GPU cloud pricing
on demand, per GPU-hour.
List prices for renting NVIDIA GPUs with no commitment, normalized to dollars per GPU-hour so an 8-GPU instance and a single-GPU pod can sit in the same column. Every figure is from the provider's own pricing page or price API, with sources and billing rules.
Facts checked against each provider's own pages
The short answer
For an H100, on-demand list prices range from $2.69 (RunPod, Community Cloud pod)[1] to $12.29 (Microsoft Azure, Standard_ND96isr_H100_v5)[2] per GPU-hour. Specialist GPU clouds and serverless platforms price far below the hyperscalers' 8-GPU instances, and the cheapest rows are RunPod Community Cloud, which connects individual compute providers to users through a peer-to-peer system[3].
The per-hour number is only half the cost. Billing granularity decides what idle time costs: Modal[4] and RunPod[1] bill per second, AWS per second with a 60-second minimum[5], Google Cloud per second after a 1-minute minimum[6], and DigitalOcean per second with a 5-minute minimum[7]. Google Cloud's accelerator-optimized prices are for the whole machine, including its vCPUs, memory and local SSD[8], so a per-GPU figure derived from them is not a GPU-only price.
On-demand price per GPU-hour
Prices are on-demand list prices in USD with no commitment, for the region shown. Where a provider sells a multi-GPU instance, the per-GPU figure is the instance price divided by its GPU count and is marked "derived". Per-second prices are multiplied by 3,600. Rows are sorted cheapest first within each GPU.
| GPU | Provider | Product | GPUs | $/unit-hour | $/GPU-hour | Region |
|---|---|---|---|---|---|---|
| B200 | RunPod | Community Cloud pod[1] | 1 | $5.98 | $5.98 | Not stated (31 regions) |
| Modal | Serverless GPU function[4] | 1 | $6.25 | $6.25derived: $0.001736/s × 3,600 | Base price; pinning a region costs 1.15–1.75× | |
| RunPod | Secure Cloud pod[1] | 1 | $7.99 | $7.99 | Not stated (31 regions) | |
| CoreWeave | NVIDIA HGX B200 instance[9] | 8 | $68.80 | $8.60derived: $68.8 ÷ 8 | North America (Europe lists the same on-demand price) | |
| RunPod | Instant Cluster (no commitment)[1] | 1 | $9.59 | $9.59 | Not stated (31 regions) | |
| AWS | p6-b200.48xlarge[10] | 8 | $113.93 | $14.24derived: $113.9328 ÷ 8 | US East (N. Virginia) | |
| H200 | RunPod | Community Cloud pod[1] | 1 | $3.59 | $3.59 | Not stated (31 regions) |
| DigitalOcean | GPU Droplet (HGX H200), per GPU[7] | 1 | $4.47 | $4.47 | Not stated on pricing page | |
| Modal | Serverless GPU function (H200 SXM)[4] | 1 | $4.54 | $4.54derived: $0.001261/s × 3,600 | Base price; pinning a region costs 1.15–1.75× | |
| RunPod | Secure Cloud pod[1] | 1 | $5.29 | $5.29 | Not stated (31 regions) | |
| CoreWeave | NVIDIA HGX H200 instance[9] | 8 | $50.44 | $6.31derived: $50.44 ÷ 8 | North America (Europe lists the same on-demand price) | |
| RunPod | Instant Cluster, H200 SXM (no commitment)[1] | 1 | $6.35 | $6.35 | Not stated (31 regions) | |
| RunPod | Serverless flex worker (H200 tier)[1] | 1 | $6.99 | $6.99 | Not stated (31 regions) | |
| AWS | p5en.48xlarge[10] | 8 | $63.30 | $7.91derived: $63.296 ÷ 8 | US East (N. Virginia) | |
| Microsoft Azure | Standard_ND96isr_H200_v5[2] | 8 | $84.80 | $10.60derived: $84.8 ÷ 8 | East US 2 (no East US price listed) | |
| Google Cloud | a3-ultragpu-8g (whole machine)[8] | 8 | $84.81 | $10.60derived: $84.806908493 ÷ 8 | Not stated on page (no region selector for this table) | |
