Compute Markets Futures and Indices
A primer on how compute is rented, indexed, and traded
Compute, or computing power, means the ability to process data. This paper focuses on two ways to buy it: renting graphics processing unit (GPU) capacity, typically measured in GPU-hours and priced per GPU-hour, and paying for usage or output, such as the tokens billed by an artificial intelligence (AI) inference service. In 2026, several US exchanges announced futures tied to GPU rental rates.1,2,3 As of November 12, 2026, none of these futures had started trading. The CFTC's review of the first contracts filed for approval runs to January 8, 2027, so a first launch could come early in 2027.4
Anyone who knows an older commodity market, and power in particular, will find much of this familiar. Compute is sold as capacity over time. Some of it is firm and some is interruptible. Some contracts are take-or-pay. Prices are collected into indices, and futures settle financially against those indices. But the vocabulary is different, a few familiar words mean something else, and no central operator publishes a single market price. This paper is a map for anyone new to the compute market. It defines each term as it appears and assumes no background in power. For readers who do have that background, Section 3 translates the main ideas into power market terms.
The paper starts with physical rentals, then moves to price indices, derivatives, and financing. Section 1 explains how compute is rented and defines the four rental structures used throughout. Section 2 lays out the market in three layers: physical supply, price indices, and derivatives. Section 3 is the translation for power readers. Section 4 looks at what the indices actually measure and at the futures built on them. Section 5 asks what the standard unit of compute could be. Section 6 covers perpetual contracts. Section 7 discusses which tools fit a buyer, a seller, a trader, and a lender. Section 8 looks ahead to one possible direction, a system operator for compute. One thread runs through all of it: there is no single compute price, and the structured compute market is at a nascent stage that is currently dominated by the GPU-hour construct.
[SECTION 01]
1. How compute capacity is rented
Since the mid-2000s, businesses have been able to rent compute capacity from cloud providers instead of buying and running the equipment themselves.5 The sellers fall into two groups. Hyperscalers are the large general cloud providers, such as Amazon Web Services (AWS), Google Cloud, Microsoft Azure, and Oracle. Neoclouds are newer specialists, such as CoreWeave, Lambda, and Nebius, whose main business is renting GPUs.
Servers use central processing units (CPUs) for general-purpose computing. AI workloads also use accelerators, specialized chips that perform the heavy arithmetic faster than CPUs. GPUs are the most common accelerators, but others include custom chips developed by cloud providers, such as Google's tensor processing units (TPUs). This paper focuses on GPUs because the rental market, its indices, and the proposed futures discussed here are largely quoted in GPU-hours, and most of the chips tracked by the benchmarks in Table 5 are NVIDIA GPUs. Where a point applies more broadly, the paper uses the term accelerator.
The unit of sale is time on a machine. For AI work, it is the GPU-hour: one GPU of a named model, such as an NVIDIA H100 or B200, for one hour. A GPU-hour measures capacity, not output. The customer pays for the hour whether the chip runs flat out or sits half idle, much as a capacity payment in power is owed whether or not the plant is dispatched.
The way that time is sold has changed in steps, and each step added commitment. AWS launched its Elastic Compute Cloud (EC2) in 2006 with one way to buy: rent a server by the hour, with no commitment, and stop whenever you like. In 2009, it added Reserved Instances, which gave a lower rate in return for a one- or three-year commitment. In 2010, it began renting servers fitted with NVIDIA GPUs. In 2023, with demand for GPUs running ahead of supply, it introduced Capacity Blocks: GPU capacity reserved for a fixed future window, paid for in full in advance, and not cancellable.6,7,8,9,10 Table 1 sets out the sequence, including the price indices and derivatives that arrived in 2025 and 2026.
Table 1. How the compute capacity market was built
| Year | What was added | Why it matters |
|---|---|---|
| 2006 | On-demand rental. AWS launches EC2: servers rented by the hour with no commitment.6 | Compute becomes a service priced by time, not by useful output. |
| 2009 | Commitment. AWS adds Reserved Instances: a lower rate for a one- or three-year commitment.7 | Price starts to depend on the term the buyer commits to. |
| 2010 | GPUs. AWS begins renting servers fitted with NVIDIA GPUs.8 | The hardware that AI later depends on becomes rentable. |
| 2023 | Prepaid GPU reservations. AWS adds Capacity Blocks.9 | Firm, non-cancellable GPU capacity appears as GPUs become scarce. |
| 2025 | Price indices. Silicon Data launches a daily GPU rental index, and others follow.11 | Rental prices become visible to people outside each deal. |
| 2026 | Derivatives. CME, ICE, and Nodal announce GPU rental futures, and the CFTC asks for public comment.1,12 | Price risk can, in principle, be transferred without renting a machine. |
Today's rental products are easiest to sort with three questions borrowed from older commodity markets. Is the capacity firm, meaning the provider guarantees it will be there when the customer wants it? Is it interruptible, meaning the provider can take it back while the customer is using it? And is it take-or-pay, meaning the customer owes payment for a committed term whether or not it uses the capacity? The answers give four common structures. Table 2 sets them out with their tenors and examples, and the definitions below the table are used in the rest of the paper. The column on who offers each structure shows who sells it, not market share.
Table 2. Compute capacity rental structures: firmness, interruptibility, and take-or-pay
| Structure | Firm? | Interruptible? | Payment obligation | Tenor | Who offers | Example products |
|---|---|---|---|---|---|---|
| On-demand, pay as you go | No; best effort when requested | No | Pay only while running. Not take-or-pay.13,14 | None; billed by the second, minute, or hour, depending on the provider | Hyperscalers and neoclouds | Lambda On-Demand Cloud instances; Nebius pay as you go GPU virtual machines15,16 |
| Spot or preemptible | No | Yes, at short notice | Pay only while running, at a variable discount. Not take-or-pay.14 | None; billed by the second, minute, or hour, depending on the provider | Hyperscalers and some neoclouds | Google Cloud Spot VMs; Vast.ai interruptible instances17,18 |
| Capacity reservation, cancellable | Yes | No | Pay for the reserved capacity whether used or not, until cancelled. No committed term, so not take-or-pay.19 | No minimum term for this structure; billed while held | Mainly hyperscalers | Microsoft Azure On-demand capacity reservations; Oracle Cloud Infrastructure capacity reservations20,21 |
| Capacity reservation, non-cancellable | Yes | No | Pay for the full term whether used or not, often in advance. Take-or-pay.10 | Fixed term, from days to several years22,23 | Hyperscalers, neoclouds, and brokers | AWS EC2 Capacity Blocks; CoreWeave multiyear reserved capacity contracts10,23 |
Firm: the provider guarantees that the capacity will be available when the customer wants to use it.
Interruptible: the provider may take the capacity back while the customer is using it.
Take-or-pay: the customer owes payment for the contracted capacity over a committed term whether or not it uses it, and the provider must make that capacity available.
On-demand: capacity rented with no commitment and paid for only while it runs. It is not normally subject to preemption, but it is not firm at the time of request: the provider supplies it on a best-effort basis and may have none to offer when the customer asks. In this paper, On-demand always means this structure. A few providers also use the words in the names of cancellable reservations, such as Microsoft Azure's On-demand capacity reservation, where they refer only to the billing rate.13,20
Spot or preemptible: spare capacity sold at a discount that the provider can reclaim at short notice. In compute, spot does not mean the current market price.17
Cancellable capacity reservation: capacity the provider holds for the customer, who pays for it whether used or not but can cancel at any time. It is firm while held, but with no committed term it is not take-or-pay.
