Professional clients and eligible counterparties only. Compute contracts are not yet available and are subject to regulatory approval.
Launching, subject to regulatory approval
Five-year timed perpetual futures on NVIDIA H100, H200, B200 and B300 capacity. Settle in cash or take delivery of verified GPU nodes. Prices reference the Compute Desk GPU rental indices. Trading is planned to take place on the MiFID II multilateral trading facility operated by PM MTF Ltd, subject to regulatory approval.
USD per GPU-hour · US one-year rental
Levels as displayed on compute-desk.com on 2 November 2026. Not a trading price and not a forecast. DESKH100 is calculated and published by General Index; DESKH200, DESKB200 and DESKB300 by Compute Desk (The Compute Index, Inc.). [Display licence to be confirmed.]
01 Contracts
Each GPU model is listed in a cash-settled and a physically delivered version, so there are eight instruments in all. You can choose pure price exposure, or a contract that can end in real, verified capacity.
Instrument codes and terms are indicative and subject to regulatory approval.
NVIDIA Hopper
Deliverable grade (indicative) 80 GB HBM3 · SXM 8‑GPU HGX node
Hopper-generation capacity, priced against the H100 benchmark. Hedge it in cash, or settle into verified 8-GPU HGX nodes.
Cash-settled
CMH100C
Funding and final settlement in cash against DESKH100.
Physically delivered
CMH100P
Monthly delivery of verified HGX H100 nodes through COMPUTE Clear. Funding against DESKH100.
NVIDIA Hopper
Deliverable grade (indicative) 141 GB HBM3e · SXM 8‑GPU HGX node
The Hopper footprint with 141 GB of HBM3e per GPU. Manage exposure to memory-heavy inference capacity, cash-settled or delivered.
Cash-settled
CMH200C
Funding and final settlement in cash against DESKH200.
Physically delivered
CMH200P
Monthly delivery of verified HGX H200 nodes through COMPUTE Clear. Funding against DESKH200.
NVIDIA Blackwell
Deliverable grade (indicative) c. 180 GB HBM3e · SXM 8‑GPU HGX node
Blackwell capacity on a five-year series. Hedge the cost or the revenue of B200 GPU-hours against the B200 benchmark.
Cash-settled
CMB200C
Funding and final settlement in cash against DESKB200.
Physically delivered
CMB200P
Monthly delivery of verified HGX B200 nodes through COMPUTE Clear. Funding against DESKB200.
NVIDIA Blackwell Ultra
Deliverable grade (indicative) Up to 288 GB HBM3e · SXM 8‑GPU HGX node
The newest GPU in the set. Manage exposure to Blackwell Ultra capacity before you commit capital.
Cash-settled
CMB300C
Funding and final settlement in cash against DESKB300.
Physically delivered
CMB300P
Monthly delivery of verified HGX B300 nodes through COMPUTE Clear. Funding against DESKB300.
The deliverable grade applies to the physically delivered series. The reference GPU variant of each index is set by Compute Desk’s index specification and may differ from the deliverable grade.
Each series takes a maturity suffix, for example CMH100C-[YYMM]. Normally one series per instrument is listed. During the final [12] months, the successor series trades alongside the maturing series so that positions can be rolled. One contract is one 8-GPU node for 730 hours (5,840 GPU-hours). Notionals use the illustrative index levels displayed on 2 Nov 2026.
About the reference indices. Each Compute Desk index is a daily benchmark of the prevailing US price for a one-year reserved rental of one NVIDIA GPU, in USD per GPU-hour. Under Compute Desk’s published methodology it combines completed transactions with published neocloud list prices (hyperscaler list prices are excluded), adjusts observations to a reference GPU and contract term, and smooths the result. Daily values are calculated at 16:30 New York time on US business days. Index methodology (opens in a new tab)
| GPU | Cash-settled | Physically delivered | Reference index | Index calculated and published by |
|---|---|---|---|---|
| NVIDIA H100 | CMH100C | CMH100P | DESKH100 | General Index |
| NVIDIA H200 | CMH200C | CMH200P | DESKH200 | Compute Desk |
| NVIDIA B200 | CMB200C | CMB200P | DESKB200 | Compute Desk |
| NVIDIA B300 | CMB300C | CMB300P | DESKB300 | Compute Desk |
Under consideration for later phases (subject to demand, the availability of a suitable reference index and regulatory approval): rack-scale NVIDIA GB200 and GB300 NVL72 systems, AMD Instinct accelerators, and region-specific delivery. No further contracts are committed at this stage.