| H100 SXM / HGX | RunPod | Community Cloud pod[1] | 1 | $2.69 | $2.69 | Not stated (31 regions) |
| Modal | Serverless GPU function (H100 SXM5)[4] | 1 | $3.95 | $3.95derived: $0.001097/s × 3,600 | Base price; pinning a region costs 1.15–1.75× | |
| RunPod | Secure Cloud pod[1] | 1 | $4.39 | $4.39 | Not stated (31 regions) | |
| DigitalOcean | GPU Droplet (HGX H100), per GPU[7] | 1 | $4.41 | $4.41 | Not stated on pricing page | |
| RunPod | Instant Cluster (no commitment)[1] | 1 | $4.79 | $4.79 | Not stated (31 regions) | |
| CoreWeave | NVIDIA HGX H100 instance[9] | 8 | $49.24 | $6.16derived: $49.24 ÷ 8 | North America (Europe lists the same on-demand price) | |
| AWS | p5.48xlarge[10] | 8 | $55.04 | $6.88derived: $55.04 ÷ 8 | US East (N. Virginia) | |
| AWS | p5.4xlarge[10] | 1 | $6.88 | $6.88 | US East (N. Virginia) | |
| Google Cloud | a3-highgpu-8g (whole machine)[8] | 8 | $88.49 | $11.06derived: $88.490000119 ÷ 8 | Not stated on page (no region selector for this table) | |
| Google Cloud | a3-megagpu-8g (whole machine)[8] | 8 | $93.40 | $11.68derived: $93.400712807 ÷ 8 | Not stated on page (no region selector for this table) | |
| Microsoft Azure | Standard_ND96isr_H100_v5[2] | 8 | $98.32 | $12.29derived: $98.32 ÷ 8 | East US | |
| H100 PCIe | RunPod | Community Cloud pod[1] | 1 | $1.99 | $1.99 | Not stated (31 regions) |
| RunPod | Secure Cloud pod[1] | 1 | $2.89 | $2.89 | Not stated (31 regions) | |
| H100 NVL | RunPod | Community Cloud pod[1] | 1 | $2.59 | $2.59 | Not stated (31 regions) |
| RunPod | Secure Cloud pod[1] | 1 | $3.19 | $3.19 | Not stated (31 regions) | |
| Microsoft Azure | Standard_NC40ads_H100_v5[2] | 1 | $6.98 | $6.98 | East US | |
| Microsoft Azure | Standard_NC80adis_H100_v5[2] | 2 | $13.96 | $6.98derived: $13.96 ÷ 2 | East US | |
| A100 80GB | Modal | Serverless GPU function[4] | 1 | $2.50 | $2.50derived: $0.000694/s × 3,600 | Base price; pinning a region costs 1.15–1.75× |
| CoreWeave | NVIDIA A100 instance (80 GB)[9] | 8 | $21.60 | $2.70derived: $21.6 ÷ 8 | North America (Europe lists the same on-demand price) | |
| RunPod | Serverless flex worker (A100 80 GB tier)[1] | 1 | $2.99 | $2.99 | Not stated (31 regions) | |
| AWS | p4de.24xlarge[10] | 8 | $27.45 | $3.43derived: $27.44705 ÷ 8 | US East (N. Virginia) | |
| Microsoft Azure | Standard_NC24ads_A100_v4 (PCIe)[2] | 1 | $3.67 | $3.67 | East US | |
| Microsoft Azure | Standard_ND96amsr_A100_v4[2] | 8 | $32.77 | $4.10derived: $32.77 ÷ 8 | East US | |
| Google Cloud | a2-ultragpu-1g (whole machine)[8] | 1 | $5.07 | $5.07 | Not stated on page (no region selector for this table) | |
| Google Cloud | a2-ultragpu-8g (whole machine)[8] | 8 | $40.55 | $5.07derived: $40.550383123 ÷ 8 | Not stated on page (no region selector for this table) | |
| A100 40GB | Modal | Serverless GPU function[4] | 1 | $2.10 | $2.10derived: $0.000583/s × 3,600 | Base price; pinning a region costs 1.15–1.75× |
| AWS | p4d.24xlarge[10] | 8 | $21.96 | $2.74derived: $21.957642 ÷ 8 | US East (N. Virginia) | |
| Microsoft Azure | Standard_ND96asr_v4[2] | 8 | $27.20 | $3.40derived: $27.197 ÷ 8 | East US | |
| Google Cloud | a2-highgpu-1g (whole machine)[8] | 1 | $3.67 | $3.67 | Not stated on page (no region selector for this table) | |
| Google Cloud | a2-highgpu-8g (whole machine)[8] | 8 | $29.39 | $3.67derived: $29.38708 ÷ 8 | Not stated on page (no region selector for this table) | |
| A100 SXM / PCIe (memory not stated) | RunPod | Community Cloud pod, A100 PCIe[1] | 1 | $1.19 | $1.19 | Not stated (31 regions) |