Non-cancellable capacity reservation: capacity reserved for a fixed term and paid for in full whether used or not. It is both firm and take-or-pay.
Table 2 illustrates a clear direction. The market began with capacity that was neither firm nor committed, and as GPUs became scarce, it moved toward capacity that is both. Some of the largest publicly described contracts are multiyear, non-cancellable reservations. One large neocloud, for example, describes its customer contracts as take-or-pay commitments for reserved capacity with no right to terminate for convenience.23
[SECTION 02]
2. The compute market in three layers
The market now has three layers, each built on the one below (Table 3). The first is physical supply: providers rent machines under the structures in Table 2. The second is price indices: firms that collect rental prices and publish a number, usually daily, for each GPU model. The third is derivatives: financial contracts whose payoff depends on one of those indices. In this paper, a derivative means any such contract, whether a future, a swap, or an option. It moves price risk from one party to another. By itself, it does not give anyone a machine. Some firms work in more than one layer: some index providers also broker or clear physical capacity. There are many other compute brokers and marketplaces that do not publish an index, and this paper does not survey them.
The same GPU names appear in all three layers. An H100 rental, an H100 index, and an H100 future are three different things. The rental gives access to a chip. The index is a published price. The future is a way to fix, or to bet on, where that published price will be. Holding a cash-settled future does not secure a machine.
Table 3. The compute market in three layers, as of November 12, 2026
| Layer | What it is | Who is in it | Closest power market counterpart |
|---|---|---|---|
| 1. Physical supply (capacity only) | Rental of machines (capacity) under the four structures in Table 2. Offers differ in more than price: a customer may get a whole physical server to itself (bare metal) or a share of one managed by software (a virtual machine), with different memory, networking, and support. | Hyperscalers: AWS, Google Cloud, Microsoft Azure, Oracle.14,20,21,24 Neoclouds: CoreWeave, Lambda, Nebius, and others. Brokers and bilateral contracts. | Generation and the physical contracts that sell it: capacity agreements |
| ▼ Indices observe the prices paid or offered in layer 1 | |||
| 2. Price indices (capacity and usage/output) | Published prices per GPU-hour (capacity) for a named GPU model are the most common. Other benchmarks track a GPU basket or price per unit of energy. An index is not a contract and gives no access to a machine. Table 5 compares selected benchmarks. Token prices reflect billed usage and are not standardized across models or billing categories (input, cached input, output). | GPU rental indices: Computable, Compute Desk, Compute Screener, Ornn, SemiAnalysis, Setara, Shanghai Metals Market (SMM), Silicon Data, and The Compute Indices. GridCompute tracks marketplace asking prices.25 LO:TECH's Super Intelligence Compute Index (SICI) tracks a basket of GPU rental prices.26 Energy-normalized index: NativX COIL.27 Token prices: indices from Compute Screener, Ornn, and Silicon Data, and provider price lists from OpenRouter.28,29,30,31 | Zonal or regional capacity price and the assessments published by ISOs or RTOs. Hub energy price and index assessments published by reporting agencies. |
| ▼ The rental derivatives shown here reference capacity benchmarks | |||
| 3. Derivatives (capacity only) | Financial contracts referencing capacity prices. The futures announced by CME, ICE, and Nodal are cash-settled: no GPU-hours change hands. One proposed venue plans a link to physical delivery (Section 4). | Futures, announced but not yet trading: CME with Silicon Data, ICE with Ornn, ICE with NativX, and Nodal with Compute Desk.1,2,3,27 Perpetuals: Architect AX (Bermuda)32,33, Injective (onchain)34, and Bitget (premarket).35 Architect separately plans US futures on Compute Desk indices.36 Event contracts: Kalshi launched GPU rental price event contracts in 2026; current availability not verified.37 Bilateral: swaps and options; FalconX reported executing an Ornn H100 swap.38 | Financially settled power futures, swaps, and options |
The CFTC's request for public comment raised a question any trader will recognize: can the indices draw on enough reliable market data to resist manipulation?12 Table 3 lists announced futures and other derivatives; Table 5 compares selected rental benchmarks. Section 4 explains how the futures use those benchmarks.
[SECTION 03]
3. A power trader's translation
Most of the compute market can be mapped onto ideas a power trader already uses. Table 4 gives the translation and says how far each analogy can be trusted. Three of these comparisons need a fuller explanation, which follows the table.
Table 4. Power market concepts and their compute counterparts
| Power concept | Compute counterpart | How well the analogy holds |
|---|---|---|
| Capacity: megawatts made available, paid for through a capacity payment | The GPU-hour: one accelerator made available for an hour | Holds well. Both pay for availability, not for what is produced (i.e., energy volume). An idle GPU-hour, like an undispatched megawatt, is still paid for and cannot be stored for later. |
| Energy: megawatt-hours delivered | Tokens: the units of text an AI model reads and writes, used to bill AI services | Holds in part. Both measure variable usage or output. But a megawatt-hour is the same everywhere, while token prices differ by provider, model, and billing category. |
| Firm service and interruptible (non-firm) service, as with firm and non-firm transmission | Firm: capacity reservations. Interruptible: spot or preemptible capacity | Holds well. A capacity reservation provides contractual assurance, subject to its terms. Spot or preemptible capacity can be taken back at short notice. On-demand sits in between: it is not interrupted once running, but it is not guaranteed to be available. |
| Take-or-pay contracts, such as tolls and long-term capacity deals | Non-cancellable capacity reservations | Holds well. The buyer pays for the term whether or not it uses the capacity. |
| The spot market: day-ahead and real-time prices set by a system operator | No counterpart. The nearest thing is the On-demand price each provider posts | Breaks. There is no central market and no single clearing price. In compute, spot means interruptible capacity, not the market price. |
| Location: prices differ by hub and node | GPU model first, region second | Holds loosely. An H100 hour and a B200 hour are separate products with separate prices, as two hubs are. Regional compute hubs have yet to develop (Section 4). |
| Hub indices and price reporting agency assessments | GPU rental price indices | Holds, with a warning. Compute indices come from private firms using different methods (Table 5). |
| Financially settled futures on a hub price | Cash-settled futures on a GPU rental index | Holds well. Both pay the difference between the traded price and the index. Neither delivers anything physical. |
| Basis risk between the hub and your node | The gap between the index and your own invoice | Holds well. The hedge works only as far as the index matches your GPU model, region, and rental structure. |
First, the GPU-hour is a capacity product and billed inference usage, measured in tokens, is closer to an energy product. A GPU-hour is time on a chip, and it says nothing about how much useful work gets done. A token is a small unit of text, and AI services are commonly billed by the number of tokens an AI model reads and writes. A token price therefore measures the use of compute, much as a megawatt-hour measures the use of generation. The difference is that a megawatt-hour is the same whoever produces it. Token prices depend on the provider, model, and billing category. Input tokens are usage, not output, and the published token price need not identify the hardware running the model. That makes tokens much harder to standardize and to index than energy, a point Section 5 returns to. Token price lists and indices exist, and Table 3 notes them, but they are not the focus of this paper.31,39
Second, the word “spot” has different meanings. A power trader hears spot and thinks of the real-time price. In compute, spot or preemptible capacity is the interruptible product in Table 2, sold at a discount because it can be taken back. The compute price that behaves most like a power spot price, the price to get a machine now with no commitment, is the On-demand price.