02 How it works
A timed perpetual trades like a perpetual future. There are no monthly expiries to roll, and a daily funding payment keeps the price close to the index. Each series also has a fixed final maturity five years after listing, so every position has a defined end date.
One series per GPU and settlement variant runs for five years. For most of a series’ life, liquidity is concentrated in one order book per instrument instead of being spread across a strip of expiries.
Once per business day, the average futures price is compared with the latest published Compute Desk index value. When futures trade above the index, longs pay shorts. When they trade below, shorts pay longs. The rate is capped at ±0.50% per day indicative.
Each series matures on the [last business day] of the 60th month after listing. Open positions then settle in cash against the average of the final [5] index values, plus a final funding payment. A successor series is listed [12] months before maturity so positions can be rolled in an orderly way.
A timed perpetual tracks the current level of the index. It does not lock in a price for a future date. Over time, funding payments can be material, and they belong in any hedging calculation.
Illustrative. Not based on market data.
Futures above index
Futures below index
Funding is paid between members holding opposite positions. It is not a fee to the venue. It is settled once per business day as a cash adjustment, separately from variation margin.
A positive rate means longs pay shorts; a negative rate means shorts pay longs. “Index” is the most recent index value published before the funding window, normally the previous US business day’s 16:30 New York value. On days with no new index value, the last published value is used [to be confirmed in the Rulebook]. The futures price is sampled every minute in a published funding window [times to be confirmed], using the last trade where it lies within the best bid and offer and the mid-price otherwise. Minutes without a two-sided market are excluded.
| Index, DESKH100 (as displayed on 2 Nov 2026) | 2.920 |
|---|---|
| Average futures price, funding window | 2.9215 |
| Premium = (2.9215 − 2.920) ÷ 2.920 | +0.0514% |
| Daily settlement price | 2.922 |
| Funding per contract = 0.000514 × 2.922 × 5,840 | USD 8.77 |
| 10 contracts | Long pays USD 87.70 Short receives USD 87.70 |
If the average futures price were 2.940, the premium would be +0.685%, so the cap applies and the rate is +0.50%. At a settlement price of 2.941, that is USD 85.88 per contract. Variation margin is separate: if the settlement price rises from 2.922 to 2.937, a 10-contract long receives 0.015 × 5,840 × 10 = USD 876.00.
03 Calculator
Convert between GPUs and contracts, see the notional, and estimate the funding paid or received on a position. The calculator uses the indicative contract terms and the illustrative index levels displayed on 2 Nov 2026.
Illustrative only. Based on indicative terms: 5,840 GPU-hours per contract, a USD 0.001 tick and daily funding capped at ±0.50%. Funding per contract is rounded to the cent. This is not a quote, forecast or offer, and it excludes fees, margin and clearing costs.
04 Physical delivery
Holders of the physically delivered contracts can settle some or all of a position by delivery of real GPU capacity in monthly cycles. Capacity is delivered and verified through COMPUTE Clear, the physical settlement service operated by Compute Desk. It checks that the hardware delivered matches the contract specification.
Indicative. Delivery arrangements are subject to the Compute Markets Rulebook, regulatory approval and final agreement with Compute Desk.
In each monthly cycle’s nomination window [dates to be confirmed], longs nominate to take delivery and shorts nominate to make delivery, in whole contracts. Delivering members register their 8-GPU nodes with COMPUTE Clear in advance.
Nominations are matched [pro rata]. If they do not balance, the imbalance is published and eligible members can cover it in a balancing window. Matched positions close at the delivery price [definition to be confirmed, e.g. the daily settlement price on the matching day]. Unmatched nominations lapse and the positions stay open.
