Where is the cheapest place to rent an H100 in the cloud?

By public on-demand prices, the cheapest cloud with stock right now is Hyperstack at $3.20/GPU·h. Out-of-stock low prices never rank — "$1.95 with zero cards rentable" is not a market price.

How much more expensive are cloud GPUs than marketplaces like vast.ai?

Clouds typically charge 3–10x more: you get SLAs, invoices and guaranteed stock, while marketplace machines come from individual hosts. The two calibers live in separate tables (/en/cloud vs /en/spread) and are never mixed.

Is RunPod or Lambda Labs cheaper?

It varies by model — check the /en/cloud table and each model page. On some models RunPod is ~10% lower, on others Lambda wins. Only in-stock quotes count.

Cloud GPU pricing · on-demand layer

Same GPU, cheapest cloud provider on demand

Where is an H100 cheapest to rent on demand: currently Hyperstack at $3.20/GPU·h (2 of 14 clouds have stock; table updated 2026-11-13). Cheapest in-stock cloud for every other model is ranked row by row below.

These are cloud providers' public on-demand prices (Azure / Lambda / Nebius / RunPod / Voltage Park…) — a different commodity from the marketplace listings on /spread: clouds charge 3–10x more but come with SLAs, invoices and guaranteed stock.

Model	In stock / all clouds	Cheapest (in stock)	In-stock median	Priciest (in stock)	Spread
A100 SXM 80GB nv.a100.sxm.80g	4 / 9	Excess Supply $1.35/GPU·h	$1.68	$3.28	143%
A6000 48GB nv.a6000.pcie.48g	2 / 7	Massed Compute $0.57/GPU·h	$1.25	$1.93	238.6%
A100 PCIe 80GB nv.a100.pcie.80g	2 / 7	Crusoe $2.00/GPU·h	$2.30	$2.60	30%
A100 SXM 40GB nv.a100.sxm.40g	2 / 6	Denvr Dataworks $1.40/GPU·h	$1.70	$1.99	42.1%
H100 SXM 80GB nv.h100.sxm.80g	2 / 14	Hyperstack $3.20/GPU·h	$3.65	$4.09	27.8%
NVIDIA V100 PCIE 16G nv.v100.pcie.16g	2 / 4	Verda $0.39/GPU·h	$1.36	$2.34	500%
H100 PCIe 80GB nv.h100.pcie.80g	2 / 8	Hyperstack $2.50/GPU·h	$4.25	$5.99	139.6%

Why only 7 models? The floor for ranking is "in stock at ≥2 clouds right now". A listed-but-out-of-stock price never ranks first — "$1.95/h with zero cards rentable" is not a market price, it's a trap. Out-of-stock prices still appear honestly on each model page; they just don't get first place.

Method: on-demand only (spot is a different pricing logic); every price normalized to $/GPU/hour; stock status is the source's current call and expires — every row carries its observation time. Marketplace (vast.ai / Clore.ai) device listings are a separate table: /en/spread. Machine-readable: GET /v1/cloud (free).