Third, and most important, compute has no system operator for either capacity or usage/output. Organized US power markets have a grid operator, centralized dispatch, and published energy prices at pricing nodes. Some power system operators, such as PJM, ISO-NE, NYISO, MISO, and CAISO, also administer capacity markets or resource adequacy programs, some through auctions and some through bilateral contracts. Compute has many separate capacity contracts between providers and customers, most of them private. Someone has to collect those prices and decide how to turn them into one number, and different index providers do it differently. That is why the next section will address how the indices are built before turning to the futures that settle against them.
[SECTION 04]
4. Indices and the futures that settle against them
A price index is a published number for a defined product: a GPU model, a region, a rental structure, and a time period, calculated by a stated method. The most important design choice is the kind of price that goes in. A transacted price is what a buyer actually paid. An offer price is what a seller is asking, whether on a public price list or as a live quote that a buyer can accept. Power and gas traders know the same distinction from price reporting agencies, which build their assessments from reported trades, bids, and offers.
Table 5 compares selected compute rental benchmarks on that basis. They are listed alphabetically, and they fall into three groups. Ornn uses transacted prices only. SMM's mainland China series is based mainly on transacted prices that SMM collects, as SMM has confirmed to me in conversation and by email. SMM publishes separate spot assessments, listing prices, and a composite. Its public overview refers to transaction data and representative listings but does not specify their relative weights.40 NativX describes its benchmark as a volume-weighted fix but does not publicly specify the full calculation.41 Compute Desk and Silicon Data blend transacted prices with offers. Computable, Compute Screener, GridCompute, LO:TECH's SICI, SemiAnalysis, Setara, The Compute Indices, and SMM's Southeast Asia series use offer prices. GridCompute reports the median asking price for a given GPU on Vast.ai; SICI combines offers from several marketplaces into a single index quoted in points. Neither measures completed trades.25,26 The index prints in Table 5 should not be confused with a live bid or offer.
Table 5. Selected compute rental benchmarks, as of November 12, 2026
| Provider | Price type and rental structure | GPU models covered | Geography | Linked derivatives |
|---|---|---|---|---|
| Computable42,43 | Offer prices from published price lists; On-demand only | H100, H200, B200, B300 | Provider panel; region not stated in the sources reviewed | None identified |
| Compute Desk44 | Transacted prices blended with neocloud list (offer) prices; one-year non-cancellable reservations | H100, H200, B200, B300 | United States | Proposed Nodal futures; proposed futures on Architect's US exchange |
| Compute Screener30 | Offer prices from price lists and marketplace quotes, adjusted to one standard eight-GPU US server; On-demand | H100, H200, A100, B200, B300, and several smaller GPUs | United States reference; prices from other regions are adjusted to it | None identified |
| GridCompute25 | Hourly median of Vast.ai On-demand asking prices for each GPU; cloud rate cards tracked separately | H100, H200, A100, B200, B300, L40S, RTX variants, and others | Vast.ai listings; no fixed region stated | None identified |
| LO:TECH SICI (Injective)26 | Marketplace asking prices for On-demand and spot capacity. Fixed weights; rebased to 100 and averaged over seven days | H100 SXM 80GB, B200, H200, A100 SXM 80GB, RTX 4090 | Compute marketplaces; no fixed region stated | Injective perpetual launched October 202634 |
| NativX COIL27,41 | Volume-weighted fix, according to NativX; eligible inputs and calculation details not verified from public materials | Not tied to a GPU model; priced per unit of energy used by the computing equipment | Not stated in public materials | Proposed ICE futures |
| Ornn45 | Transacted prices only; On-demand | H100, H200, B200, A100, RTX PRO 6000 | Global contributors, tagged by continent | Proposed ICE futures; Architect AX perpetual |
| SemiAnalysis46 | Offer prices collected from providers and marketplaces; the H100 index adds analyst surveys of contract prices; mixes On-demand and term reservations | H100, A100, B200, one composite index each | Mixed cloud sources; no fixed region | None identified |
| Setara47,48 | Offer prices from posted price lists, lowest price per provider, equally weighted; On-demand only | H100, H200, A100, B300, L40S, L40, A6000 | Global series per GPU model; US East and US West series for A100 | None identified; published as a pre-production series |
| Shanghai Metals Market (SMM)49,50 | Mainland China: mainly transacted prices, with some public listings, for terms of days to months. Southeast Asia: retail offer prices for On-demand, spot or preemptible, and one- and three-year commitments | Mainland China: dozens of GPU models, including H100, H20, A100, RTX 4090, and Huawei Ascend 910B. Southeast Asia: mainly H100 and H200 | Mainland China; separate series for Singapore, Malaysia, and Indonesia | None identified |
| Silicon Data51,52 | Transacted prices plus verified executable offers; spot, On-demand, and reservation records are adjusted to an On-demand basis | H100, H200, A100, B200, B300, MI300X | Global sources; some separate regional and hyperscaler series | Proposed CME futures use the H100 and B200 neocloud indices53 |
| The Compute Indices54 | Offer prices, with live marketplace quotes weighted twice as heavily as list prices; On-demand, with separate spot series | H100 headline; H200, B200, B300, and A100 where data allow | European Union and European Economic Area headline; US and global H100 series | None identified |
The indices are generally not like for like, even for one GPU model, and the price type is the first reason. It explains how two indices for the same chip can disagree without either being wrong. Rental structure is the second reason. On-demand capacity, spot capacity, and reservations of different terms are different products with different prices, so an index built on one structure will not match an index built on another. Each index also draws its own sample: which transactions or offers it collects, from which providers, in which places, and at what times. A gap between two indices may therefore reflect what each one measures, not a disagreement about the market. Before relying on an index, check three things: which price type it uses, which rental structure it prices, and whether both match your own exposure.
Geography is another design choice, and it involves a trade-off. An index that pools prices from several regions (e.g., the United States, Europe, the Middle East, and Asia) into one global number is a useful gauge of the general trend. But it may be too broad for anyone to trade and settle against, because no participant's actual cost follows a world average: power prices, chip supply, and demand all differ by region. A regional or local index that reflects the conditions of its own market is likely to make a better contract. Power markets show the pattern. The contracts that trade in large volume are tied to regional hubs, such as the Western Hub in PJM, the grid operator for much of the eastern United States, and the North Hub in ERCOT, the operator in Texas. Around each hub are many individual pricing points, called nodes, and traders manage the price difference between a node and its hub with instruments such as financial transmission rights (FTRs) and congestion revenue rights (CRRs). Compute could follow the same path. Electricity is a large part of the cost of running a data center, and data centers cluster where power is available, so one reasonable place to start would be to define compute regions around the liquid power hubs. This is my suggestion, not a description of any index or contract offered today.