COMPUTE Clear checks that the hardware made available matches the contract specification for the series before delivery is confirmed.
The seller makes the allocated nodes available to the buyer through COMPUTE Clear. [Access model, e.g. bare-metal or single-tenant: to be confirmed.]
Each contract delivers 730 consecutive hours of capacity from [delivery start]. The buyer’s payment is released to the seller against verified delivery [payment and settlement agent to be confirmed]. Remedies for partial or failed delivery are set out in the Rulebook.
Deferral fee indicative [0.25%] of notional per cycle. When nominations to take delivery exceed nominations to make delivery, open short positions pay a deferral fee to open long positions. When the reverse is true, longs pay shorts. Positions created by balancing nominations receive the fee.
One contract delivers one 8-GPU node for 730 hours. The final grade, substitution and tolerance rules, acceptance tests and delivery locations will be set out in the contract specifications and the MTF Rulebook.
| GPU | Node | GPU memory | GPU interconnect | Network fabric per GPU | Facility |
|---|---|---|---|---|---|
| H100 | NVIDIA HGX H100, 8 × SXM | 80 GB HBM3 | 4th-gen NVLink via NVSwitch, 900 GB/s per GPU | ≥ 400 Gb/s, InfiniBand NDR or RoCE, 1:1 GPU‑to‑NIC | Tier III-equivalent¹ |
| H200 | NVIDIA HGX H200, 8 × SXM | 141 GB HBM3e, 4.8 TB/s | NVLink via NVSwitch, 900 GB/s per GPU | ≥ 400 Gb/s, 1:1 GPU‑to‑NIC | Tier III-equivalent¹ |
| B200 | NVIDIA HGX B200, 8 × SXM | c. 180 GB HBM3e (c. 1.44 TB per node) | 5th-gen NVLink, 1.8 TB/s per GPU | ≥ 400 Gb/s, 1:1 GPU‑to‑NIC | Tier III-equivalent¹; power and cooling for c. 14 kW‑class nodes |
| B300 | NVIDIA HGX B300, 8 × SXM (Blackwell Ultra) | Up to 288 GB HBM3e (at least 2.1 TB per node) | 5th-gen NVLink, 1.8 TB/s per GPU | ≥ 800 Gb/s (ConnectX-8 class), 1:1 GPU‑to‑NIC | Tier III-equivalent¹ |
05 Participants
The firms that build, finance, consume and price GPU capacity all carry exposure to the same rental curve. A venue only works when these natural hedgers meet professional liquidity providers on equal terms.
A Lenders and underwriters
Banks and debt funds financing GPU fleets and data centres, equipment lessors, the sponsors behind them, and firms underwriting compute-backed risk.
Cash-settled
B Compute sellers
Data-centre and colocation operators, cloud GPU platforms and firms reselling reserved capacity.
Cash-settled Physically delivered
C Compute buyers
Model developers, software companies running inference at scale, and corporate AI and research teams.
Cash-settled Physically delivered
D Professional trading firms
Proprietary trading houses, electronic liquidity providers, and discretionary or systematic funds active in commodity and macro markets.
Cash-settled Physically delivered
Examples are illustrative, not recommendations. Firms that are not authorised under MiFID II should take their own advice on whether they need authorisation, or can rely on an exemption, before trading on their own account.
06 Specifications
The two settlement variants share one contract design. They differ only in how a position can end. Every term below is indicative and is not available for trading.