| RunPod | Community Cloud pod, A100 SXM[1] | 1 | $1.39 | $1.39 | Not stated (31 regions) | |
| RunPod | Secure Cloud pod, A100 PCIe[1] | 1 | $2.09 | $2.09 | Not stated (31 regions) | |
| RunPod | Instant Cluster, A100 SXM (no commitment)[1] | 1 | $2.15 | $2.15 | Not stated (31 regions) | |
| RunPod | Secure Cloud pod, A100 SXM[1] | 1 | $2.19 | $2.19 | Not stated (31 regions) | |
| L40S | RunPod | Community Cloud pod[1] | 1 | $0.79 | $0.79 | Not stated (31 regions) |
| RunPod | Secure Cloud pod[1] | 1 | $1.09 | $1.09 | Not stated (31 regions) | |
| DigitalOcean | GPU Droplet, per GPU[7] | 1 | $1.57 | $1.57 | Not stated on pricing page | |
| AWS | g6e.xlarge[10] | 1 | $1.86 | $1.86 | US East (N. Virginia) | |
| Modal | Serverless GPU function[4] | 1 | $1.95 | $1.95derived: $0.000542/s × 3,600 | Base price; pinning a region costs 1.15–1.75× | |
| CoreWeave | NVIDIA L40S instance[9] | 8 | $18.00 | $2.25derived: $18 ÷ 8 | North America (Europe lists the same on-demand price) | |
| L4 | RunPod | Community Cloud pod[1] | 1 | $0.44 | $0.44 | Not stated (31 regions) |
| RunPod | Secure Cloud pod[1] | 1 | $0.69 | $0.69 | Not stated (31 regions) | |
| Google Cloud | g2-standard-4 (whole machine)[8] | 1 | $0.7068 | $0.7068 | Iowa (us-central1), page default | |
| Modal | Serverless GPU function[4] | 1 | $0.7992 | $0.7992derived: $0.000222/s × 3,600 | Base price; pinning a region costs 1.15–1.75× | |
| AWS | g6.xlarge[10] | 1 | $0.8048 | $0.8048 | US East (N. Virginia) | |
| A10G / A10 | AWS | g5.xlarge (A10G)[10] | 1 | $1.01 | $1.01 | US East (N. Virginia) |
| Modal | Serverless GPU function (A10)[4] | 1 | $1.10 | $1.10derived: $0.000306/s × 3,600 | Base price; pinning a region costs 1.15–1.75× | |
| Microsoft Azure | Standard_NV36ads_A10_v5 (A10)[2] | 1 | $3.20 | $3.20 | East US | |
| T4 | Google Cloud | T4 attached to an N1 VM (GPU only; VM billed separately)[8] | 1 | $0.35 | $0.35 | Iowa (us-central1), page default |
| AWS | g4dn.xlarge[10] | 1 | $0.526 | $0.526 | US East (N. Virginia) | |
| Microsoft Azure | Standard_NC4as_T4_v3[2] | 1 | $0.526 | $0.526 | East US | |
| Modal | Serverless GPU function[4] | 1 | $0.5904 | $0.5904derived: $0.000164/s × 3,600 | Base price; pinning a region costs 1.15–1.75× | |
| RTX 6000 Ada | RunPod | Community Cloud pod[1] | 1 | $0.74 | $0.74 | Not stated (31 regions) |
| RunPod | Secure Cloud pod[1] | 1 | $1.09 | $1.09 | Not stated (31 regions) | |
| DigitalOcean | GPU Droplet, per GPU[7] | 1 | $1.57 | $1.57 | Not stated on pricing page | |
| RTX 4000 Ada | Linode (Akamai Cloud) | g2-gpu-rtx4000a1-s[11] | 1 | $0.52 | $0.52 | Base price (some regions differ) |
| Linode (Akamai Cloud) | g2-gpu-rtx4000a2-s[11] | 2 | $1.05 | $0.525derived: $1.05 ÷ 2 | Base price (some regions differ) | |
| Linode (Akamai Cloud) | g2-gpu-rtx4000a4-s[11] | 4 | $2.96 | $0.74derived: $2.96 ÷ 4 | Base price (some regions differ) | |
| DigitalOcean | GPU Droplet, per GPU[7] | 1 | $0.76 | $0.76 | Not stated on pricing page |
Billing granularity and minimums
| Provider | How GPU time is billed |
|---|---|
| AWS | Per second with a 60-second minimum (Linux On-Demand)[5] |
| Google Cloud | 1-minute minimum, then per second (applies to vCPUs, GPUs and memory)[6] |
| Microsoft Azure | — |
| CoreWeave | Priced per instance-hour; invoices count hours, usage attributed every 30 seconds[12] |
| RunPod | Per second, no commitment (Pods)[1] |
| Modal | Per second; containers stay up 60 s after the last input by default, which is billed[4] |
| DigitalOcean | Per second with a 5-minute minimum; powered-off GPU Droplets are still billed[7] |