Governance is a separate question from price type. The International Organization of Securities Commissions (IOSCO) publishes principles for how a financial benchmark should be designed, calculated, and overseen.55 The compute indices are young, and their providers are at different stages of building that kind of governance. Each provider sets out its own position in its methodology, including whether the index is intended for settling contracts, and a user should read that alongside the price type. A provider that also arranges physical deals sees transacted prices directly, which can strengthen its index, but it must also manage the potential conflict between publishing a benchmark and earning fees on trades priced against it. This paper does not rank the indices. Which of them become the market's benchmarks will be decided by the participants who use them and by regulators such as the CFTC, whose review of the proposed futures includes the robustness of the indices behind them.
That brings us to the futures. A future is a standardized contract, traded on an exchange and cleared, that fixes a price today for a defined quantity over a future period. Futures settle in one of two ways. A physically delivered future ends with the seller delivering the commodity. For a cash-settled future held to expiry, the buyer's cumulative gain or loss is the final settlement price less the trade price, multiplied by the contract size. The seller has the opposite result. Gains and losses are settled in cash as the position is marked to market, not paid in one lump sum at expiry. This is not a spread between two products. It is one traded price compared with one index. Power traders use the same design in financially settled power futures, which pay against the average hub price for the delivery month and deliver no electricity.56,57,58
The futures announced by CME, ICE, and Nodal are all cash-settled. The proposed CME/Silicon Data and ICE/Ornn contracts have monthly maturities and reference rental rates for either the H100 or the B200. CME specifies a fixed 730 GPU-hours per contract; ICE/Ornn uses the actual number of hours in the contract month. Both would settle against the monthly average of a rental index for that chip from the exchange's index partner, Silicon Data for CME and Ornn for ICE.1,2,53,59 ICE has separately announced cash-settled futures on NativX's COIL.27 COIL is not tied to one chip. NativX describes a volume-weighted fix, but its public materials do not establish which observations are eligible or how each chip contributes.41 A buyer who holds a matching future can offset part of a rise in rental costs if the index rises with its rental price, but still has to rent the machine. Physical delivery is possible in principle, and one route to it has been announced. In July 2026, Architect and Compute Desk announced ComputeConnect, a service designed to facilitate exchange for physical (EFP) transactions in GPU capacity. An EFP is a privately arranged trade that pairs a futures transaction with a corresponding physical transaction, a mechanism long used in energy and agricultural markets. ComputeConnect is intended to let customers of Architect's US exchange pair a futures transaction with a separately arranged physical GPU capacity trade, with Compute Desk's platform handling physical settlement.36 Architect's US venue, American Innovation Exchange, is a CFTC designated contract market.60 Its compute futures have been announced but have not launched.
The contracts differ in what they settle against, and the difference follows directly from Table 5. The CME contracts reference Silicon Data's neocloud On-demand benchmarks.53 The ICE/Ornn contracts reference Ornn's benchmarks of executed On-demand GPU rentals. The Nodal contracts would settle against Compute Desk indices, which track one-year non-cancellable reservations in the United States and blend transacted prices with neocloud list prices.3,44 A buyer whose real exposure is a one-year reservation and a buyer who rents On-demand need different contracts.
The launch timing of the proposed futures is uncertain. CME, through NYMEX, asked the CFTC to approve its two Silicon Data futures before listing, and the review deadline has been extended to January 8, 2027. The review could end sooner if the CFTC approves the contracts earlier, or if another CFTC registrant, such as ICE or Nodal, self-certifies, obtains approval for, or lists a comparable contract first.4,61,62 A compute hedge therefore has three clocks that must line up (Table 6): how the capacity is rented, how the index is observed, and when the derivative matures. Where they do not line up, the hedger carries basis risk, just as a power hedger does when the hedge settles at the hub and the plant sits at a node.
Table 6. Three clocks in a compute hedge
| Clock | The question to ask | Why it matters |
|---|---|---|
| Rental structure | Is the capacity rented On-demand, under a cancellable reservation, or under a non-cancellable reservation? | Helps determine when payment is owed; check separately whether the rate is fixed or floats. |
| Index observation | Does the index use transacted or offer prices, and is it calculated hourly, daily, or as a monthly average? | Sets which market the index actually represents. |
| Derivative maturity | Does the contract expire each month, or never? | Sets when the hedge pays, and whether it has to be rolled or funded. |
[SECTION 05]
5. The search for a standard compute unit (capacity or usage/output)
Every index and every future has to say exactly what is being priced. In established commodity markets, a contract specifies what it prices: a megawatt of generation capacity, a megawatt-hour at a named hub, or a barrel of a named crude oil. Compute has no agreed grade yet, and choosing one is harder than it looks, because accelerators differ in both speed and memory, and both matter to the work they can do. I see four options today, and others may emerge. The first three are generally capacity constructs and the fourth one is a compute usage or output construct.
The first option is to name the accelerator for GPU-hour capacity, which is the dominant way in today's markets. Nearly every index in Table 5, and the proposed CME/Silicon Data, ICE/Ornn, and Nodal/Compute Desk GPU rental futures, prices one hour of a specific GPU model, and each model gets its own index and its own contract. Even then, the index provider has to fix a reference configuration, because the same chip is rented in servers with different memory, networking, and service terms. Compute Desk adjusts prices to a reference GPU variant and contract term, and Compute Screener adjusts every price to one standard eight-GPU server in the United States.30,44 Naming the chip is simple and easy to observe. Its cost is fragmentation. Trading is split across models, each contract ages as its chip is superseded, and a hedger running a different chip carries basis risk.
The second option is to price what the hardware can do, whichever chip does it. Computer architects already measure a chip along two lines. The first is computing speed, counted in floating-point operations per second (FLOPS), where a floating-point operation is the basic arithmetic step in AI work. The second is memory bandwidth, the amount of data the chip can move in and out of its memory each second. A traded unit could be built from each: a fixed computing rate held for one hour, for example one exaFLOP (10¹⁸ floating-point operations) per second, abbreviated here as FxH, and a fixed rate of memory traffic held for one hour, for example one terabyte per second, abbreviated here as BxH. Speed and memory have to be measured together because an AI workload uses both at once, and the attainable rate depends on the workload's arithmetic intensity and the hardware's compute and bandwidth limits, as described by the roofline model.63,64 That is also the difficulty. The two are produced jointly by the same chip, so adding separate prices could double count the same service. The cited telemetry is averaged over time; a settlement design would still need to specify how usage is attributed and verified.65 For now, these measures are better suited to describing a delivered service or a product's specifications than to defining a traded unit.
The third option is to price energy. In my view, this is a variant of the capacity approach. NativX describes its unit, COIL, as 20 kilowatt-hours (kWh) of metered energy used by the computing equipment, delivered as compute on any chip of any generation, and ICE has announced futures on it.27,41 The appeal is a single unit that does not expire with each chip. Table 7 is my own illustration of what one COIL could mean in practice. It rests only on public descriptions and on my own assumptions. NativX has not reviewed or confirmed it, and its detailed methodology is not public. I offer the table to start a discussion, not as a statement of how COIL is actually priced.