| Term | Cash-settled | Physically delivered |
|---|---|---|
| Contract type | Timed perpetual future: perpetual-style daily funding, with a fixed final maturity for each series | |
| Underlying | NVIDIA H100, H200, B200 and B300 GPU capacity | |
| Instrument codes | CMH100C · CMH200C · CMB200C · CMB300C | CMH100P · CMH200P · CMB200P · CMB300P |
| Series code | Instrument code plus maturity suffix, e.g. CMH100C-[YYMM] | |
| Reference index | DESKH100 · DESKH200 · DESKB200 · DESKB300: Compute Desk daily US one-year rental price per GPU-hour | |
| Use of the index | Funding and final settlement | Funding, and final settlement of positions not settled by delivery |
| Contract unit | 5,840 GPU-hours: one 8-GPU node for 730 hours | |
| Price quotation | USD per GPU-hour, to three decimal places | |
| Minimum price increment | USD 0.001 per GPU-hour = USD 5.84 per contract | |
| Minimum order size | 1 contract (8 GPUs) | |
| Trading hours | [To be confirmed] | |
| Daily settlement price | Volume-weighted average price in the closing window [15:45–16:00 CET/CEST]; fall-backs set out in the Rulebook | |
| Funding | Once per business day. Rate = (average futures price over the funding window − index) ÷ index, capped at ±0.50% [product-specific]. Positive: longs pay shorts. Negative: shorts pay longs | |
| Index value used for funding | The most recent index value published before the funding window, normally the previous US business day’s 16:30 New York value. On days with no new index value, the last published value is used [to be confirmed in the Rulebook] | |
| Variation margin | Daily, in USD; funding settled as a separate cash adjustment in the same cycle | |
| Clearing and margin | [Arrangement to be confirmed] | |
| Series and maturity | Normally one series per instrument. Final maturity on the [last business day] of the 60th month after listing. During the final [12] months, the successor series trades alongside the maturing series so that positions can be rolled | |
| Physical delivery | Not applicable | Monthly cycles: 730 consecutive hours of verified 8-GPU capacity per contract, through Compute Desk COMPUTE Clear |
| Delivery price | Not applicable | Price at which matched positions close for delivery [definition to be confirmed, e.g. the daily settlement price on the matching day] |
| Deferral fee | Not applicable | Indicative [0.25%] of notional per cycle, paid to the side whose delivery nominations are not met |
| Final settlement | Cash, against the average of the final [5] published index values, plus a final funding payment | As cash-settled, for any positions remaining after the last delivery cycle |
| Position management | Venue position management controls apply. EU commodity-derivative position limits and position reporting may apply [under review] | |
| Fees | [To be published] | |
Indicative terms. Final specifications are set out in the MTF Rulebook and contract specifications and are subject to regulatory approval.
07 Why an EU MTF
Compute Markets is designed to run on the multilateral trading facility operated by PM MTF Ltd under MiFID II. The venue’s rules are written down, published, and applied to every member in the same way.
Orders will interact under published, non-discretionary rules. The operator may not trade against members’ orders with its own capital.
Membership criteria are transparent, non-discriminatory and based on objective criteria, and will be published in the Rulebook.
Quotes and trades are made public under MiFIR, subject to any waivers and deferrals that apply.
Arrangements to monitor orders and trades, and to detect and report suspected market abuse under the Market Abuse Regulation (MAR).
Open to professional clients and eligible counterparties. Retail clients are not admitted.
A Cyprus investment firm supervised by the Cyprus Securities and Exchange Commission, within the EU single rulebook: MiFID II, MiFIR, MAR and, where applicable, EMIR. [client to confirm wording and that the authorisation covers operation of an MTF]
08 Membership
Access to the venue is governed by transparent, non-discriminatory rules based on objective criteria, which will be published in the Rulebook [link when available]. Members must qualify as professional clients or eligible counterparties.
Investment firms, credit institutions and other firms that:
Firms that are not authorised under MiFID II should take their own advice on whether they need authorisation, or can rely on an exemption, before trading on their own account. [Eligible jurisdictions, including the position of US persons, are subject to legal review.]
Membership enquiry
Tell us about your firm and the contracts you are interested in. We will send the information pack and explain the next steps.
Prefer email? membership@computemarkets.eu placeholder
09 FAQ
If your question is not covered here, include it in your membership enquiry.
A timed perpetual future works like a perpetual future. There is no monthly roll, and a daily funding payment between buyers and sellers keeps the futures price close to the Compute Desk index. Unlike a true perpetual, each series has a fixed final maturity five years after listing. At maturity, open positions settle in cash against the index, and a successor series is listed in advance so that positions can be rolled.
