Energy & compute · Index design
One number for what a rentable chip-hour earns, and two ways to take it apart. The index is built the way a stock index is built — every rentable chip-hour, weighted by what the chip is worth — and cut two ways that add back to the same total: by what the unit is, and by which piece of the money it is. This prototype sets out the design, lets you move every input and watch the index respond, reconstructs two years of history from the anchors we have, and maps what the market already supplies to each piece.
The unit is one chip-hour of rentable capacity: one accelerator or one general-purpose processor, for one hour, at a named operator and site. It is the unit every marketplace quotes and every futures contract is built on — CME Group’s contract is 730 of them, one chip for one month.
Its price has four pieces. Rent is what the hour sells for. Power is electricity at the site’s own meter. Space is the rack, the cooling and the operation of the site. Value is what the chip is worth and how fast that falls. Rent less power less space is the yield; yield plus the change in value is the return on owning the hour.
| Piece | What it is | What moves it | Roughly how big today |
|---|---|---|---|
| Rent | What the hour sells for | Chip supply, contract term, generation turnover, how full the fleets are | $2.63 on the posted-rate index; $3.46 at a US neo-cloud on demand; $1.38 realised by the largest operator across its whole fleet |
| Power | Electricity at the site’s own meter | Gas, weather, congestion into the load pocket, capacity and transmission charges | $0.08 an hour in Texas, $0.18 in Northern Virginia, delivered est.; a tail hour has reached $1.64 |
| Space | Rack, cooling, site operation | Colocation rates by market | $0.34 to $0.48 an hour |
| Value | What the chip is worth and how fast that falls | New generations, resale prints, term curves | About $20,000 for an H100; falling roughly 14 percent a year on recent trend, 40 percent in seventeen months for the generation before it |
Compute cannot be stored. An hour not sold is gone; there is no tank, so “carry” means renting forward and re-letting. The index therefore prices the hour, not the chip in a warehouse.
Rent is quiet; usage is not. Seven-day changes in the rental index run under one percent, the same chip flips between profit and loss on whether it is busy, and token prices — what the chip’s output sells for — have moved five-fold in a day. The biggest input to the return is the one no publisher reads.
Tier dispersion dwarfs everything else. The same H100 hour costs 2.6 times more at a hyperscaler than at a neo-cloud, and neo-cloud rents across seven regions range $2.00 to $3.30. Power, by comparison, is 2 to 7 percent of a US neo-cloud hour — a share that climbs steeply where rents are lower and power dearer, as the regional inputs below will show you.
The index is built the way the S&P 500 is built — one number, decomposable to one share — and that is what makes it useful to anyone who trades: every sub-index is a leg somebody is naturally long, somebody is naturally short, and an instrument either exists for or can be written. Six rules do the work.
Accelerators and general-purpose processors, wherever they sit. A chip that is contracted but idle is still rentable capacity; a chip that is not offered for rent is not in the index. Coverage is a diagnostic beside the index, because nobody knows what share of the cash market any publisher sees.
Rent, less delivered power, less space, gives the yield in dollars per hour. The change in value — the chip’s price times its rate of decline, spread over the hours in a year — is added. The result is the return on owning that hour.
Each segment’s return in dollars is summed across its chips and its hours, and the fleet total is divided by what the fleet is worth. The headline is an annualised return on the capital tied up in rentable chips, and a total-return level accrues at that rate.
Cut one: accelerator or general-purpose; Frontier, Mid or Economy grade; region. Cut two: rent, power, space, value. Each cut sums to the headline exactly, so a position in any sub-index is a position in a known share of the whole.
Frontier is the generation that sets the price per reference token; Mid is the one before it; Economy the one before that. A generation steps down a grade the day the next one crosses 10 percent of rentable capacity. Rubin is shipping; when it crosses, Hopper becomes Economy and Ampere leaves the index — a dated event, 2027 on our estimate.