Table 7. One COIL on three chips: the author's illustration, not confirmed by NativX
| GPU model | Assumed IT power per GPU | GPU-hours in one 20 kWh COIL | Assumed rental price per GPU-hour | Implied value of one COIL | Implied value per kWh IT |
|---|---|---|---|---|---|
| A100 | 0.81 kW | 24.7 | $1.5 | $37.0 | $1.9 |
| H100 | 1.28 kW | 15.6 | $2.5 | $39.1 | $2.0 |
| B200 | 1.79 kW | 11.2 | $4.1 | $45.6 | $2.3 |
Assumptions: IT power per GPU is the published maximum power of an eight-GPU NVIDIA DGX server (about 6.5 kW for the A100 generation, 10.2 kW for H100, and 14.3 kW for B200) divided by eight, so it includes each GPU's share of the rest of the server.66 Rental prices are illustrative, not quotes. GPU-hours per COIL is 20 kWh divided by IT power per GPU. Implied value per COIL is GPU-hours times the rental price, and implied value per kWh IT is the rental price divided by IT power per GPU.
Three things stand out. First, at the assumed power draws, 20 kWh corresponds to about 25 GPU-hours on an A100 but about 11 on a B200. A buyer who is not told which chip will serve its COIL does not know how much work it will get. Second, the implied value of a COIL is much closer across chips, roughly $37 to $46 on these assumptions, because newer chips draw more power and also rent for more. This is an illustration of how energy normalization can narrow the gap on these inputs, not evidence that it consistently does so. Third, the result depends on choices that only the methodology can settle, such as how much of a server's power is counted for each GPU and how differences in data center efficiency are treated. I would welcome NativX's corrections on any of these points.
The fourth option is to price token usage or output rather than machine time. In power, the heat rate measures how much fuel a plant burns to produce a megawatt-hour. It lets traders compare plants and trade the spread between fuel and power. Tokens per megawatt-hour is closer to the inverse of a heat rate: how much AI output a data center produces from the energy it uses. It is an appealing measure and, for now, an impractical one, because the token is not a fungible unit. Different AI models split the same text into different numbers of tokens.67 They give answers of different lengths to the same question. Billing rules differ too: OpenAI bills reasoning tokens as output tokens even though the raw reasoning tokens are not returned to the user.68 Input tokens and output tokens are priced differently. A token from one model does not necessarily provide the same value as a token from another. Output per unit of energy also depends on the chip and on how the software is run. A benchmark of output per unit of energy needs a defined workload and disclosed model, hardware, and operating conditions, which is why token indices are harder to build than rental indices.
None of these four options has won so far. Until one does, the practical rule is the one from Section 4: know exactly which unit an index prices before relying on it.
[SECTION 06]
6. Perpetuals: a hedge with no expiry date
A perpetual future is a derivative with no expiry date. Because it never settles against the index at maturity, it needs another way to stay close to it, and it uses a periodic funding payment between the two sides. Typically, positive funding means buyers pay sellers; negative funding reverses the payment. The venue's formula determines the rate and direction. The design comes from cryptocurrency markets. In May 2026, the CFTC allowed a bitcoin perpetual to be listed as a future but said perpetuals on other asset classes would be reviewed one by one, so that decision does not approve a compute perpetual in the United States.69 One example is Architect's compute perpetual, linked to Ornn's H100 index. Architect announced it in January 2026 and said in April 2026 that it was live on AX, its Bermuda venue, which is regulated in Bermuda and, according to Architect, is not available to US persons.32,33,70 I have found no published volume or open interest for it, so how much it trades is unverified. Injective launched an H100 rental perpetual in 2025 using Squaretower's price feed.34 In October 2026, Injective launched a basket perpetual based on LO:TECH's SICI.26 Bitget announced H100 and B200 premarket perpetuals in September 2026.35
A perpetual may suit an eligible buyer who knows it will need compute but not when, if the contract is liquid and the index tracks its costs. Suppose an AI lab expects to rent 1,000 H100 GPU-hours in the coming months and worries that the price will rise first. It buys a perpetual at $3.00 per GPU-hour, sized to 1,000 GPU-hours. Two months later, it has a start date and can sell the perpetual at $4.00. It now does two things. It closes the perpetual and collects about $1,000, the $1.00 rise on 1,000 GPU-hours, before funding payments and fees. And it rents the capacity. If it can also book a non-cancellable reservation for the same hours at $4.00 per GPU-hour, the gain on the perpetual brings its effective cost back to about $3.00, and the reservation itself now fixes its price, so it needs no further hedge. If it rents On-demand instead, its price keeps floating, and it would need to keep a hedge in place, most naturally with monthly futures for the months it expects to run. These are illustrative numbers, not market prices.
The limits follow from the design. A perpetual protects against a change in price, not against a high price that persists: once the position is closed, it pays nothing more, however long rents stay at $4.00. While it is open, it can pay or receive funding and ties up margin, and a sharp move against the position can force it to be closed. And like a cash-settled future, it gives no right to a machine. It is a bridge to a physical contract, not a substitute for one. A neocloud with unsold capacity can use the same tool in the other direction, selling a perpetual to protect against falling rents until the capacity is contracted.
[SECTION 07]
7. Which tool fits whom: the buyer, seller, trader, and lender test
Whether any of this is useful depends on who is asking. The test is the same for everyone. What is my actual exposure, and does the index behind the hedge match it in GPU model, region, rental structure, and timing? Table 8 applies that test to five participants.
Table 8. Exposures, tools, and what a hedge leaves uncovered
| Participant | The first question to ask | Tools that may fit | What a hedge does not cover |
|---|---|---|---|
| Compute buyer paying a floating price | Which GPU model, region, and rental structure does my invoice follow? | A future on a matching On-demand index; a perpetual while timing is uncertain | Getting the machines, service quality, and the gap between the index and the invoice |
| Buyer holding a capacity reservation | Am I paying for capacity I may not use? | A fixed price reservation already fixes the rate | Utilization and performance, which no price hedge addresses |
| Neocloud or other seller | How much of my revenue is contracted at fixed prices, and how much floats or is unsold? | Long-term take-or-pay contracts first; selling futures or perpetuals for the floating part | Idle GPUs, customer credit, and power cost |
| Trader or market maker | Can I trade the differences between indices, locations, tenors, contracts, and funding rates? | Futures and perpetuals, once they are live and liquid | Thin markets, margin calls, and the integrity of the index |
| Lender against GPUs | Which cash flows repay the loan if rental prices fall? | Contracted take-or-pay revenue and collateral; an index put option held by the borrower | Borrower credit, and whether the option seller pays |
The lender's case shows how the physical and financial layers combine. A loan against GPUs is repaid from rental revenue, so the lender's main protection is the quality of the borrower's customer contracts. One large neocloud, for example, describes financing that is secured by long-term take-or-pay contracts.23 Where revenue is not fully contracted, the borrower can buy a put option on a rental index from a third party. If the index falls below the strike price, the option pays, offsetting part of the lost revenue.
