GPUDECK · PRICE CONTEXT · LICENSED SERIES, NOT A SCRAPE
Daily on-demand observations for full cards only — fractional cloud slices are excluded — split into marketplace (Vast.ai, RunPod, CloudRift, Verda) and cloud (AWS, Azure, GCP, OCI, Nebius, Lambda, Vultr) segments, because a median across both describes nothing a renter can actually buy. Cells with fewer than 5 observations are suppressed and counted. Cards marked ● measured also have workload goodput on the observatory.
Decentralized markets (Clore, OctaSpace) consistently trade materially lower than any column below and are absent by policy — their terms bar republication (§8).
| GPU | Marketplace p50 vast/runpod tier | Cloud p50 | Lowest observed | Trend (2026-08-15 →) | Sample |
|---|---|---|---|---|---|
| RTX5090 32GB ● measured | $0.61 n=24 | — | $0.37 | -6.1% | 24 obs · 2 providers |
| RTX4090 24GB ● measured | $0.56 n=14 | — | $0.33 | +18.1% | 14 obs · 3 providers |
| RTX3090 24GB ● measured | $0.22 n=9 | — | $0.13 | 9 obs · 2 providers | |
| RTX5070Ti 16GB ● measured | $0.39 n=5 | — | $0.13 | +7.4% | 5 obs · 1 providers |
| H100 80GB | $3.25 n=18 | $4.29 n=94 | $3.25 | -5.1% | 112 obs · 5 providers |
| RTXPRO6000 96GB | $2.09 n=13 | $2.76 n=382 | $0.97 | +0.0% | 395 obs · 5 providers |
| A100 80GB | $1.69 n=12 | $4.78 n=68 | $1.39 | -1.9% | 80 obs · 4 providers |
| H200 141GB | $4.00 n=12 | $4.50 n=6 | $3.59 | +1.5% | 18 obs · 4 providers |
| L40S 48GB | $1.37 n=11 | $2.04 n=16 | $0.99 | -2.0% | 27 obs · 3 providers |
| B200 180GB | $6.11 n=10 | $6.99 n=40 | $5.95 | +0.0% | 50 obs · 5 providers |
| A6000 48GB | $0.61 n=9 | $1.09 n=36 | $0.53 | +0.0% | 45 obs · 3 providers |
| A100 40GB | $1.29 n=8 | $1.99 n=110 | $0.30 | +0.0% | 118 obs · 5 providers |
| RTX6000Ada 48GB | $1.04 n=8 | — | $0.84 | +0.0% | 8 obs · 2 providers |
| RTXPRO6000S 96GB | $1.81 n=8 | — | $1.20 | +9.2% | 8 obs · 1 providers |
| RTXPRO6000WS 96GB | $1.11 n=7 | — | $0.82 | -20.1% | 7 obs · 1 providers |
| B300 268GB | $7.50 n=6 | — | $7.50 | +0.0% | 6 obs · 1 providers |
| L4 24GB | $0.49 n=5 | $1.26 n=712 | $0.25 | +0.0% | 717 obs · 4 providers |
| H100NVL 94GB | — | $9.07 n=54 | $2.76 | -1.9% | 56 obs · 3 providers |
| RTXPRO4500 32GB | — | $3.04 n=24 | $0.72 | +0.0% | 25 obs · 2 providers |
| T4 16GB | — | $1.22 n=568 | $0.53 | +0.0% | 568 obs · 2 providers |
| A10 24GB | — | $3.64 n=272 | $1.29 | -5.3% | 272 obs · 3 providers |
| A10G 22GB | — | $2.03 n=194 | $1.01 | +0.0% | 194 obs · 1 providers |
| L40S 44GB | — | $3.61 n=120 | $1.86 | +0.0% | 120 obs · 1 providers |
| P100 16GB | — | $1.27 n=88 | $1.27 | +0.0% | 88 obs · 1 providers |
| A10 12GB | — | $2.08 n=74 | $1.60 | +0.0% | 74 obs · 1 providers |
| A10 4GB | — | $0.59 n=74 | $0.45 | +0.0% | 74 obs · 1 providers |
| A10 8GB | — | $1.18 n=74 | $0.91 | +0.0% | 74 obs · 1 providers |
| GH200 96GB | — | $2.29 n=36 | $2.29 | +0.0% | 36 obs · 1 providers |
| RTX6000 24GB | — | $0.69 n=36 | $0.69 | +0.0% | 36 obs · 1 providers |
| A16 2GB | — | $0.06 n=14 | $0.06 | +0.0% | 14 obs · 1 providers |
| A16 4GB | — | $0.12 n=14 | $0.12 | +0.0% | 14 obs · 1 providers |
| A16 8GB | — | $0.24 n=14 | $0.24 | +0.0% | 14 obs · 1 providers |
| A40 4GB | — | $0.14 n=6 | $0.14 | 6 obs · 1 providers |
Scope, stated plainly. This page covers the 11 providers in the licensed tracker. Decentralized marketplaces such as Clore and OctaSpace are not here and will not be: their terms restrict republication, and GPUDeck never publishes a price it collected itself (methodology §8). If you rent there, expect lower figures than the marketplace column shows.
These are observations, not offers. The same model spans machines, regions and reliability tiers; the lowest observed figure is a floor someone quoted, not a price everyone gets. Price alone cannot rank hardware — divide by a measured goodput on the value board before concluding anything.
Take the data: gpu-price-context.csv carries the full derived series including both segment columns.