The market quotes the price of compute in four denominations — per megawatt, per GPU-hour, per million tokens, per task — and they do not convert, because each step between them hides a factor nobody publishes under controlled conditions. This surface states each rung honestly instead of collapsing the chain into one number.
Three rungs are published as reference. The fourth is declined, in public, with the reasons. Nothing on this surface is an assessment: no rung prints a daily value and no conversion factor is published with a level.
What this is. The measurement frame, stated rung by rung: what each denomination prices, what The Grid publishes on it, and what the step to the next rung hides. Where The Grid holds data, the rung shows it with source, method, observation count and date. Where it holds none, the rung says so — an empty rung is stated, never faked.
What this is not. Not a set of assessments, not a unified $-per-anything number, and not a conversion calculator. Each arrow between rungs embeds a factor (power density, throughput for a named model under named conditions, tokens per task) that varies enough to swallow the comparison — publishing a single collapsed number would present a modelling choice as a property of the market.
The declined rung is part of the method. A price for an undefined deliverable is not a price. The Grid publishes the declination of the $/task rung with the reasons, and holds it until the market defines the unit — see rung 4.
| Rung | Denomination | Status | What The Grid publishes |
|---|---|---|---|
| Rung 1 | US$ per MW. What a megawatt of AI datacentre capacity costs to hold or rent. The denomination the power and datacentre layer quotes in — interconnection queues, leases, builds — and the rung furthest from any price The Grid collects. | Reference — stated, not assessed | Reference and explainer only. The Grid collects no capacity-cost-per-megawatt series; the empty rung is stated plainly. The conversion arithmetic renders with cited public inputs; worked conversions from stored Grid data are labelled worked conversions, never prints. |
| Rung 2 | US$ per GPU-hour. What an hour of a defined accelerator class costs on a bookable ask. This rung is the existing GAP-H100 series — the assessed spot leg of the term structure, specified at /standards/instruments/gap-h100. | Existing instrument — see the specification | No new number and no duplication: one definition stays one definition. This page links to the specification and the live surface; it republishes nothing. |
| Rung 3 | US$ per million tokens. What a million tokens costs at a provider's published list price. The denomination hosted inference is sold in — and the rung where model choice starts to hide inside the number. | Reference — stated, not assessed | A reference table of the stored token_pricing rows (OpenRouter + DeepInfra documented public APIs, one request per endpoint per daily cycle; source ratified by founder ruling 2026-08-20). Published list prices as recorded — not an assessment: no adjustment, no judgment, no panel. Source, method, observation count and date ride with the table. |
| Rung 4 | US$ per task. What a unit of completed work costs. The denomination every buyer actually wants — and the one the market has not defined. | Declined — published with the reasons | Declined, in public, with the reasons stated (notice 2026-08-20 §3). The rung is declined until the market defines the unit; a standardised, independently governed task definition reopens the question by notice. |
| Step | Factor | What it hides | Where it stands |
|---|---|---|---|
| 1 → 2 | Power density | GPUs per megawatt, inclusive of server overhead and PUE — derivable from public vendor specifications and facility figures, contested in practice. | Stated as arithmetic with cited inputs on the rung-1 page; not published as a factor. |
| 2 → 3 | Throughput | Tokens per second for a named model at named precision, batch and context. Hardware- and deployment-specific. | Anchor of record: MLPerf® Inference (MLCommons) — audited, controlled-conditions, hardware-specific. The Grid republishes no result figures; official results live at mlcommons.org. |
| 3 → 4 | Tokens per task | How many tokens a model spends to complete the same work — varies on the order of 2× between models, unpublished under controlled conditions. | Measurement exists behind a commercial licence The Grid has not purchased: [gap: licence pending]. Rung 4 itself is declined. |