There are two basic ways to arrange this. In the first, the lender lends directly to the borrower, and a separate party sells the borrower the put. In the second, an intermediary stands in the middle: it borrows from the lender, lends on to the borrower, and sells the put itself. Which structure is used depends heavily on how risk is allocated and on the credit arrangements among the counterparties. In either case, the option protects against price only. Borrower default, customers who do not pay, idle equipment, and an option seller who fails to pay are separate risks, and the option helps only if its payout arrives when it is needed and reaches the lender under the loan documents. Options of this kind are normally documented under an International Swaps and Derivatives Association (ISDA) Master Agreement, with a credit support annex where collateral is required. The loan and its security are documented separately.71
One group is so far missing from the public debate: the largest sellers. Hyperscalers and neoclouds both stand to benefit from deeper and more transparent markets, and both could use these tools. Yet the roughly 100 comment letters filed in response to the CFTC's request, which closed on October 20, 2026, came from exchanges, index providers, trading firms, investors, and smaller GPU marketplaces. I found no direct submissions from the named hyperscalers or neoclouds in the public comments reviewed.72 Silence is not opposition, and it may reflect nothing more than caution.
[SECTION 08]
8. One possible direction: a system operator for compute
Section 3 ended on the biggest difference between the two markets: compute has no system operator. It is worth asking what one would look like, because the question shows how far the market could still travel. In power, an independent system operator (ISO) or regional transmission organization (RTO) is a non-profit body that runs the grid in a region. One of the key functions of every ISO and RTO is to run daily energy markets. It uses supply offers and system conditions to dispatch available resources at least cost, subject to transmission, operating, and reliability constraints. The energy prices it publishes are the reference for the whole market.73
A compute ISO would do the same for the data centers in a region. Table 9 sets out the mapping. Data centers, large and small, would be the power plants, each with its own efficiency and cost. Electricity would be their fuel. Compute users would be the load. Fiber networks would be the transmission lines. The operator could match eligible, portable jobs to data centers on a least-cost basis, subject to hardware, latency, data location, and service constraints. Published clearing prices would describe that pool, not the entire compute market.
Table 9. A compute ISO by analogy with a power ISO for energy dispatch (usage or output)
| Power system | Compute counterpart | Comment |
|---|---|---|
| Power plants, each with its own efficiency and cost | Data centers and the accelerators in them | Newer chips and better run sites play the part of the efficient plants. |
| Fuel cost | Cost of electricity | Electricity is to a data center what fuel is to a power plant. |
| Energy volume (megawatt-hours) | Compute output | Hard to measure until the unit of output is standardized (Section 5). |
| Heat rate: fuel burned per megawatt-hour | Energy used per unit of compute output | Same as above. |
| Load: homes and businesses | Compute users running training and inference jobs | Some jobs must run now. Others can wait or move, which gives a dispatcher room to optimize. |
| Transmission and distribution lines | Fiber networks | Congestion on fiber is rarely discussed today but could matter as compute volumes grow. |
| ISO or RTO: non-profit, least-cost dispatch, reliability | A neutral compute operator | A thought experiment, not an operating regional market. |
| Locational prices from the dispatch | Transparent prices by region and class of accelerator | Would replace much of the index reconstruction described in Section 4. |
If an ISO or RTO also operates a capacity market, it usually does so through auctions, which may be annual, seasonal, or monthly. Compared with a compute ISO that dispatches compute usage or output, a compute ISO that at first operates only a capacity market might be much easier to create. A compute capacity market could be built like a power capacity market, with auctions, and it might partly replace some of the functions that several exchanges are now trying to build with GPU-hour futures and indices. This is my own thought experiment, and the obstacles are real. A compute ISO is unlikely any time soon, especially if compute capacity remains scarce for the foreseeable future. But power markets did not begin with ISOs either. They moved from bilateral contracts to organized markets over decades, and compute may be at the start of a similar path.
Conclusion
Three points sum up the map. First, compute is sold as capacity over time, and the terms of sale matter as much as the chip. On-demand, spot, cancellable reservations, and non-cancellable reservations are four different products with four different prices. Second, the indices are not interchangeable. Each measures a particular GPU model, region, rental structure, and price type, and the depth of the underlying rental market varies by benchmark. Third, the derivatives inherit all of that. A future or a perpetual is only as good a hedge as its index is a match for the hedger's own contract, and a cash-settled contract never supplies a machine. The instruments themselves are familiar from other commodity markets. The work lies in checking, each time, what the price behind them actually measures. Two open questions will shape what comes next: whether the market settles on one or more standard units, and whether it ever gains a neutral operator.
[DISCLOSURE]
Published by Rheostat for informational purposes only. Generative AI tools were used in the preparation of this paper. All final content, arguments, and data have been reviewed and edited by the author. This paper is not an offer, solicitation, or recommendation to transact, and is not investment, legal, tax or accounting advice. Market data is drawn from public sources including issuer filings and press releases.
[REFERENCES]
References
1 CME Group, CME Group and Silicon Data to Launch Compute Futures on October 5 to Unlock New Way to Hedge AI Risks, August 11, 2026. https://www.cmegroup.com/media-room/press-releases/2026/8/11/cme_group_and_silicondatatolaunchcomputefuturesonoctober5tounloc.html
2 Intercontinental Exchange, ICE Futures U.S. Exchange Notice - ORNN Compute 4Q2026, September 29, 2026. https://www.ice.com/publicdocs/futures_us/exchange_notices/ICE_Futures_US_Exchange_Notice_ORNN_Compute_4Q2026_20260929.pdf
3 Nodal Exchange, Nodal Exchange and Compute Desk to Launch Compute Futures Contracts, September 3, 2026. https://www.nodalexchange.com/nodal-exchange-and-compute-desk-to-launch-compute-futures-contracts/
4 NYMEX, Voluntary Extension of Review Period for the Silicon Data H100 and B200 Rental Index Futures, filed with the CFTC, November 4, 2026. https://www.cftc.gov/filings/documents/2026/orgdcmnymexvolexten261105.pdf
5 National Institute of Standards and Technology, The NIST Definition of Cloud Computing, SP 800-145, September 2011, section 2; accessed November 1, 2026. https://nvlpubs.nist.gov/nistpubs/Legacy/SP/nistspecialpublication800-145.pdf