The posted-rate family — Silicon Data, which the CME Group contracts settle on — reads posted rates, and no publisher in either family uses utilisation, so every hour counts as rented. The index reads posted rates for the same reason: so that it can settle on that family. The share of posted rent an operator actually earns — $1.38 against $2.63 for the largest — is carried as a diagnostic series and as an input you can move below, because it is the number the physical trade lives or dies on.
Two things about the design deserve saying plainly. First, the headline at published inputs is an upper bound: posted rates sit above what operators realise, and every hour is counted as rented. That is not a flaw to hide but the point of the second cut — the gap between the index and an operator’s realised return is the fill and utilisation risk that dominates the physical trade, and the design carries it as a named input rather than burying it. Second, the power leg is priced at the site’s meter, not at the hub. The hub price misses congestion into the load pocket, losses and the fixed charges that rise per rented hour as utilisation falls; the difference is small in a normal hour and has been two-thirds of an hour’s rent in the worst one.
Every input the index reads is a slider. Defaults are the document’s September 14, 2026 figures where it gives them, values derived from those figures where marked derived, and our working estimates where it does not est.. The headline, both cuts and the sensitivity table recompute as you move anything, and section 04’s history recomputes with you, so the inputs that have no published history — space, realisation, the power usage effectiveness of the site — are held at whatever you set here.
| Move | Headline | Change | Return per chip-hour |
|---|
Each row moves one input from its current setting and holds everything else. Changes are in percentage points of annual return on fleet value. Power usage effectiveness is the site’s total draw divided by the chips’ own draw.
| Grade | Region | Chips, millions | Rent, $ an hour | Power | Space | Yield | Value change | Return | Return on value | Share of headline |
|---|
No publisher has more than two years of daily history, and none publishes the index defined here. What we do have are anchors: the H100 index fell 40 percent between September 2024 and September 2025 and sits at $2.63 today; delivered power in Northern Virginia rose 83 percent over a recent year — we place it September 2025 to September 2026 — to about $0.18 a chip-hour (an estimate in the document too); an H100 is worth about $20,000 and has been losing roughly 14 percent a year; the generation before it lost 40 percent in seventeen months; a B200 rents at about $4,200 per contract of 730 chip-hours, or about $5.75 an hour, and on our assumption became rentable at scale from March 2025. Between those anchors the paths below are interpolated, and where a series has only one anchor it is extrapolated along an assumed path; both are marked as such. Space and the general-purpose inputs have no series in the document and are held near their current levels. The exercise is not a backtest — it is a demonstration of how the legs move the headline, and every path rescales to whatever you set in the dashboard.
Rent is stacked above zero; power, space and value are stacked below it; the line is the sum. At the default inputs the headline falls every month to September 2025 as the H100’s rent drops 40 percent, and recovers through 2026 for two reasons of roughly equal size: Blackwell’s share of the rentable fleet ramps up at a higher rent per dollar of value than the fleet it joins, and the H100 loses value faster than it loses rent, which raises the return on a shrinking base — the value leg records the loss, the rent leg records the yield on what is left.
Each leg’s bar is its contribution at the end less its contribution at the start, holding the fleet’s composition at the end; the composition bar is what changing the fleet (Blackwell entering) did to the legs on their own. The five bars sum to the change in the headline.
A benchmark tells you the price. A certificate tells you the chip is what it claims. A telemetry feed tells you it ran. A colocation survey tells you what the rack costs. None of these moves risk from one balance sheet to another, and the index needs all of them before any instrument can settle on it. The table below takes each element of the design and asks two questions: what exists today that satisfies it, and whether anything lets a holder transfer the risk in it. Most of the market is in the first column. The instruments — the futures, the yes-or-no contracts, the perpetuals, the swaps, the insurers’ cover — sit on two legs of four, and on the rent leg for two grades only.
