Index123.8
2026-08-21 04:53 UTC · 839 obs today · 16 providers
Indices

Four indices, four different questions

Compute is becoming a traded commodity: CME Group and ICE have both announced GPU futures, each settling against a different price index, and a third exchange is proposing its own. The indices do not agree, and they are not supposed to — each measures a different slice of the market. Here is what each one actually measures, what it costs to see, and where ours is weaker.

Why our H100 number is higher than theirs

Silicon Data publishes $2.53/GPU-hr for the H100; we publish $3.87. Both are correct. Theirs blends multi-year reserved contracts, which are far cheaper per hour, into one normalised rate. Ours is deliberately the on-demand price a person can click and rent today, with reserved pricing excluded because we cannot verify it consistently across providers. Our interruptible/spot median — $1.98 — brackets their blended figure from the other side. If you are hedging contracted supply, a blended index is the right tool. If you are about to rent a GPU this afternoon, ours is.

Silicon Data H100 Rental Index (SDH100RT)

Silicon Data
Measures
Listed prices normalised across rental term, cluster scale, interconnect and geography. On-demand, spot and reserved (1–60 months) are blended into one standardised rate.
Coverage
States ~150,000 daily pricing records across 50–100 platforms in 40–50 countries; data from September 2024.
Access
Free tier capped at 30 days of history with no filters; Pro $998/month; API only on Enterprise, sales-gated.
Exchange
CME Group — H100 and B200 futures announced for 5 Oct 2026, pending CFTC review
Method published
Not in full Describes a 'proprietary framework' for normalisation. Their published methodology post does not state whether the index is a mean, a median or weighted, nor how outliers are handled.

Bloomberg Terminal, Kaiko, own portal · source · verified 2026-08-17

Ornn Compute Price Index (OCPI)

Ornn
Measures
Printed transactions only — actual traded prices rather than listed rates. Described as the first compute index built solely from transactions.
Coverage
H100, H200 and B200 spot compute.
Access
Institutional; distributed via Bloomberg.
Exchange
ICE — cash-settled GPU compute futures announced, pending CFTC review
Method published
Not in full Transaction basis is stated publicly; the construction rules are not published for free.

Bloomberg Terminal · source · verified 2026-08-17

Compute Desk indexes

Compute Desk
Measures
Rental prices for H100, H200, B200 and B300.
Coverage
Four NVIDIA parts.
Access
Not publicly published.
Exchange
Architect Financial Technologies — American Innovation Exchange, a proposed US contract market, pending regulatory review
Method published
Not in full Not published publicly at time of checking.

Exchange-facing · source · verified 2026-08-17

SemiAnalysis H100 1-Year Rental Index

SemiAnalysis
Measures
One-year rental contract rates.
Coverage
H100.
Access
Paid research subscription.
Exchange
None
Method published
Not in full Described in subscriber research.

Subscription research newsletter · source · verified 2026-08-17

The Price of Compute Index

this site
Measures
On-demand list prices only, per single GPU per hour. Per-provider medians first, then the median across providers. Spot and community tiers are published separately and never blended in.
Coverage
20 parts across 16 providers, collected hourly, effectively US regions. Daily series back to January 2023 via labelled archival reconstruction.
Access
Free. Full history, open JSON API, no key, no login.
Exchange
None. We publish a measurement; we do not settle contracts and we do not publish forward curves.
Method published
Yes — to the arithmetic, with a changelog: methodology.

Where ours is weaker, plainly: we do not price reserved or committed contracts, which is where most large GPU spend actually sits; we are US-region; and we read listed prices rather than printed trades, so we measure what providers advertise, not what buyers finally paid.

Every claim on this page is taken from the operator’s or exchange’s own pages, dated at the time we checked. If we have described your index wrongly, tell us and we will correct it.