6 Amazon Web Services, Happy 20th Birthday, Amazon EC2, August 25, 2026. https://aws.amazon.com/blogs/aws/happy-20th-birthday-amazon-ec2/
7 Amazon Web Services, Amazon EC2 Introduces Reserved Instances, March 12, 2009. https://aws.amazon.com/about-aws/whats-new/2009/03/12/amazon-ec2-introduces-reserved-instances/
8 Amazon Web Services, Announcing Cluster GPU Instances for Amazon EC2, November 15, 2010. https://aws.amazon.com/about-aws/whats-new/2010/11/15/announcing-cluster-gpu-instances-for-amazon-ec2/
9 Amazon Web Services, Announcing Amazon EC2 Capacity Blocks for ML, November 1, 2023. https://aws.amazon.com/about-aws/whats-new/2023/11/amazon-ec2-capacity-blocks-ml/
10 Amazon Web Services, Capacity Blocks pricing and billing, Amazon EC2 User Guide, undated; accessed November 3, 2026 (the full price is charged at purchase and the reservation cannot be modified or canceled). https://docs.aws.amazon.com/AWSEC2/latest/UserGuide/capacity-blocks-pricing-billing.html
11 Silicon Data, Silicon Data Launches World's First Daily GPU Rental Index, May 20, 2025. https://www.silicondata.com/news-room/silicon-data-launches-world-s-first-daily-gpu-rental-index-to-bring-transparency-to-the-ai-infrastructure-economy-copy
12 Commodity Futures Trading Commission, Request for Comment on compute derivatives, Federal Register, August 21, 2026. https://www.govinfo.gov/content/pkg/FR-2026-08-21/pdf/2026-17163.pdf
13 Amazon Web Services, On-Demand Instances, Amazon EC2 User Guide, undated; accessed November 1, 2026. https://docs.aws.amazon.com/AWSEC2/latest/UserGuide/ec2-on-demand-instances.html
14 Amazon Web Services, Instance purchasing options, Amazon EC2 User Guide, undated; accessed November 1, 2026. https://docs.aws.amazon.com/AWSEC2/latest/UserGuide/instance-purchasing-options.html
15 Lambda, Billing overview, Lambda Docs, accessed November 4, 2026 (On-Demand Cloud instances are billed in one minute increments for as long as they run). https://docs.lambda.com/public-cloud/billing/
16 Nebius, Compute pricing in Nebius AI Cloud, Nebius documentation, accessed November 4, 2026 (GPU virtual machines are billed per second while running; preemptible prices are listed separately). https://docs.nebius.com/compute/resources/pricing
17 Google Cloud, Spot VMs, Compute Engine documentation, accessed November 4, 2026 (Compute Engine can preempt Spot VMs at any time to reclaim capacity). https://cloud.google.com/compute/docs/instances/spot
18 Vast.ai, Rental Types FAQ, Vast.ai documentation, accessed November 4, 2026 (interruptible instances run on a bid and can be stopped by higher bids). https://docs.vast.ai/guides/reference/faq/rental-types
19 Amazon Web Services, Capacity Reservation pricing and billing, Amazon EC2 User Guide, undated; accessed November 3, 2026 (unused reserved capacity is billed at the On-Demand rate whether or not instances run). https://docs.aws.amazon.com/AWSEC2/latest/UserGuide/capacity-reservations-pricing-billing.html
20 Microsoft Azure, On-demand capacity reservation, Azure Virtual Machines documentation, updated June 16, 2026; accessed November 4, 2026 (capacity can be reserved for any duration, created or deleted at any time, and is billed at the virtual machine rate even when unused). https://learn.microsoft.com/azure/virtual-machines/capacity-reservation-overview
21 Oracle, Capacity reservations, Oracle Cloud Infrastructure documentation, updated October 13, 2026; accessed November 4, 2026 (charges begin when the reservation is created, including for unused capacity, and stop when it is deleted). https://docs.oracle.com/en-us/iaas/Content/Compute/Tasks/capacity-reservations-create_tabs.htm
22 Amazon Web Services, Amazon EC2 Capacity Blocks for ML is now available in AWS GovCloud (US) Regions, June 12, 2026 (reservations for durations up to six months, booked up to eight weeks in advance). https://aws.amazon.com/about-aws/whats-new/2026/06/amazon-ec2-capacity-blocks-ml-govcloud/
23 CoreWeave, Investor Presentation, March 2026 (long-term take-or-pay customer contracts for committed multiyear reserved capacity with no termination for convenience; asset-level financing secured by those contracts). https://s205.q4cdn.com/133937190/files/doc_events/2026/Mar/02/March-2026-Investor-Presentation.pdf
24 Google Cloud, AI Hypercomputer consumption models, undated; accessed November 1, 2026. https://docs.cloud.google.com/ai-hypercomputer/docs/consumption-models
25 GridCompute, Market Overview and GPU Price Tracker. Accessed November 12, 2026. https://gridcompute.ai/; https://gridcompute.ai/gpu/.
26 LO:TECH, Compute now has a reference price, October 29, 2026. https://lo.tech/news/compute-now-has-a-reference-price.
27 Intercontinental Exchange, ICE and NativX to Launch Energy-Normalized Compute Futures Contracts, July 1, 2026. https://ir.theice.com/press/news-details/2026/ICE-and-NATIVX-to-Launch-Energy-Normalized-Compute-Futures-Contracts/default.aspx
28 Silicon Data, LLM Token Expenditure Index, undated; accessed November 1, 2026. https://www.silicondata.com/products/silicon-index/llm-token-expenditure-index
29 Ornn, Token Price Index documentation, undated; accessed November 1, 2026. https://data.ornn.com/docs/token-price-index
30 Compute Screener, Methodology v2.0, effective October 28, 2026. https://computescreener.com/methodology
31 OpenRouter, Models and Provider Routing documentation, undated; accessed November 1, 2026. https://openrouter.ai/docs/guides/routing/provider-selection
32 Architect, Architect partners with Ornn to launch compute futures, January 21, 2026. https://architect.co/insights/press/architect-ornn-compute-futures/
33 Architect, post on X, April 3, 2026 (states that perpetual futures on NVIDIA H100 rental prices, using Ornn's indices, are live on AX). https://x.com/Architect_Fi/status/2040068321643516034
34 Injective, Injective Releases the First-Ever Onchain NVIDIA GPU Market, August 18, 2025; and Injective Launches the Super Intelligence Compute Index, October 26, 2026. https://injective.com/blog/injective-releases-the-first-ever-onchain-nvidia-gpu-market ; https://injective.com/blog/super-intelligence-compute-index
35 Bitget, Listing of H100USDT and B200USDT Pre-Market Compute Perps, September 8, 2026. https://www.bitget.com/asia/support/articles/12560603894397
36 Architect Financial Technologies, Architect Financial Technologies Partners with Compute Desk to Launch Exchange-for-Physical Market for GPU Compute, PR Newswire, July 8, 2026. https://www.prnewswire.com/news-releases/architect-financial-technologies-partners-with-compute-desk-to-launch-exchange-for-physical-market-for-gpu-compute-302820762.html
37 Kalshi, Compute Has a Market-Implied Forward Curve Now, July 14, 2026. https://news.kalshi.com/p/gpu-compute-forward-curve
38 FalconX, FalconX Executes First OTC Compute Forward Trade, May 27, 2026. https://www.falconx.io/newsroom/falconx-executes-first-otc-compute-forward-trade
39 OpenAI, API pricing, accessed November 12, 2026. https://developers.openai.com/api/docs/pricing
40 Shanghai Metals Market, GPU Cloud Compute, accessed November 4, 2026 (the methodology uses transaction data and representative public listings). https://www.metal.com/categories/gpu-cloud-compute
41 NativX, public product overview (20 kWh per COIL and volume-weighted fix), accessed November 12, 2026. https://nativx.net/
42 Computable, GPU Index, undated; accessed November 1, 2026. https://www.getcomputable.com/
