Cloud GPU Pricing Compared

What AWS, Azure, Google Cloud, RunPod, Lambda and CloudRift charge per GPU-hour, read from each provider's own price sheet and divided down to the single card. On demand, Linux, US regions, no commitment.

Provider prices read October 15, 2026. CloudRift column live, read November 14, 2026 at 4:49 PM UTC.

What the Big Clouds Charge per GPU-Hour

One number per cell, with the instance shape it comes from underneath. Multi-GPU shapes are divided by their GPU count, and the lowest on-demand rate in each row is marked. Where a provider has no on-demand shape for a card, the cell says so instead of guessing.

On-demand price per GPU-hour by GPU and provider, read 2026-10-15
GPU	AWS	Azure	Google Cloud	RunPod	Lambda	CloudRift
NVIDIA H10080 GB	$6.88p5.48xlarge8 GPUs	$12.29ND96isr_H100_v58 GPUsNVL 94 GB single GPU: $6.98	$11.06a3-highgpu-8g8 GPUs	$3.99lowestH100 SXMPCIe: $2.89	$3.99lowestH100 SXM8 GPUssingle GPU: $4.29	not offered
NVIDIA H200141 GB	$7.91p5en.48xlarge8 GPUs	$10.60ND96isr_H200_v58 GPUsEast US 2	$10.60a3-ultragpu-8g8 GPUs	$5.29lowestH200 SXM	not offered	not offered
NVIDIA B200180 GB	$14.24p6-b200.48xlarge8 GPUs	$27.04ND128isr_NDR_GB200_v64 GPUsGB200 NVL72 slice	$8.06a4-highgpu-8g8 GPUsFlex-start only, no on-demand price	$7.99B200	$6.69lowestB200 SXM68 GPUssingle GPU: $6.99	not offered
NVIDIA A10080 GB	$3.43p4de.24xlarge8 GPUs	$4.10ND96amsr_A100_v48 GPUsPCIe single GPU: $3.67	$5.07a2-ultragpu-1g	$1.79lowestA100 SXM	$2.79A100 SXM8 GPUs	not offered
NVIDIA A10040 GB	$2.74p4d.24xlarge8 GPUs	$3.40ND96asr_v48 GPUs	$3.67a2-highgpu-1g	not offered	$1.99lowestA100 PCIe	not offered
NVIDIA L40S48 GB	$1.86g6e.xlarge8 GPUs: $3.77 each	not offered	not offered	$1.09L40S	not offered	$0.69lowest48 GB
NVIDIA V10016 GB	$3.06p3.2xlarge	$3.06NC6s_v3	$2.48N1 + V100GPU only, VM billed on top	not offered	$0.79Tesla V1008 GPUs	$0.25lowest16 GB $0.25, 32 GB $0.28
NVIDIA T416 GB	$0.53g4dn.xlarge	$0.53NC4as_T4_v3	$0.35lowestN1 + T4GPU only, VM billed on top	not offered	not offered	not offered
AMD Instinct MI300X192 GB	not offered	$6.00ND96is_MI300X_v58 GPUsEast US 2	not offered	not offered	not offered	$4.00lowestMI350X 288 GB, newer generation

AWS: EC2 on-demand pricing, us-east-1, EC2 on-demand.
Azure: Azure Retail Prices API, East US (East US 2 for H200 and MI300X), pay-as-you-go.
Google Cloud: Accelerator-optimized VM pricing, us-central1 (Iowa), on-demand.
RunPod: RunPod pricing, Secure Cloud pods, page dated October 8, 2026.
Lambda: Lambda GPU Cloud, on-demand instances.
Oracle sells the MI300X at $6.00 per GPU-hour on BM.GPU.MI300X.8 (launch announcement). CloudRift sells the newer MI350X; the row compares the nearest card.
CloudRift: live rates, the same as the pricing table.

The Same Card on CloudRift

Rates, instance sizes and regions below come from the CloudRift API each time this page is rebuilt. Availability per instance size changes by the hour, so it lives in the console rather than here.

CloudRift on-demand GPU rates, instance sizes and regions
GPU	On Demand	Cheapest Elsewhere	Per Instance	Regions
NVIDIA V100 SXM216 GB	$0.25per GPU-hour68% below Lambda	Lambda $0.79 in 8-GPU boxes, AWS and Azure $3.06	up to 8 GPUs	United States
NVIDIA V100 SXM332 GB	$0.28per GPU-hour65% below Lambda	Lambda $0.79 in 8-GPU boxes, AWS and Azure $3.06	up to 16 GPUs	United States
NVIDIA RTX 409024 GB	$0.45per GPU-hour	RunPod $0.89, community tier $0.34	up to 5 GPUs	Taiwan
NVIDIA RTX 509032 GB	$0.65per GPU-hour45% below RunPod	RunPod $1.19, community tier $0.69	up to 2 GPUs	South Korea, United States
NVIDIA L40S48 GB	$0.69per GPU-hour37% below RunPod	RunPod $1.09, AWS $1.86	up to 8 GPUs	Taiwan
NVIDIA RTX PRO 600096 GB	On requestreserved termsfrom $1.76 per GPU-hour	RunPod $2.49	up to 2 GPUs	United States
AMD MI350X288 GB	$4.00per GPU-hour33% below the MI300X elsewhere	MI300X on Azure and Oracle $6.00	up to 8 GPUs	Italy

