Vast.ai Observatory
How to read this
Vast.ai is the only per-host order-book marketplace in the panel — every individual machine listing is exposed via the API, with hundreds to thousands of distinct host offers per chip, repriced continuously by hosts. This page surfaces what the median alone omits: distribution shape, quality bands, $/perf frontier, supply structure, geography, egress and storage costs, network quality, long-tail catalog, and the time-series of the spread itself. Observatory data — not part of the CCIR rate methodology. Vast is the reference marketplace, not a tier.
Per-GPU $/hr histogram
— last snapshot · 2026-05-13 —Definitions & methodology
Every offer live on Vast at the snapshot for the selected chip, binned by per-whole-GPU
$/hr (log scale on x). The vertical amber line marks the median; the
dashed grey marks the median interruptible-bid floor. Where the
histogram has a long upper tail, that's the trust premium — buyers
pay 5-50× the floor for verified datacenter offers vs deverified
hosts. per_gpu_dph = dph_total / num_gpus / gpu_frac.
Verified vs deverified bands
— last snapshot · 2026-05-13 —Definitions & methodology
Median per-GPU $/hr for offers tagged verified (audited
datacenter hosts) vs deverified (hosted, unverified).
The premium varies sharply by chip — RTX 4090 verified is ~2× deverified
but H100 SXM verified can be cheaper than deverified.
Top 20 by DLPerf per $/hr
— —| $/hr | DLPerf/GPU | DLPerf/$/h | GPUs | Verified | Country | Reliability |
|---|
Definitions & methodology
DLPerf is Vast.ai's composite deep-learning benchmark — a synthetic mixed-precision matmul/conv score that proxies training throughput. Shown per GPU here so it lines up with the per-GPU $/hr column. DLPerf/$/h is bundle perf-per-dollar; higher is more performance per dollar of total spend on the offer (independent of unit choice — same ratio whether you compute per-GPU or per-bundle).
Multi-GPU offer structure
— last snapshot · 2026-05-13 —Definitions & methodology
Count of live offers by GPU count (1×, 2×, 4×, 8×, other) for the selected chip. Reveals whether the chip has scale-up supply or only single-GPU rentals — material for buyers planning multi-node training.
Country / region distribution
— —| Country / Region | Offers | Median $/hr |
|---|
Egress + storage costs
Per-offer distribution · last snapshot · 2026-05-13Definitions & methodology
Egress cost ($/TB) and storage cost ($/GB-mo) distributions across live offers, exposed per-offer in the Vast API. Egress ranges $0–$50/TB across the panel.
Internet-down distribution
— last snapshot · 2026-05-13 —Definitions & methodology
Bandwidth distribution is bimodal: consumer-hosted GPUs at gigabit-class, proper datacenter offers at multi-Gbps. Same chip, very different connectivity — matters for training jobs that pull from cloud storage.
Every chip Vast carries
—| Chip | Offers | Verified | Min | Median | Max | Spread× |
|---|
Distribution evolution over 30 days
— Tracks the chip selected at the top of the page —Definitions & methodology
Five lines show how the per-GPU $/hr distribution shifts over time: p10 (cheapest 10%), p25, p50 (median), p75, p90 (most expensive 10%). When spread widens, supply is fragmenting; when it narrows, the market is converging.
Collected every three hours alongside /marketplace while live; collection has been paused since 2026-05-13 pending the pipeline migration, and every figure on this page reflects the last published snapshot. Vast bundles API authenticated via CCIR research key — values are public list-asks as Vast publishes them. Distribution shape captured at the cycle Vast served the query; offers churn faster than this cadence so the histogram is a snapshot, not a long-window aggregate.