a lender |
|---|---|---|---|
| CME (NYMEX) | H100, B200 on demand | on demand listings, proprietary "market relevance" weights | most GPU revenue is committed contracts, not on demand |
| ICE Futures US | Ornn OCPI | transactions; also settles Kalshi and Architect perps | transaction volume disclosed by the administrator, not audited |
| Nodal (EEX) | Compute Desk | private OTC transactions, described as IOSCO aligned | thin, early, and paired with power on the same venue |
All three price one thing: a GPU hour of a named SKU, delivered on a cloud, for a short tenor. None prices the board. A lender who holds a lien on 512 boards in a Reno colo needs the second number.
What a listing index measures
The Safewaterline on demand rate for the H100 was about $2.90 an hour in October 2026, a median of public neocloud listings, against hyperscaler list prices above $7. The gap between the two lines is the point: the listing median is the marginal hour, not the contracted hour. One-year contract rates in 2026 sat at $1.70 to $2.35. SemiAnalysis.
For a lender the on demand rate is still the right input. It is the rate a unit earns after the contract ends, which is the rate that sets the tail of the residual curve. It is not the rate that sets the value of a unit inside a contract. The Safewaterline curve uses it for the tail and leaves the contract to the desk.
What the credit market prices
The large facilities are priced on the offtaker's credit and the tenor of its contract. CoreWeave's DDTL series went from about 11 to 14% all in (2023) to SOFR plus 225 basis points at DDTL 4.0, rated A3, then widened to SOFR plus 450 and 550 for DDTL 5.0 and 5.5. The spreads track who the customer is, not what the boards are worth. Datagravity.
Below that tier sits a long tail: 50 or more lenders in USD.AI style and second tier facilities, where the hardware is the only credit. These lenders need an advance table by model and age, a liquidation haircut by channel, and a monitor that tells them the boards are still in the rack. The register on this portal shows that most filed agreements do not state an advance rate at all.
Insurers have begun to fill the gap from the other side. Barkr sells warrantied forward valuations backed by Munich Re, covering up to about 85% of an assigned value, and has covered more than $200M. The Insurer. An insurer that writes residual cover needs the same curve a lender needs, and has the same lack of comps.
Where recovery data comes from
Three sources exist, none complete.
- Broker and ITAD prints. American Compute's Rack Report indexes realised resale from brokers and ITADs monthly, with liquidity grades. It is the closest thing to a comp series. Rack Report.
- Bankruptcy sales. Public but rare. The Core Scientific restructuring produced some GPU lot prices; the amounts are small relative to the market.
- The lender's own liquidations. The only source with channel, days to sale and condition recorded. It belongs to whoever runs the sale.
The third source is the one a credit desk generates by doing its job. It is also the one that cannot be scraped, which is why it can be a benchmark. The firewall on this site exists so that it can be published: a comp is generalised and consented before it leaves the desk.
What the desk publishes, and what it does not
| Published | Not published |
|---|---|
| Fair value, orderly and forced liquidation value per model and age | An exchange settled rental index. The on demand rate here is a listing median, published with every listing behind it. |
| An advance per unit, as a share of orderly value | A forward curve. Three administrators already compete on it. |
| A risk score per model and per fleet type, with reasons | Token prices, benchmark tests, delivered compute telemetry. |
| The register of filed facilities, with quotes | Anything about a Safewaterline borrower. |
Sources
Exchanges: CNBC on CME · ICE and Ornn · Nodal and Compute Desk. Credit: Datagravity, how the GPU became collateral · The Insurer on residual value insurance · American Compute Rack Report. Market structure: SemiAnalysis.