GPU VRAM Rate Monitor

Exact VRAM sizing for open-weight models (DeepSeek · Qwen · Mistral) matched against live, verified on-demand rates from RunPod and Vast.ai. Every number below is fetched from vendor APIs the moment this page was built — no cached markups, no invented discounts.

Live marketplace snapshot

Observed (UTC): 2026-11-18T14:11:56+00:00Snapshot SHA-256: a80cc0e6d59b…Freshness guard: auto-grays after 72h

Terminal speed-math — no install needed

Size any model from a real terminal, straight off this page's live snapshot:

curl -fsSL https://gpuvram.pages.dev/vram-cli.py | python3 - qwen3-8-27b --ctx 32768

Full live GPU rate table — RunPod vs Vast.ai

GPUVRAM (GB)RunPod Secure ($/hr)RunPod Community ($/hr)Vast.ai Verified ($/hr)Lowest ($/hr)Deploy
B300 288GB288$8.99$6.94—$6.94RunPodVast
MI300X 192GB192$2.39$0.50—$0.50RunPodVast
B200 180GB180$7.99$5.98—$5.98RunPodVast
H200 NVL 143GB143$3.79$0.50—$0.50RunPodVast
H200 140GB141$5.29$3.59—$3.59RunPodVast
H100 NVL 94GB94$3.19$2.59—$2.59RunPodVast
A100 PCIe 80GB80$2.09$1.19$0.87$0.87VastRunPod
A100 SXM 80GB80$2.19$1.39$0.52$0.52VastRunPod
H100 PCIe 80GB80$2.89$1.99—$1.99RunPodVast
H100 SXM 80GB80$4.39$2.69$2.27$2.27VastRunPod
A40 48GB48$0.59$0.35$0.47$0.35RunPodVast
L40 48GB48$0.82$0.69$0.46$0.46VastRunPod
L40S 48GB48$1.09$0.79$0.77$0.77VastRunPod
RTX 6000 Ada 48GB48$1.09$0.74$0.52$0.52VastRunPod
RTX A6000 48GB48$0.69$0.33$0.40$0.33RunPodVast
A100 SXM 40GB40—$1.00—$1.00RunPodVast
RTX 5090 32GB32$1.39$0.69$0.44$0.44VastRunPod
RTX 3090 24GB24$0.50$0.22$0.27$0.22RunPodVast
RTX 4090 24GB24$0.89$0.34$0.39$0.34RunPodVast
RTX A5000 24GB24$0.27$0.16—$0.16RunPodVast
RTX A4000 16GB16$0.25$0.17$0.09$0.09VastRunPod

— means no live on-demand stock from that channel at snapshot time. Vast rates are the lowest verified, rentable single-GPU on-demand bundle.

What standing still costs you

Run one H100 SXM 80GB around the clock (730 h/month): RunPod Secure bills $3204.70/mo, while the cheapest verified Vast.ai bundle bills $1656.30/mo — a real gap of $1548.40 every single month, or $18580.84 per year, for the identical GPU class.

By comparison, the hyperscalers' list rates for on-demand A100/H100-class instances (AWS p4d/p5, GCP A3) commonly land in the several-dollars-per-GPU-hour range with rigid 1-year terms — before egress fees. The community/verified spot channels above rent the same silicon by the hour with zero commitment. Check the live table before any monthly commitment; the gap is real and measurable.

FAQ

How is the KV cache computed? With grouped-query attention, only the KV heads carry cache: 2 (K and V) × hidden layers × kv heads × head_dim × context × sequences. We read layer/head counts straight from each repo's config.json on Hugging Face.

Why 16-bit weights for some models and not others? The calculator lets you pick FP16, FP8, or 4-bit; parameter counts come from the actual safetensors totals, so a 4-bit quant of a 32B model is sized from its true 32B params, not a marketing number.

Are these prices really live? Yes. Every rate in the table was returned by RunPod's GraphQL endpoint or Vast.ai's verified on-demand bundle API at the snapshot timestamp above, and the SHA-256 fingerprint of the raw response file is published so you can audit us.