We logged every GPU provisioning run against Kilawatt Cloud's gateway over two days (Sep 6–7) — real submissions to real providers, measured from request to confirmed-ready state, then automatic teardown. No staged runs, no cherry-picking. Here's what actually happened.
The numbers
Provider Runs Median Fastest Slowest
RunPod 8 7.65s 5.8s 8.8s
Vast.ai 15 64.5s 20.7s 197.0s
Lambda 1 153.9s — —
RunPod: tight and predictable
Every single RunPod run landed inside a 3-second window — 5.8s to 8.8s, all on RTX 4090s. That consistency is the interesting part, not just the speed. A provider that's fast and predictable is easier to build reliability guarantees on top of than one that's fast on average but spiky.
Vast.ai: the spread tells its own story
Vast.ai's 15 runs ranged from budget cards (GTX 1050 Ti, GTX 1060, GTX 1660S) up to H200 and B200 — so the wide spread (20.7s–197s) isn't noise, it's mostly explained by card tier and host state:
• Warm hosts, high-end cards: H200 runs landed at 20.7s and 21.0s
• Cold-start penalty: one B200 run hit 197.0s because the host had to pull the container image first — the same card warm-started at 26.0s and 41.3s
• Budget cards, cold hosts: the GTX 1660S runs clustered around 121–128s
Cold start vs. warm host is roughly a 5-9x difference on the same hardware. Worth knowing if you're timing anything against a Vast.ai node cold.
Lambda: too early to call
One logged run at 153.9s. That's a single data point, not a trend — we're not going to dress it up as one. More runs needed before this is representative of anything.
Why this matters for automatic failover
Kilawatt's gateway routes across all four providers with automatic failover — if one is slow or unavailable, the job lands on the next. Numbers like these are what makes that routing decision meaningful instead of arbitrary: knowing RunPod is consistently fast lets the router lean on it for latency-sensitive jobs, while Vast.ai's breadth of card tiers makes it the better fit when price matters more than speed.
Full run log with every timestamp is available on request. Real over convincing.
kilawattcloud.dev
We post the raw runs as they happen, not just the summaries — follow @kilawattcloud on X if you want the numbers before they make it into a report like this one.
Top comments (0)