| Index element | What exists that satisfies it | What it leaves out | Risk transfer? |
|---|---|---|---|
| Rent leg · Mid and Frontier grade H100, B200 | Two index families built differently: Silicon Data reads posted rates across many platforms and publishes a term curve to 36 months, a forward curve since April and residual values, with an independent calculation agent appointed this month; Ornn and Compute Desk read executed deals (Compute Desk’s from verified private transactions, distributed on Bloomberg and Refinitiv). SemiAnalysis surveys prices out to five years. | Utilisation — no publisher uses it. The families print apart and the gap is structural; methodology changes have moved an index 3 to 7 percent per event and one history was restated 35 to 40 percent. Two years of daily history at best. | Yes, this leg CME Group’s H100 and B200 futures from October 5 (pending review) on Silicon Data; Kalshi and Polymarket US yes-or-no contracts, Architect’s offshore perpetuals and Nodal’s announced futures on the executed-deal family; bilateral swaps on either index from any willing balance sheet. |
| Rent leg · Economy grade A100 | Silicon Data publishes posted-rate indices for every major chip, and its term curve prices the A100 in contango ($1.37 for 36 months against $1.61 spot); Kalshi lists rental-price contracts for five Nvidia chips, which are not named. | No listed future: CME Group lists H100 and B200 only. Whether the executed-deal family covers the A100 is not public. One publisher’s history has been restated by 35 to 40 percent; which series is not public. | Partly Prediction-market contracts at small size, if the A100 is among Kalshi’s five chips; bilateral swaps. |
| Rent leg · General-purpose processor | Posted rates on the cloud spot markets and bare-metal hosts; server processor prices (up 10 to 20 percent since March, six-month lead times) as the value anchor. | No published index series with history in chip-hour terms; nothing listed. Agentic workloads are moving work to the processor, which is why the design carries the sub-index. | No A watch item. |
| Rent leg · tier and region | Hyperscaler sub-series and neo-cloud series for the same chip (a 2.6-times gap); country series (US $3.46, Germany $1.34, Japan $1.88 at neo-clouds). | The hyperscaler series moves in steps and reads like a list price; ex-US panels are thin — one series printed flat for seven weeks then jumped. The dashboard applies the country prints as regional factors for want of anything better. | No Bilateral only: a hyperscaler commitment on one side, a marketplace on the other. |
| Power leg | Hub power futures and hourly contracts; transmission-congestion contracts between hub and bus; Nodal Exchange, which already holds most North American power open interest and has announced compute futures beside it for 2026; capacity and transmission charges in the utility tariff. | The index needs the price at the site’s own meter. The hub misses congestion into the load pocket, losses and fixed charges, and no rental index reads the meter. A three-year study finds no stable co-movement between power and compute, so a margin offset between the two will credit diversification only. | Yes, at the hub The deepest cleared market of any leg — for the hub component. The basis from hub to bus is a structured position. |
| Space leg | Colocation rates by market: rack, cooling and site operation at $0.34 to $0.48 a chip-hour. | No published index in chip-hour terms; no history the document can cite; the design holds it near-constant. | No A lease, not an instrument. |
| Value leg | Silicon Data’s residual values; resale prints, including Compute Exchange’s used and refurbished hardware market; the compute insurers’ cover prices for residual value; the depreciation the term curve implies. | The sources disagree: the term curve expects new chips to get cheaper to rent and old ones to hold, while residual values have done the reverse — the H100 losing roughly 14 percent a year against 40 percent in seventeen months for the generation before it. Resale prints are sparse. Performance can be certified per chip (SiliconMark); value cannot. | Structured Residual-value, outage and credit cover from the compute insurers; a lender’s collateral desk on the other side. No screen. |
| Weights · rentable capacity by grade and region | The fleet loans and securitisations that fund the operators; marketplace listings; the exchange-for-physical network’s adjustment tables for chip, memory and location — the first public basis grid this market has had — once the network launches. | Nobody knows what share of the cash market any publisher sees — the regulator has asked in writing. Our fleet counts are estimates. | Not applicable Weights are measured, not traded. |