| Linode (Akamai Cloud) | Hourly, rounded up; GPU plans have no monthly cap since July 1, 2026[13] |
| Hetzner | Dedicated GPU servers billed monthly or hourly plus a one-time setup fee (EUR, excl. VAT)[14] |
| Fly.io | No GPU prices on the current pricing page; GPU Machines continue but no new version is planned[15] |
| Railway | No GPU instances offered[16] |
How to read the price ranges
A100 80GB: $2.50 (Modal, Serverless GPU function)[4] to $5.07 (Google Cloud, a2-ultragpu-8g (whole machine))[8]. L40S: $0.79 (RunPod, Community Cloud pod)[1] to $2.25 (CoreWeave, NVIDIA L40S instance)[9]. L4: $0.44 (RunPod, Community Cloud pod)[1] to $0.80 (AWS, g6.xlarge)[10]. H200: $3.59 (RunPod, Community Cloud pod)[1] to $10.60 (Google Cloud, a3-ultragpu-8g (whole machine))[8].
Three things explain most of the spread. First, what's bundled: an AWS p5.48xlarge or a Google a3-highgpu-8g is priced as a whole machine, with its vCPUs, memory and local SSD included; Modal's GPU price, by contrast, excludes CPU and memory, which are billed separately per second[4]. Second, interruption and placement: Modal's listed GPU prices are preemptible base prices, and pinning a region multiplies them by 1.15× to 1.75×[4]. Third, where the hardware sits: RunPod's Secure Cloud operates in T3/T4 data centers, while Community Cloud connects individual compute providers to users through a vetted peer-to-peer system[3].
Some hardware isn't sold on demand everywhere. Google Cloud's B200 machine (a4-highgpu-8g) shows no on-demand price, only Flex-start, Calendar and Spot[8]. Hetzner's GPU line is dedicated servers with RTX PRO Blackwell cards billed monthly or hourly in euros plus a setup fee, so it isn't in the per-GPU-hour table[17]. Railway says it does not offer GPU instances[16].
Choosing a procurement model before a provider
Serverless GPU functions (Modal, RunPod Serverless) fit bursty inference: you pay per second while a container is up and nothing when it has scaled to zero, but you pay for cold starts and the idle window before scale-down. On-demand pods and instances (RunPod Pods, DigitalOcean GPU Droplets, single-GPU cloud VMs) fit steady single-GPU work. 8-GPU instances (CoreWeave HGX, AWS P5, Azure ND, Google A3) fit multi-GPU training and large-model serving that needs fast GPU-to-GPU links.
The GPU capacity guide on our blog covers capacity blocks, quotas and cold starts in more depth. The RunPod vs Modal and CoreWeave vs RunPod pages compare those providers on everything beyond price.
Run the same model on more than one GPU cloud
AW3 Host deploys to CoreWeave, RunPod and Modal alongside AWS, Google Cloud, Azure and Kubernetes from one dashboard and API. With the same inference service on two GPU providers, weighted routing shifts traffic between them, failover with 60-second health checks moves traffic when one runs out of capacity or fails, and canary rollouts let you move a model to new hardware a step at a time.
Daily cost sync and budget policies put GPU spend from every connected provider in one view with alerts. AW3 Host doesn't get you quota or capacity on any provider, and it doesn't change their prices.
Frequently asked questions
- What is the cheapest on-demand H100?
- On list prices checked 16 November 2026, the lowest H100 SXM figure in this table is Community Cloud pod on RunPod at $2.69 per GPU-hour. RunPod lists H100 PCIe pods from $1.99/hr on Community Cloud.
- How much does an H100 cost per hour on AWS?
