What we’re measuring: time from submitting a provisioning request to a genuinely usable instance (SSH actually reachable, not just an API saying “running”). Every test instance is torn down immediately after, so cost per run stays near zero.
Round 1 — cheapest available GPU:
• RunPod (RTX 4090): 7.2s-8.0s across 6 runs
• Vast.ai (mixed cards): 38.6s-575.0s across 7 runs, high variance since Vast is a peer-to-peer marketplace of many different hosts
Round 2 — pinned to the actual hardware our customers rent (H200, B200) on Vast.ai:
• H200: 93.1s cold start, then 21.0s and 20.7s once the host was warm
• B200: 197.0s cold start, then 41.3s and 26.0s once warm
The pattern we found interesting: the first run on any given host pays a real tax for the container image download. Once cached, subsequent runs land consistently in the 20-40s range regardless of GPU tier.
Sample size is still small (3 runs per GPU on the hardware-tier tests), so treat this as directional, not final. We’re running more to see if the cold-start pattern holds up. Full methodology and charts at kilawattcloud.dev if you want to dig in.
Top comments (0)