Billed per second

No minimum spend

No egress charges

Elsewhere means the lowest dedicated-host on-demand rate for the same card among the providers in the first table, and the percentage is measured against it. Where RunPod's shared community tier is lower than our rate, no percentage is shown and the column says so.

Cost per Token Changes the Ranking

An hourly rate says little until you know what the card does in that hour. These are the published benchmarks behind the comparisons above, with the pricing basis each one used. We have no V100 or MI350X benchmark yet, so those rows stand on price alone.

RTX PRO 6000 Beats the H100 on a Single GPU

28% lower cost per token on single-GPU inference. Once a model needs 8-way tensor parallelism over NVLink, the H100 and H200 pull 3 to 4 times ahead. That test priced the cards at RunPod rates, not ours.

Read the benchmark

RTX 4090, 5090 and PRO 6000 on Our Own Rates

On a 24 GB-class model and one GPU: $0.048, $0.040 and $0.043 per million tokens. The 5090 is the cheapest per token at that size, and the PRO 6000 takes over as models grow.

Read the benchmark

B200 Leads on 16K-Context Inference

Up to 4.9 times the throughput of the RTX PRO 6000 and the best cost efficiency in the set. That comparison used estimated run costs rather than cloud list prices, which is why it is not in the table above.

Read the benchmark

What the List Price Leaves Out

Why two honest comparisons of the same GPU can disagree by a factor of two.

Commitments: Google's 8-GPU H100 a3-megagpu-8g is $93.40 an hour on demand, $64.21 with a one-year commitment and $40.65 with three. AWS and Azure sell the same kind of one- and three-year terms.
Shapes You Cannot Buy on Demand: Google sells the B200 only through its Dynamic Workload Scheduler, and its single-GPU H100 shapes have left the on-demand list. At the hyperscalers the eight-GPU box is increasingly the unit.
Spot and Preemptible Rates: Google's spot price for that same 8-GPU H100 shape was $50.44 an hour on the day we read it. Spot capacity can be taken back at any time, so it suits batch work, not serving.
The Bundle Around the GPU: A hyperscaler GPU price buys a fixed set of vCPUs, memory and local disk with it. Google's attachable V100 and T4 prices exclude the N1 virtual machine they attach to, so the real hourly bill is higher.
Egress and Minimums: The hyperscalers bill data leaving the region per gigabyte on top of the GPU. CloudRift bills per second of runtime, with no minimum spend and no egress charges.

Pricing FAQ

Questions About Cloud GPU Prices

On October 15, 2026: $6.88 per GPU-hour on AWS (p5.48xlarge, us-east-1), $12.29 on Azure (ND96isr_H100_v5, East US) and $11.06 on Google Cloud (a3-highgpu-8g, Iowa), all on demand for the eight-GPU shape. Azure's single H100 NVL shape is $6.98. RunPod and Lambda list the H100 SXM at $3.99.

Among the providers on this page, the V100 on CloudRift at $0.25 per GPU-hour, below the hyperscalers’ T4 at $0.35 to $0.53. For current-generation cards, the RTX 4090 at $0.39 and the RTX 5090 at $0.65 on CloudRift, or RunPod’s community tier, which lists the same cards from $0.34 and $0.69 on shared hosts.

The hyperscalers fold networking, storage, support and committed-use discounts into a higher list price and sell most accelerators in eight-GPU units. Specialist clouds run leaner, buy consumer and workstation cards the hyperscalers do not stock, and price per GPU. The same H100 ranges from $3.99 to $12.29 per hour on this page.

Yes. Every number is the pay-as-you-go Linux price with no commitment: AWS us-east-1, Azure East US (East US 2 for the H200 and MI300X, which East US does not sell), Google Cloud us-central1. Spot, reserved and committed-use rates are lower and noted where we read them.

CloudRift sells the same V100 and L40S cards at $0.25 to $0.63 per GPU-hour against $1.86 to $3.06 at the hyperscalers, and consumer-class RTX cards they do not offer at all. It does not sell the H100, H200 or B200 on demand. Current availability per instance size is in the console.

The provider columns were read on October 15, 2026 and are re-checked quarterly; the date is printed at the top of the page. The CloudRift column comes from our API on every rebuild and is at most five minutes old.

Live availability

See What Is Available in the Console

The console shows the same hourly rates as this page with current availability per instance size. Reserved terms for the cards above are quoted on request.

Open Console See All Rental Prices