43 Computable, GPU Index Methodology, undated; accessed November 3, 2026 (On-Demand list prices only; spot, preemptible, and reserved tiers are not eligible). https://github.com/getcomputable/gpu-index/blob/main/METHODOLOGY.md
44 Compute Desk, Index Methodology, undated; accessed November 4, 2026 (US one-year reserved rental prices; completed transactions blended with published neocloud list prices). https://www.computedesk.com/index-methodology
45 Ornn, OCPI Methodology, first published July 24, 2026; accessed November 1, 2026 (only realized rental prices are included; offered prices are excluded; On-Demand rentals only). https://data.ornn.com/methodology
46 SemiAnalysis, GPU Spot and Contract Composite Index methodology, undated; accessed November 1, 2026. https://gpu-index.semianalysis.com/
47 Setara Financial Corp., Setara Announces the Release of its Initial GPU Price Index, PR Newswire, November 3, 2026. https://www.prnewswire.com/news-releases/setara-announces-the-release-of-its-initial-gpu-price-index-302927135.html
48 Setara Financial Corp., Setara GPU Index and Methodology pages, accessed November 3, 2026 (pre-production index from posted On-Demand list prices; lowest eligible price per provider; equally weighted average of at least three providers). https://www.setarafinancial.com/gpu-index
49 Shanghai Metals Market, Computing Power Price Index methodology, October 2026; accessed November 1, 2026. https://www.metal.com/methodology/gpu-cloud-compute
50 Shanghai Metals Market, Launch of SMM Southeast Asia Compute Rental Retail Price, August 28, 2026 (retail prices for Singapore, Malaysia, and Indonesia, published as OnDemand, Preemptible, 1YrCommit, and 3YrCommit price points). https://www.metal.com/announcements/104085449-smm-announcementlaunch-of-smm-southeast-asia-compute-rental-retail-price
51 Silicon Data, Silicon Index family, undated; accessed November 1, 2026. https://www.silicondata.com/products/silicon-index
52 Silicon Data, Silicon Data Index Governance and Methodology, undated; accessed November 3, 2026 (inputs include transactions and directly executable listings; spot and reserved records are normalized to an On-Demand equivalent). https://www.silicondata.com/documents/silicon-data-index-governance.pdf
53 CME Group, Compute Futures, undated; accessed November 1, 2026. https://www.cmegroup.com/markets/energy/power/compute-futures.html
54 The Compute Indices, Methodology v0.11.1, effective October 11, 2026; accessed November 1, 2026. https://thecomputeindices.com/methodology.html
55 International Organization of Securities Commissions, Principles for Financial Benchmarks, July 2013. https://www.iosco.org/library/pubdocs/pdf/ioscopd415.pdf
56 Commodity Futures Trading Commission, Futures Market Basics, undated; accessed November 1, 2026. https://www.cftc.gov/LearnAndProtect/EducationCenter/FuturesMarketBasics/index2.htm
57 CME Group, Cash Settlement vs. Physical Delivery, March 25, 2024. https://www.cmegroup.com/education/articles-and-reports/agriculture-cash-settlement-vs-physical-delivery
58 CME Group, Compute Futures fact card and Mark-to-Market, accessed November 12, 2026. https://www.cmegroup.com/markets/energy/files/compute-futures-fact-card.pdf ; https://www.cmegroup.com/education/courses/introduction-to-futures/mark-to-market
59 Intercontinental Exchange, ICE Ornn B200 contract specification, undated; accessed November 1, 2026. https://www.ice.com/products/83050009
60 CFTC, American Innovation Exchange, LLC, Designated Contract Market record (originally designated as IMX Health, LLC on January 18, 2024), accessed November 12, 2026. https://www.cftc.gov/IndustryOversight/IndustryFilings/TradingOrganizations/48926; CFTC designation announcement, January 18, 2024. https://www.cftc.gov/PressRoom/PressReleases/8849-24; Architect acquisition announcement, May 28, 2026. https://architect.co/insights/press/architect-american-innovation-exchange-compute/
61 Commodity Futures Trading Commission, Designated Contract Market Products filings: NYMEX Silicon Data H100 and B200 Rental Index Futures, both listed as approval pending. Accessed November 12, 2026. https://www.cftc.gov/IndustryOversight/IndustryFilings/TradingOrganizationProducts/62544; https://www.cftc.gov/IndustryOversight/IndustryFilings/TradingOrganizationProducts/62545.
62 Electronic Code of Federal Regulations, 17 CFR 40.3(c) and (d), voluntary submission of new products for Commission review and approval. Accessed November 12, 2026. https://www.ecfr.gov/current/title-17/chapter-I/part-40/section-40.3.
63 S. Williams, A. Waterman, and D. Patterson, Roofline: An Insightful Visual Performance Model for Multicore Architectures, Communications of the ACM, 52(4), April 2009, pages 65 to 76. https://doi.org/10.1145/1498765.1498785
64 R. Pope and others, Efficiently Scaling Transformer Inference, Proceedings of Machine Learning and Systems 5 (MLSys 2023). https://proceedings.mlsys.org/paper_files/paper/2023/hash/c4be71ab8d24cdfb45e3d06dbfca2780-Abstract-mlsys2023.html
65 NVIDIA, Data Center GPU Manager (DCGM) documentation, Profiling, accessed November 12, 2026 (profiling metrics are averages over a time interval and do not attribute activity to a particular piece of work). https://docs.nvidia.com/datacenter/dcgm/latest/learn/modules/profiling.html
66 NVIDIA, DGX A100, DGX H100/H200, and DGX B200 user guides, Power Specifications (maximum system draws of 6.5, 10.2, and 14.3 kW, respectively), accessed November 12, 2026. https://docs.nvidia.com/dgx/pdf/dgxa100-user-guide.pdf ; https://docs.nvidia.com/dgx/dgxh100-user-guide/introduction-to-dgxh100.html ; https://docs.nvidia.com/dgx/dgxb200-user-guide/introduction-to-dgxb200.html
67 OpenAI, Understanding and Counting Tokens, accessed September 2026. https://help.openai.com/en/articles/4936856-understanding-and-counting-tokens
68 OpenAI, Reasoning models, OpenAI API documentation, accessed November 12, 2026 (raw reasoning tokens are not returned; they are billed as output tokens and counted in API usage metadata). https://developers.openai.com/api/docs/guides/reasoning
69 Commodity Futures Trading Commission, Policy Statement Concerning the Listing of Perpetual Contracts, adopted May 29, 2026. https://www.cftc.gov/media/14066/PerpetualContractsPolicyStatement052926/download
70 Architect, AX Expands Access to Eligible Individual Investors and Omnibus Brokerage Accounts, July 21, 2026. https://architect.co/insights/press/architect-ax-individual-investors-omnibus/
71 International Swaps and Derivatives Association, Best Practices for the OTC Derivatives Collateral Process, June 30, 2010. https://www.isda.org/a/S0iDE/isda-best-practices-for-the-otc-derivatives-collateral-process.pdf
72 CCIR, Compute Derivatives at the CFTC: Letters and Statements, a third-party compilation of docket CFTC-2026-1850, as of November 10, 2026 (lists 97 comment letters, with docket entries numbered up to 109; counts differ with the treatment of duplicates and late filings; comments closed October 20, 2026). https://ccir.io/research/compute-derivatives-cftc
73 Federal Energy Regulatory Commission, Energy Primer, January 2024. https://ferc.gov/sites/default/files/2024-01/24_Energy-Markets-Primer_0117_DIGITAL_0.pdf
[ABOUT]
About the author
Cheng Chang is a director at Rheostat and founder of Myriad Galaxies, an advisory firm focused on energy and AI infrastructure. He brings more than 20 years of experience in energy investment, commercial development, and commodity derivatives to his work on compute pricing and trading.