| Grade rules · the regrade calendar | Shipment and capacity data for the next generation; SemiAnalysis cluster quality grades; per-chip performance certificates; a reference-token conversion between generations. | The 10-percent threshold is our rule; nobody publishes the share of rentable capacity by generation on a schedule. | At listing The grade-against-grade spread (B200 against H100 at CME Group; Blackwell against Hopper at Nodal). |
| Realisation · posted against realised, and usage | Performance certificates certify performance; telemetry vendors (Clockwork, Hydra’s Brokkr, Fluidstack’s Lighthouse and the standard exporters) record fleet telemetry; a proposed usage-record standard reconciles telemetry and invoice with facility power as the third check. | Usage — whether the chip was actually busy — is recorded by nobody independent, and it is the multiplier the index does not see: $1.38 realised against $2.63 posted for the largest operator. | Structured, later A swap on metered hours or a floor written against a usage record — first half of 2027 on our estimate, and never before there is a meter. |
| Term structure · the expected path of rent | Silicon Data’s term curve to 36 months and forward curve; Ornn’s forward curves on its own index; the prediction-market ladders; the registered funds’ disclosed roll calendars; the futures strip once listed. | The H100 term rate ($2.29 for 36 months against $2.63 spot) expects new chips to get cheaper to rent and the A100 ($1.37 against $1.61) to hold — the opposite of what residual values have done — and no instrument yet arbitrates the two. Kalshi’s whole compute notional through July 27 was $4.4 million; Polymarket US’s about $285,000. | At listing The strip, 36 months out; today, the ladders as a sensor. |
Sources are the discussion document’s tables on where compute trades and what holds it up, and its spread table for the term, regional, regrade and processor facts, as of September 14, 2026. “Structured” means a counterparty and a contract are needed; “at listing” means the CME Group or Nodal contract is needed. Where a venue’s terms are not yet public the table says so.
Read down the last column and the shape of the market is clear. The rent leg is served by two index families and, from October, by a cleared future — but for two grades, on one family, at posted rates. The power leg has the deepest market of all, at the hub rather than the meter. The value leg has several answers that do not agree and an insurer willing to write cover on one of them. Space has a price and no index; the general-purpose processor has a price and no publisher. And realisation — the one input that separates the index from what an operator actually earns — has certificates for performance, telemetry for the fleet, and no independent record of usage at all. The index can be assembled today from published inputs for its rent, power and value legs, with the caveats above; what it cannot yet be assembled from is the number the physical trade lives or dies on.
That is why the design carries realisation as a named input rather than an assumption, prices power at the meter rather than the hub, and keeps a general-purpose sub-index before there is anything to settle it on. An index that only reflects what can be hedged today would be a rental index with a different name; the point of this one is to put a number beside every piece of the money so that the pieces without an instrument are visible, sized and, in time, written.
An operator is long every leg at once and short none of them by choice. A lender against a fleet is long the value leg and cares about nothing else until it moves. A lab renting capacity is short the rent leg. A site owner is long space and short power. A hyperscaler is long the tier gap. Each of them has a natural counterparty on the other side of one sub-index, and the design’s job is to make that sub-index a number both can settle on. The spreads between the pieces — term against spot, physical against listed, one index family against the other, power against compute, posted against realised — are the subject of a companion piece; here it is enough to say that each is a pair of sub-indices from the trees above.
Three things the design needs next, in order. Power priced at the site’s own bus, so the power leg is the operator’s actual bill and not the hub’s approximation of it. An independent usage record, so realisation becomes a series rather than a slider. And standard paper with collateral both ways, so that a sub-index can be written into a contract without negotiating every block from scratch. None of the three is an index problem in the narrow sense; all three decide whether the index is a benchmark or a description.