1–8 GPU Instances. Minute billing and no egress fees claimed. Stock not observed.
Source ↗A price list is not a reservation.
Choose the access terms your job needs. Live stock was not checked.
Start without a long commitment
First-come and on-demand offers keep the commitment short. They do not establish that your exact region has capacity.
Instant Clusters. Catalog statement, not an observation of available stock.
Source ↗On-demand clusters. Based on real-time availability.
Source ↗Hold capacity for a longer run
Read the payment term and the exact service commitment before planning a multi-week job.
16 × H100 · 2 weeks to 1 year. Plus applicable taxes. Different product from single GPU Instances; terms of one year or longer require a quote.
Source ↗Reserved clusters. Capacity locked for duration; quote must specify exact term.
Source ↗Reserved Clusters. No contract uptime target, remedy or regional stock captured.
Source ↗Use interruptible capacity deliberately
Ask how your job resumes, what happens to storage, and how the rate can change while it runs.
Interruptible instances. A vendor price description, not a measured savings claim.
Source ↗Preemptible GPU spot prices, except flat-rate L40S. Published floors are not quotes or upper limits.
Source ↗a3-highgpu-1g/2g/4g. Restriction is not stated here for 8 GPU variant.
Source ↗On-demand GPU offering. Does not establish capacity or a reservation guarantee.
Source ↗Need a full VM?
A saved Vast.ai discussion asks for custom ISO and persistent boot controls. Current support for those exact controls is still unknown here.
Read the feedback boundary →