Vast.ai offer search

Filters & search

GPU GPU models — select one or more; none = any

Loading models…

GPUs min max

VRAM/GPU ≥ GB total ≥ GB

Price

$/hr total ≤ $/hr per GPU ≤

Market

Serving traps Whole host only (gpu_frac = 1) Not power-capped (≥95% of model best) Free host ports ≥ (need ≥20 to expose a server) Measured mem bandwidth ≥ GB/s NVLink present (bw_nvlink > 0) No active display on GPU

Host quality Reliability ≥ 0–1 Verified datacenter only Rentable only (uncheck = full market, mostly unavailable) Min remaining uptime, days Region contains

Machine

Disk ≥ GB CPU RAM ≥ GB

vCPU ≥ CUDA ≥

Down ≥ Mb/s Up ≥ Mb/s

Static IP

Scan depth Grid — narrow queries see more of their range Passes — unions repeat scans; a weaker lever than grid

— — price per GPU Sort by

How this reads the API — and what it still can't promise

The row count tells you nothing about completeness. bundles/ caps responses near 64, answers "truncated": false while doing it, and rejects offset/page/skip. But the ceiling isn't even fixed — one identical query repeated 8× returned 62/62/64/64/64/64/62/64. And a response under the cap is still a rotating sample: over machine_id in [0, 6275) the API returned exactly 33 rows on 12 consecutive identical calls while the union of those 12 reached 58. The stable count is the trap — it reads as corroboration.
So the scan narrows instead of repeating. Same span: 1 wide call → 33 rows, 4 sub-ranges → 87, 16 → 98, 64 → 109; twelve repeats of the wide call only reached 58. This page scans a fine grid of narrow machine_id ranges, which is why Grid is the main quality dial and passes is the weaker one. Price bisection (the obvious idea) stalls on price ties; keyset pagination is exact about ordering but must stop on a short page, and a short page isn't complete.
rentable is ON by default, and you almost always want it. Only 4.0% of listings are rentable (369 of 9,289 scanned) — the rest are real hardware with no free capacity right now. Trying to rent one returns no_such_ask — not available. This page originally defaulted it off, on the reasoning that the excluded offers had higher median reliability (0.9949 vs 0.989) so the flag couldn't be about availability. That inference was wrong: a host is reliable and full at the same time. Uncheck it only for price research or a market census, where unavailable capacity still counts.
Coverage is a percentage, never a guarantee. A zero-result search means "nothing matched in this instant", not "this configuration doesn't exist" — and a result that stops growing has not thereby been confirmed. Raise the grid before believing a zero.
Power caps are published, and common. gpu_max_power is per-offer; hosts run cards well under stock TDP. A 250 W cap on a 350 W 3090 costs roughly a third of decode throughput, so the % column compares each offer against the best figure observed for that same model in these results.
gpu_frac < 1 means you share the host. The Slice column shows what you actually get, e.g. 2 of 8. Fractional is the common case, not the exception.
gpu_mem_bw is measured, not datasheet, and the ratio to spec differs per model, so BW% is relative to a healthy card of that same model and VRAM size rather than a hardcoded spec table that would go stale and miss variants.
Those reference figures use the mode / p90, not the max, because one outlier poisons a max. Across 1,343 RTX 4090 offers stock is 450 W (865 of them) but the highest seen is 600 W — scoring against that marked 92% of normal cards as capped. The mode gives 450 W and 30.6% capped. Cards above their reference are factory-overclocked and correctly show over 100%.