A timed perpetual tracks the current level of the index. It does not lock in a price for a future date.
Once per business day, the average futures price over a published funding window is compared with the most recent index value published before the funding window, normally the previous US business day’s 16:30 New York value. On days with no new index value, the last published value is used [to be confirmed in the Rulebook]. The difference, as a percentage of the index, is the funding rate, capped at ±0.50% indicative. Funding per contract is the rate × the daily settlement price × 5,840 GPU-hours. A positive rate means longs pay shorts, and a negative rate means shorts pay longs. Funding is exchanged between members and is settled as a cash adjustment separate from variation margin. See the worked example.
Both share the same contract unit, funding mechanism and five-year maturity. A cash-settled contract only ever settles in money, against the index. A physically delivered contract also allows holders to settle some or all of a position by delivery of real GPU capacity in monthly cycles: 730 consecutive hours of a verified 8-GPU node per contract, delivered and verified through Compute Desk COMPUTE Clear. Any positions remaining at maturity settle in cash.
The contracts reference the Compute Desk GPU rental price indices: DESKH100, DESKH200, DESKB200 and DESKB300. Each measures the US price of a one-year reserved rental of one NVIDIA GPU, in USD per GPU-hour, and is calculated daily. DESKH200, DESKB200 and DESKB300 are calculated and published by Compute Desk (The Compute Index, Inc.). DESKH100 is calculated and published by General Index. The methodology is published at compute-desk.com/index-methodology (opens in a new tab).
Because the indices measure US one-year rental prices, they can differ from the price of the capacity a member actually buys, sells or delivers. See the risk disclosure.
Investment firms, credit institutions and other firms that meet the published access criteria (good repute, trading competence, adequate organisation and sufficient resources) and qualify as professional clients or eligible counterparties under MiFID II.
Professional clients include, in summary, regulated financial firms; large companies meeting at least two of a EUR 20 million balance sheet, EUR 40 million net turnover and EUR 2 million own funds; governments, central banks and supranational institutions; other institutional investors whose main activity is investing in financial instruments; and clients who ask to be treated as professional and pass the required assessment.
Eligible counterparties are a narrower group of financial institutions and public bodies, such as investment firms, credit institutions, insurers, UCITS and pension funds and their managers, national governments and central banks. This summary is for information only, and your firm is responsible for confirming its own classification.
No. Compute Markets is for professional clients and eligible counterparties only. It is not intended for retail clients, and retail clients must not rely on this website.
The compute contracts are forthcoming. Their admission to trading is subject to regulatory approval, and no launch date can be confirmed at this stage. [Target launch window: to be confirmed.] If you ask for launch updates in the membership enquiry form, we will keep you informed.
[Fee schedule to be published.] Trading, membership, connectivity and any delivery-related fees will be set out in a published fee schedule before launch.
[Connectivity and API documentation to be published.] Admission includes access to a test environment and conformance testing of API connectivity, pre-trade risk limits, the kill switch and any trading algorithms.
Under the indicative terms, positions are subject to daily variation margin in USD, and daily funding is settled as a separate cash adjustment. The clearing and margining arrangement, including any central counterparty, is [to be confirmed] and will be set out in the Rulebook. The venue operator does not act as counterparty to members’ trades. COMPUTE Clear is a delivery and verification service, not a central counterparty.
Possibly. Firms that deal on their own account as members of an MTF may need authorisation under MiFID II unless an exemption applies, for example for trades that objectively reduce commercial risk or under the ancillary-activity exemption. Please take your own legal advice. We will check authorisation or exemption status during due diligence.
Important notice
This website and the services described on it are directed only at firms that qualify as professional clients or eligible counterparties under Directive 2014/65/EU (MiFID II). They are not intended for retail clients, and retail clients must not rely on them.
Compute Markets contracts are not yet available, and their admission to trading is subject to regulatory approval. Nothing on this website is an offer or solicitation in any jurisdiction where it would be unlawful, and it is not directed at persons in [excluded jurisdictions, subject to legal review].
