What a hedge would have done

A buyer of 1,344 GPU-hours a week, 2025-04-14 to 2026-11-02, at a clearing price equal to the anchor (modelled; nobody traded these weeks).

What hedging changed, for a buyer of 1,344 GPU-hours a week

The surprise is the part of a week’s bill not known when the auction opened. All figures are modelled at a clearing price equal to the anchor; the sensitivity table below moves it.

Not covered: 2026-06-01 (average $3.250, band $2.256 to $3.050). Above the cap the buyer paid the market again.

The rule

Index

H100 on-demand composite, USD per GPU-hour

Data

2025-04-01 to 2026-11-10; 536 days with an index value, 204 without (never filled)

Weeks

69 kept of 82 candidate Mondays; 13 dropped (8 stale anchor, 5 few prints)

Band

15% of the anchor either side. anchor = last daily index print; mid = anchor rounded down to tick; half = anchor * the band's share rounded down to tick; floor = mid - half; band = 2 * half; cap = floor + band

LONG price p

LONG price p = half, so the forward floor + p equals mid (the anchor on the tick grid). This is modelled: nobody traded these weeks.

Cost

unhedged weekly cost = fixing * the hours used a week (the buyer pays the index every hour, so the week's average price is the fixing); hedged = (fixing - LONG payout + p) * the hours used a week Usage: 1,344 GPU-hours a week.

Not modelled

no real order flow; asks and rate cards, not trades; sample length; provider survivorship and composition; gaps; daily prints, not the on-chain cadence; usage and fees; hyperscalers are excluded from the index. Caveats in full

Coverage by band width

If the auction had cleared elsewhere

Every week

What this does not show

• No real order flow. Nobody traded these weeks. The clearing price is assumed (at the anchor, or shifted ±2% and ±5%); a real auction could clear further away, or not at all in a quiet week.
• Asks and rate cards, not trades. The index is the median of providers' posted on-demand prices. Posted prices are sticky: a provider can keep a list price for months while discounts, availability and negotiated deals move underneath. Low measured volatility partly reflects that stickiness, not a calm market. Where it overlaps (91 days from 2026-08-12) the composite sits on average 28% above Ornn's transaction-based H100 index (data/derived/validation.json): this is a list-price index, and a fixing on transaction prices would sit elsewhere and might move differently.
• Sample length. The sources reach back to 2024-11-01, but before 2025-04-01 fewer than 4 market providers are in the data, so the index (and the backtest) covers 2025-04-01 to 2026-11-10: 69 usable weeks, one gentle decline and one sharp rise. The 2024 to early-2025 price fall is not in the sample. Tail estimates (the 95% band) rest on a handful of weeks.
• Provider survivorship and composition. The composite uses whoever is in the sources on a day: the set changed on 243 index days. Providers enter (Akash from 2025-07, Cudo and Vast.ai from 2026-01, others in 2026-10) and change their SKU mix; the median can step when the set changes, without any provider changing price. Providers that left the market before or during the window are not represented.
• Gaps. Days without at least 4 providers have no index value and are never filled. Weeks with fewer than 4 daily values or a stale anchor are dropped, and dropped weeks may not be random (scrapes fail more often when things change).
• Daily prints, not the on-chain cadence. The contract may print more often than once a day; one daily snapshot is a coarse stand-in, taken at whatever hour the source ran.
• Usage and fees. Flat usage of 8 GPUs every hour; no auction fees or forfeits, no cost of locked capital. Weeks with too few prints are dropped, not settled as void (spec 0.5 unwinds a void series at its auctions' average traded price, and there are no trades here to compute one).
• Hyperscalers are excluded from the index. and appear only in the basis-risk table.

Figures from Weir’s backtest of its own index on public provider prices, as of Fri 13 Nov, 08:08 UTC, when this page was published; the rule is set out above.