Cloud GPU pricing for LLM inference

Providers and sources

• Lambda (GPU cloud) · Price per GPU-hour falls slightly with larger instances; the range runs from the largest instance to a single GPU.
• CoreWeave (GPU cloud) · Published per instance (North America, on-demand); per-GPU prices here are the instance price divided by its GPUs.
• RunPod (GPU cloud) · Pods. Secure Cloud runs in RunPod's partner data centers; Community Cloud is cheaper capacity from vetted hosts.
• Nebius (GPU cloud) · Prices exclude taxes, including VAT.
• Together AI (GPU cloud) · GPU clusters. Reserved prices shown are for 91–180 day reservations.
• DigitalOcean (GPU cloud) · GPU Droplets, billed per second with a 5-minute minimum. Prices effective 1 August 2026.
• Hyperstack (GPU cloud)
• AWS (Hyperscaler) · EC2 Capacity Blocks for ML: GPUs reserved for a fixed window, priced per accelerator-hour, US East (N. Virginia) unless stated. AWS's on-demand GPU prices are not on a page we can read, so they are not listed.

Not listed yet

• NVIDIA: NVIDIA does not publish per-hour prices; DGX Cloud Lepton routes to partner clouds.
• Google Cloud: its GPU price tables did not load from the official pricing page.
• Microsoft Azure: its pricing page shows no prices without its interactive calculator.
• Oracle Cloud: its GPU pricing page could not be read.
• Vultr, Scaleway, OVHcloud: their pricing pages could not be read or did not list GPU prices.
• Crusoe: its pricing table did not tie each price to a GPU clearly enough to record.

Spotted a price change or a missing provider? Tell us.

Quick answers

How much does an H100 cost per hour?

H100: from $2.50 per GPU-hour (Hyperstack) on-demand, among the providers listed here, up to $6.16. Marketplace capacity is cheaper: RunPod Community Cloud lists an H100 at $1.99. Prices differ with the form factor (SXM, PCIe, NVL) and whether the GPU comes in an 8-GPU node.

What is the cheapest GPU for serving an LLM?

It depends on the model's size. A model has to fit in GPU memory with room for its KV cache: a 70B model at 8-bit needs about 80 GB, so one H100 or A100 80GB; larger models need H200 (141 GB), B200 (180 GB) or several GPUs. Current lowest on-demand prices: H200: from $3.99 per GPU-hour (Hyperstack); B200: from $6.00 per GPU-hour (Hyperstack); A100: from $1.35 per GPU-hour (Hyperstack); L40S: from $1.09 per GPU-hour (RunPod).

Why are the hyperscalers' prices higher?

The AWS prices listed are Capacity Blocks, GPUs reserved for a fixed window, and they include the full 8-GPU instance with its CPUs, memory and networking. Specialised GPU clouds usually list lower per-GPU prices. Google Cloud, Azure and Oracle are not listed because their GPU prices could not be read from their pricing pages.

How were these prices collected?

By hand, from each provider's own public pricing page, linked in every row. A price is recorded only as the provider states it; where a provider publishes a price per instance, the per-GPU price is the instance price divided by its GPUs and is marked as derived. Prices change often: check the provider before you buy.