Exascale AI Index Methodology

The Exascale AI Index is a daily reference price for AI inference, expressed as the number of AI credits redeemable per one United States dollar at print time. The index is computed from observed transactions across the venue's constituent sub-credit markets, weighted by inference category, with statistically robust filtering for outliers and minimum-volume requirements.

The index is intended as the canonical reference for cash- or physically-settled derivatives on AI compute, including the Exascale AI-INDEX spot and forward contracts. Its design borrows from established commodity benchmarks (LBMA Gold Price, Brent Dated, CME Henry Hub) and from equity-index methodology (S&P, MSCI) where applicable to a continuously-traded, multi-constituent underlying.

This document is normative. Any divergence between this document and an operational implementation is to be treated as a defect in the implementation. The methodology committee (§12) is the authority of last resort on interpretation.

The index draws on six observable input streams, each capturing a distinct inference category. Five are spot prices on Exascale sub-credit markets; the sixth is a derived metric of weighted-average inference cost, computed from observed capacity utilisation across the GPU-credit underlying.

• Text credit spot. Median execution price of TEXT-INDEX trades over the calculation window.
• Speech credit spot. Median execution price of SPEECH-INDEX trades over the calculation window.
• Image credit spot. Median execution price of IMAGE-INDEX trades over the calculation window.
• Video credit spot. Median execution price of VIDEO-INDEX trades; thinner book, see §5.
• Niche credit spot. Aggregate of long-tail inference categories (embedding, classification, fine-tune).
• Inference cost average. Capacity-weighted USD-per-token cost derived from H100 and H200 GPU-credit utilisation; see Appendix A of the full whitepaper.

Constituent weights — effective 2026-04-01

Weights are set by the methodology committee at each quarterly review (§8). Weights reflect transaction-volume share over the preceding 90 calendar days, adjusted for inference-category representativeness. The current schedule is reproduced below.

The index value V(t) at print time t is the trimmed, volume-weighted sum of constituent observations across the calculation window [t − Δ, t]:

The trimmed mean operator excludes the lowest 2.5% and highest 2.5% of observations within the window before the volume-weighted sum is taken. The operator is applied jointly across all constituents (rather than per-constituent) so that genuine cross-asset moves are preserved while idiosyncratic spikes are suppressed.

A reference implementation in Python, with deterministic test vectors covering 18 published prints, is distributed with the JSON download above. The reference implementation is the authority for any computational ambiguity in the prose.

The 95% trimmed mean was selected after comparative back-testing against Huber M-estimation, MAD-based winsorisation, and the unmodified arithmetic mean across 2,194 simulated prints drawn from observed sub-credit volatility. The trimmed mean is the simplest estimator that meets two requirements: (i) a published, auditable cut-off; and (ii) bounded influence under a single-counterparty wash-trading attack up to 2.4% of window volume.

The 2.5% trimming threshold may be revised by the methodology committee with not less than 30 calendar days' notice. Any revision triggers a new methodology version (§8).

A valid print requires at least N = 240 observations across the calculation window, of which not fewer than 20 must originate from each constituent except video credit spot, which has a lower floor of 8 observations in recognition of its thinner book.

If the volume floor is not met at any constituent, the print is marked provisional and republished with the next-day reconciliation. If the floor is not met at the aggregate level, no print is issued; the previous print remains the prevailing reference and the methodology committee convenes within four business hours under standing protocol VP-01.

In the 18 months of pre-launch observation, the aggregate floor was met on 100.0% of business days; constituent floors were met on 99.6% of business days, with three provisional prints, each reconciled within one publication cycle.

The index incorporates the following defences against attempted manipulation. Each is documented in greater detail in the corresponding annex of the full whitepaper.

1. Cross-validation against multiple input sources. Each constituent's spot price is corroborated against the time-weighted mid-quote on the same market; large divergences trigger source isolation.
2. Outlier exclusion via trimmed mean. See §4. Bounded influence of any single counterparty's window activity.
3. Volume threshold for valid prints. See §5. A thin book defers to the prior print rather than admitting a thin observation.
4. Surveillance for marking-the-close patterns. Pattern matching against last-minute order-book pressure, with manual review for matches above the surveillance threshold.
5. Cryptographic audit chain on all prints. Each daily print is hashed and chained to the prior print; the chain root is published to a third-party transparency log within 60 seconds of issuance.
6. Member-trading restrictions. Methodology committee members are subject to a 5-day cooling-off window around methodology revisions and may not hold proprietary positions in any constituent during their tenure.

The Exascale AI Index is published once per ICE business day at 17:00 UTC. Publication latency is targeted at < 60 seconds from the close of the calculation window; the observed median latency in 2026-Q1 was 14 seconds.

A continuous intraday indicative value is also computed at five-second intervals and disseminated over the market data feed under the symbol AI-INDEX.IV. The intraday indicative is for reference only and does not constitute a published print.

Holiday and exceptional-event handling follows the schedule maintained at exascale.com/calendar. No print is issued on days where the venue is closed.

Methodology revisions follow semantic versioning with the following correspondence: major changes (formula structure, constituent set) carry a 90-day consultation period; minor changes (weight schedule, trimming threshold) carry a 30-day notification period; patch changes (typographical, clarification of intent) take effect at publication and do not re-rebase historical prints.

The methodology is reviewed quarterly by the Exascale Methodology Committee (§12) and audited annually by an independent third-party benchmark administrator. The audit covers (a) conformity of the operational implementation to this document, (b) integrity of the cryptographic audit chain, and (c) governance practices of the methodology committee.

The audit chain is rooted in a third-party transparency log operated by the Cloudflare Merkle Town service (RFC 6962). Independent verifiers may reconstruct the entire history of prints from the published audit roots and the per-print JSON manifests distributed under data.exascale.com/index/v1/.

The chart below renders the full series of daily index prints from launch (2025-10-15) to the most recent business day. The index was normalised to a value of 1.0000 at launch; the most recent print is 1.0024, a year-to-date change of +4.18%.

Prints prior to launch are not included; the consultation-period pilot (versions 0.9 — 0.9.4) was a paper exercise and did not include settlement. Researchers requesting the pilot series should contact methodology@exascale.com.

Per-constituent observed prices, window-volume figures, and exclusion counts are published with each print under the JSON manifest. The manifest is fully signed and chained to the audit log described in §9, and is downloadable on a five-second delay over the public data API.

Subscribers to the institutional data feed receive the full per-fill detail in real time, including taker / maker flags, counterparty bucket (member / non-member), and the maker–taker contribution to each constituent's window-volume.

The methodology committee is responsible for (a) the quarterly review of constituent weights, (b) extraordinary reviews triggered under the volume-floor protocol (§5), and (c) approval of any methodology revision. Membership is mixed internal and external; the external chair holds a casting vote.

Committee minutes for non-confidential portions of each session are published at exascale.com/methodology/committee/minutes within ten business days of the session.

Any market participant, vendor, or member of the public may submit a complaint regarding the integrity of a print, the conduct of the methodology committee, or the operation of the audit chain. Complaints are received at complaints@exascale.com and acknowledged within two business days.

Substantive complaints are investigated within ten business days. Where the investigation cannot be completed within the standard window, the complainant is notified of an extended timetable. Resolution outcomes are published in anonymised form at exascale.com/methodology/complaints.

General methodology enquiries: methodology@exascale.com. Audit-chain and reconciliation enquiries: audit@exascale.com. Media enquiries should be routed through press@exascale.com.

Exascale Markets · 200 Pine Street · San Francisco, CA 94104 · United States. For corporate registration and counterparty diligence, see exascale.com/legal/entity.