# A graded registry of GPU cloud prices and fabric disclosure

> Source: <https://computeassay.com/>
> Published: 2026-09-21 19:58:07+00:00

The delivery layer for compute

# The exchanges built a price for compute.

Nobody defined the product.

Cash-settled GPU futures are listing on four venues. All of them price the GPU-hour. None can deliver one — because a GPU-hour is not a fungible unit, and whether one cluster can substitute for another is a networking question the financial layer cannot answer. Compute Assay is the specification, registry, and weekly report that define the deliverable.

Substitutability is a *relation* between two fleets with respect to a workload class — not a property of a fleet. Eight H100s on oversubscribed Ethernet and eight in a rail-optimized non-blocking InfiniBand plane are identical goods for batch inference, different goods for large-model serving, and not comparable for frontier training. The indexes count chips. This describes deliverable capacity. — [ASSAY-1 §0.1](https://computeassay.com/spec.html)

## One chip. One week. 3.8× price spread.

This is not a ripoff chart — it is an undefined-product chart. The green dots publish enough about their network fabric to be underwritten; the red ones publish nothing. The spread is what an unwritten specification costs.

Excluded from the chart and spread: TensorDock's "starting at" $2.25; Voltage Park's "starting at" $1.99. Floor and teaser rates — where the bookable price is quote-gated or varies by host — are listed on the [H100-SXM page](https://computeassay.com/gpus/h100-sxm.html) with that label; they never anchor a headline number here.

[ASSAY-1 — the specificationThe record format, headline metrics (CSD·NBD·DCB·PFI), workload classes, substitutability grades, and verification harness that define compute as a deliverable good. Versioned like an RFC. Free forever.](spec.html)

[The Delivered Compute ReportWeekly. What GPU capacity actually exists, what it costs, and whether it can do the job the price implies. Every row sourced.](report.html)

[Compute has no kilowatt-hourThe frame: why compute is pre-kWh, and the three-step path from a unit of capacity to a unit of delivered intelligence.](frame.html)

[The Disclosure LeaderboardProviders ranked by disclosed training-class capacity and fabric transparency — what they state in writing, not measured performance. Re-ranked weekly from public sources.](https://computeassay.com/leaderboard.html)

[Run-cost calculatorWhat your training run costs this week, with a disclosed training-class band — from live tracked prices.](https://computeassay.com/calculator.html)

[Open dataThe full graded registry as JSON/CSV, free under CC BY 4.0. Every row sourced.](https://computeassay.com/data.html)

### Get the Delivered Compute Report

One email a week: headline prices, deliverability grades, one teardown, and the gap between the index and what you can actually take delivery of.

[Subscribe free](https://computeassay.substack.com/subscribe)

Live-API prices (28 listings: AWS, Azure, Oracle, Vast) refreshed **2026-09-10**; the other 105 from the weekly survey, most recently **2026-09-14**. Every row carries its own *as of* date — staleness is disclosed, never hidden.

## What an assay is

In every physical commodity market, an *assay* is the test that determines grade and purity — the reason a buyer in one port can trust a cargo loaded in another. Compute has never had one. Eight H100s on shared Ethernet and eight H100s in an NVLink domain with non-blocking InfiniBand sell under the same name at prices that differ by multiples, and the difference is written nowhere a contract can cite. ASSAY-1 is that test, for compute: published free, graded conservatively from what sellers state in writing, and revised in the open.