GPU prices, per GPU-hour

What it costs to rent a GPU by the hour, and what the same card costs to buy outright. Node-priced offerings are divided by their GPU count, so every figure on this site is directly comparable. Each price links to the provider page it was read from.

As of September 8, 2026, the median on-demand price of an NVIDIA H100 SXM across the 20 providers we track is $3.97 per GPU-hour — see the Cloud GPU Price Index. Split by provider type, the median is $9.90 across the 4 hyperscalers (AWS, Azure, Google Cloud, Oracle) and $3.67 across 12 specialist GPU clouds — hyperscalers list the H100 at 170% more.

Cheapest listed on-demand rates today: NVIDIA H100 SXM $1.79 per GPU-hour on Cudo Compute (no-kristiansand-1, GPU only); NVIDIA A100 80GB $0.75 per GPU-hour on Vast.ai (floor); NVIDIA RTX 4090 $0.34 per GPU-hour on Vast.ai (floor).

Cheapest current rate by model

Split by hyperscaler, specialist cloud and marketplace, on-demand only: cheapest on-demand rate by provider type. Purchase price against rental: buy vs rent.

Lowest tracked per-GPU rate across all providers and price types.
GPU	VRAM	Cheapest / GPU-hour	Provider	Type	Retail, new
NVIDIA GB200 NVL72	186 GB	$6.22	Microsoft Azure	spot	—
NVIDIA GB300 NVL72	288 GB	$4.31	DataCrunch	spot	—
NVIDIA B300	288 GB	$3.75	DataCrunch	spot	—
NVIDIA H100 NVL	94 GB	$2.59	RunPod	on-demand	—
NVIDIA GH200	96 GB	$2.29	Lambda	on-demand	—
NVIDIA H200	141 GB	$2.00	DataCrunch	spot	$36,150
NVIDIA H100 PCIe	80 GB	$1.75	Hyperstack	reserved	$31,447.50
NVIDIA B200	192 GB	$1.64	Google Cloud	spot	—
NVIDIA H100 SXM	80 GB	$1.06	Google Cloud	spot	$29,950
NVIDIA RTX PRO 6000 Blackwell	96 GB	$0.94	DataCrunch	spot	$12,554.02
NVIDIA L40	48 GB	$0.69	RunPod	on-demand	—
NVIDIA L40S	48 GB	$0.69	DataCrunch	spot	$8,900
NVIDIA RTX 6000 Ada	48 GB	$0.52	DataCrunch	spot	$7,299
NVIDIA A100 80GB	80 GB	$0.43	Google Cloud	spot	$14,599.50
NVIDIA RTX 5090	32 GB	$0.38	Vast.ai	on-demand	$4,434
NVIDIA RTX 4090	24 GB	$0.34	Vast.ai	on-demand	$2,732.40
NVIDIA L4	24 GB	$0.33	Vast.ai	on-demand	—
NVIDIA A100 40GB	40 GB	$0.31	Google Cloud	spot	—
NVIDIA RTX A6000	48 GB	$0.30	DataCrunch	spot	—
NVIDIA V100	16 GB	$0.09	DataCrunch	spot	—

Providers

Provider	Regions	Models tracked	Notes
RunPod	—	17	—
Vast.ai	—	12	auction pricing, API only
DataCrunch	—	12	—
CoreWeave	—	10	Rate card priced per 8-GPU HGX node - parser stores per_node with gpu_count, DB derives per-GPU. GB200 and B300 on-demand are quote-only; B300 spot is published.
Hyperstack	—	9	—
Modal	—	9	per-second serverless
Lambda	—	8	Pricing page lists per-GPU rates for 1x/2x/4x/8x configs (old per-node note is stale). 1-Click Cluster reserved rates are contact-sales, not scraped.
Oracle Cloud (OCI)	—	8	Public price list API, already per GPU per hour, single global list price. Cloud@Customer SKUs skipped.
Paperspace	Paperspace, GPU Droplets	8	Two product lines at different prices, kept apart by the region label: Paperspace machines (docs pricing tables, H100 $5.95) and GPU Droplets (digitalocean.com, H100 $4.41 on-demand, $3.26 reserved 12mo). Nodes recorded per_node. AMD MI300X/MI325X and the P/Quadro series are not tracked.
Nebius	—	6	—
Jarvislabs	—	6	—
Amazon Web Services	—	6	EC2 on-demand, Linux, us-east-1 only; instance price stored per node with GPU count (p5 = 8x H100). Spot needs signed API calls - not recorded.
OVHcloud	—	5	US catalog in USD (L4/L40S); H100/A100/H200 come from the French catalog in EUR, shown with the ECB rate.
Scaleway	—	5	Prices in EUR (fr-par, pl-waw), read from the pricing page data; shown with the ECB daily rate. Not in the Price Index (USD listings only).
Google Cloud	—	5	Cloud Billing Catalog API (needs GCP_API_KEY). Plain per-GPU-hour SKUs in US regions, cheapest region kept; commitments/reserved/DWS skipped. GPU component only - A3/A2 VMs bill CPU and RAM separately. Some spot SKUs carry usageType OnDemand; type is decided from the description.
Microsoft Azure	—	5	Retail Prices API, Linux VMs; cheapest region per SKU kept, region = armRegionName. ND/NC series priced per VM, stored per node with GPU count. Spot recorded; Low Priority skipped.
GMI Cloud	—	4	—
Cudo Compute	—	3	Public machine-types API. gpuPriceHr is the GPU component only - vCPU/RAM/storage billed separately, a runnable instance costs more.
Together AI	—	3	—
Vultr	—	1	—
TensorDock	—	0	NOT SCRAPED: marketplace API returns an empty hostnode list after the Voltage Park acquisition (checked Sep 2026). Re-check later.
Crusoe	—	0	Blackwell quote-only
Latitude.sh	—	0	bare metal
Massed Compute	—	0	—
Salad	—	0	consumer GPUs only
Voltage Park	—	0	—
Sesterce	—	0	needs headless browser

GPU comparisons

H100 SXM vs H100 PCIe · H100 SXM vs H200 · H200 vs B200 · B200 vs B300 · GB200 NVL72 vs B200 · A100 80GB vs H100 SXM · A100 40GB vs A100 80GB · H100 SXM vs B200 · L40S vs RTX A6000 · RTX 4090 vs RTX PRO 6000 Blackwell · RTX PRO 6000 Blackwell vs L40S · RTX 4090 vs L40S

Provider comparisons

Two providers side by side on the GPUs both list — the cheaper one per GPU, with every price linked to its source.

Data updated: September 8, 2026