Skip to content

Reserved capacity

Reserved capacity is a block of GPUs held for your team for a fixed term. It is how a lab gets a rack for a training run without waiting in a queue or watching the on demand price. This page covers terms, what you get, and how to sign.

Terms

Minimum term is 1 month. Maximum is 12 months. GPU count is 1 to 8 on C1. The price is 1.89 USD per GPU hour, fixed for the term, invoiced at signature and prorated to the day for partial months. A full rack for one month is 8 × 24 × 30 × 1.89, which is 10,886.40 USD for a 30 day month.

Terms renew only if you ask. We send a note 14 days before the end. If you do not renew, the GPUs return to on demand and your shared storage stays for 7 more days.

What you get

The GPUs are yours. Jobs on them start immediately. Nobody else is scheduled on them while the reservation is active, except batch jobs when your GPUs are idle, and those are preempted with 60 seconds of notice the moment you submit.

You can release idle reserved GPUs to batch in the console. For each hour a released GPU runs batch work, your next invoice is credited 0.60 USD. You can take the GPU back at any time.

Reserved teams get the SLA with 99.5 % monthly availability, credits on breach, a 4 hour support response and a phone number for hardware incidents.

What it is good for

A training run that needs 8 GPUs on one NVLink fabric for weeks. A production inference endpoint that must not wait in a queue. A lab that wants a predictable monthly number for its compute budget. A team that wants to run batch evaluation for free in the gaps of its own reservation.

What it is not good for

Bursty work that needs 8 GPUs one day and none the next. Use on demand. Work that can wait and tolerate restarts. Use batch, at less than half the on demand price.

How to sign

Request access and choose the hours per month you expect. We reply within two business days with available start dates. The agreement is 3 pages. Once signed, the reservation appears in your console and the API within one business day.

Payment is by bank transfer or card, in USD, at the start of the term. Multi month terms can be invoiced monthly on request.

Beyond one rack

If you need more than 8 GPUs, tell us. The next cluster is listed when C1 is sold out, and a signed reservation for the next cluster is how it gets sold out. See regions.

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