The compute spot index.

Methodology, in full.

Basket: ten fixed models spanning datacenter, workstation and consumer silicon, each with multi-provider coverage at inception:

Construction: per model, per half-hour snapshot, we take each provider’s median listed $/GPU/hr — medians, so no single strange listing matters. Between adjacent snapshots a model compares only providers present on both sides; the index step is the geometric mean of those model ratios, and levels chain from 100 at inception. A new feed joining the panel therefore cannot move the index — that is the difference between an index and a coverage counter. The headline uses on-demand listings only (firm prices); the same construction over interruptible listings is published as the idle sub-index, volatile by nature. Steps with fewer than five chainable models carry flat and are counted openly in the payload.

Revision policy: the basket and construction are frozen as v1; any change arrives as an announced version bump, never a silent edit.

Why now.

Compute is becoming a traded commodity: CME Group and Silicon Data list the first regulated compute futures on October 5, 2026 — contracts settling on proprietary H100 and B200 rental indexes. This index is the open counterpart: methodology in full view, every input on our public tape, per-model series for the contracts’ own underlyings readable straight from /api/history. When someone asks “what did GPU spot do this month”, the answer shouldn’t live behind a license.

Honesty box.

The index is young — its tape starts 23 Aug 2026, and the page you are reading will still say so when that sentence is embarrassing. It measures listed marketplace prices, not negotiated contracts or hyperscaler rates. An index over listing medians moves when cheap inventory is absorbed as well as when hosts reprice: both are real market signal, and we name both.