Skip to content

Capacity reports

Once a month we publish a capacity report for each cluster. It is built from the telemetry feed and the billing ledger, and it is the record behind the numbers on the site. Reports are published on the 3rd business day of the following month.

What a report contains

Hours available: GPU hours the cluster had in the month, which is 8 GPUs times the hours in the month, minus announced maintenance.

Hours by plan: how many of those hours ran reserved, on demand and batch work, and how many sat idle.

Utilization: mean SM utilization across all GPUs for the month, and the mean during hours that had a job running.

Uptime: the availability figure used for the SLA, with every unavailable minute listed with its cause.

Queue: median and 95th percentile wait for on demand and batch jobs, in minutes.

Price comparison: our three plan prices next to the list prices at the four largest providers for a comparable H100 instance, as published on their pricing pages on the first of the month. We do not name the providers in the report.

Next cluster: the commissioning checklist for the next rack with each step marked done or not done.

Example, 2026-08

This is what a report looks like. C1 was commissioned on 2026-08-19, so the month is partial. The figures below are simulated and marked as such. preview

LineValue
Hours available2,496
Reserved1,152
On demand588
Batch431
Idle325
Utilization, all hours84.1 %
Utilization, busy hours91.7 %
Uptime99.94 %
Median on demand wait3 min
95th percentile on demand wait41 min
Median batch wait12 min

The preview figures follow the same shape as a real month. They are replaced with measured figures in the first report after preview ends.

Reading it with the API

GET /v1/capacity?report=2026-08 returns the report as JSON. The response has the same lines as the table above, plus the per day breakdown and the per minute unavailability list. See the API reference.

Why we publish it

A price only means something if you can see what it bought. The report shows how full the rack was, how long jobs waited and how often it was down. If the numbers get worse, you will see it here before you see it on an invoice.

RACK SYSTEMS, us-west. Console and API access is by invitation during preview.