- In US East (N. Virginia), a p5.48xlarge with 8 H100 GPUs is $55.04 per instance-hour on demand for Linux, which is $6.88 per GPU-hour. A single-GPU p5.4xlarge is also listed at $6.88 per hour.
- How does Modal bill for GPUs?
- Per second while a container runs, including model load time. Containers stay up 60 seconds after the last input by default, and that time is billed. An H100 is listed at $0.001097 per second, about $3.95 per hour, as a preemptible base price; pinning a region costs 1.15× to 1.75×.
- Why is the per-GPU price for a hyperscaler 8-GPU machine so much higher?
- Because it isn't a GPU-only price. Instances like AWS p5.48xlarge or Google a3-highgpu-8g bundle vCPUs, memory and local SSD with the GPUs, and the per-GPU figure here is simply the machine price divided by eight.
How the numbers were gathered
All prices are on-demand list prices in USD for Linux, read on 16 November 2026 from: CoreWeave's pricing page; RunPod's pricing page (dated November 12, 2026); Modal's pricing page; the EC2 On-Demand price data that aws.amazon.com/ec2/pricing/on-demand/ loads for US East (N. Virginia); Google Cloud's accelerator-optimized pricing page; the Azure Retail Prices API (Consumption, Linux); DigitalOcean's GPU Droplet pricing page; and the Linode API's plan types.
Per-GPU prices for multi-GPU units are the unit price divided by its GPU count; Modal's per-second prices are multiplied by 3,600. Both are marked "derived" in the table. Where a provider doesn't state the region a price applies to, the table says so. RunPod's A100 pods are grouped separately because the pricing page names them A100 SXM and A100 PCIe without stating memory.
Not included: spot, preemptible-discount and committed-use prices (except Modal, whose list prices are preemptible), storage, egress, and taxes. Fly.io's GPU documentation now redirects to a February 2025 post saying GPU Machines aren't being removed but won't get a new version, and its pricing page lists no GPU prices, so Fly has no rows.
Sources
- RunPod: Pricing (page dated November 12, 2026). https://www.runpod.io/pricing (checked 2026-11-16)
- Microsoft: Azure Retail Prices API (Virtual Machines, Consumption, Linux). https://prices.azure.com/api/retail/prices (checked 2026-11-16)
- RunPod Docs: Pods overview. https://docs.runpod.io/pods/overview (checked 2026-11-16)
- Modal: Pricing. https://modal.com/pricing (checked 2026-11-16)
- AWS: EC2 On-Demand pricing. https://aws.amazon.com/ec2/pricing/on-demand/ (checked 2026-11-16)
- Google Cloud: Compute Engine pricing. https://cloud.google.com/products/compute/pricing (checked 2026-11-16)
- DigitalOcean: GPU Droplet pricing. https://www.digitalocean.com/pricing/gpu-droplets (checked 2026-11-16)
- Google Cloud: Accelerator-optimized machine pricing. https://cloud.google.com/products/compute/pricing/accelerator-optimized (checked 2026-11-16)
- CoreWeave: Pricing. https://www.coreweave.com/pricing (checked 2026-11-16)
- AWS: EC2 On-Demand price data for US East (N. Virginia), Linux (file loaded by the EC2 pricing page). https://b0.p.awsstatic.com/pricing/2.0/meteredUnitMaps/ec2/USD/current/ec2-ondemand-without-sec-sel/US%20East%20(N.%20Virginia)/Linux/index.json (checked 2026-11-16)
- Linode API: plan types. https://api.linode.com/v4/linode/types (checked 2026-11-16)
- CoreWeave Docs: Capacity plans. https://docs.coreweave.com/platform/capacity-plans (checked 2026-11-16)
- Akamai TechDocs: How billing works. https://techdocs.akamai.com/cloud-computing/docs/understanding-how-billing-works (checked 2026-11-16)
- Hetzner: GEX131. https://www.hetzner.com/dedicated-rootserver/gex131/ (checked 2026-11-16)
- Fly.io Blog: We Were Wrong About GPUs (2025-02-14). https://fly.io/blog/wrong-about-gpu/ (checked 2026-11-16)
- Railway: AI agent workers guide (GPU availability). https://railway.com/guides/ai-agent-workers (checked 2026-11-16)
- Hetzner: GPU dedicated servers. https://www.hetzner.com/dedicated-rootserver/matrix-gpu/ (checked 2026-11-16)
Your next deploy
goes everywhere.
Connect a provider, point a service at it, and ship. Add the second provider when you're ready.
No card needed · Free tier · Pay as you go after that