GPU cloud pricing FAQ

The CoreWeave H100 SXM published list rate in June 2026 is $6.155 per GPU-hour, billed per-GPU on the HGX H100 8x node configuration, sourced from the CoreWeave pricing page. Reserved-capacity contracts (1-3 year) price below the list rate but require direct negotiation. This is the canonical CoreWeave H100 SXM rate referenced across this site; if any page quotes a different number, the vendors.ts data file (the single source of truth) wins.

Top-traffic vendors (CoreWeave, Lambda, RunPod, AWS, Azure, GCP) are re-verified monthly. The long tail (Crusoe, Hyperstack, Together AI, DigitalOcean, Modal, Replicate, etc.) is re-verified quarterly. Every page footer carries the most recent verified date.

Two reasons. First, the headline list rate at AWS, Azure, and GCP is positioned for spot-buyers and is the highest tier the vendor offers; actual prices paid by enterprise customers under a Savings Plan, Reserved Instance, or Capacity Blocks for ML commitment are materially lower. Second, hyperscaler rates bundle integration depth (IAM, S3, EFA, SageMaker, Foundry) that the specialist clouds do not match. The integration premium is real but it is not always justified for greenfield AI workloads.

Only when your workload shape benefits from it. For continuous training, per-second billing typically prices at a premium versus the same GPU rented per-hour. For inference that processes a few requests per minute, per-second billing means you pay only for the seconds the GPU is doing useful work, which can be much cheaper than a per-hour instance that sits idle most of the time. Cold-start billing on serverless platforms is the gotcha.

Specialist clouds (CoreWeave, Lambda, Together, Crusoe, DigitalOcean) typically publish 30 to 60 percent off on-demand for 1 to 3 year reservations. Hyperscalers offer 30 to 50 percent off via Reserved Instances or Savings Plans. Capacity Blocks for ML on AWS reserves dated H100 / A100 windows in advance at a premium relative to on-demand but with a capacity guarantee. Enterprise EA contracts at all three hyperscalers add a further committed-spend discount.

The current scope is Nvidia GPUs (H100, H200, A100, L40S, L40, A40, A10, A10G, T4, RTX). Google TPUs are mentioned on the Google Cloud page because they share a control plane. AWS Trainium and Inferentia, Intel Gaudi, and AMD MI300X are out of scope for the current release; we plan to add MI300X coverage when more vendors publish public per-hour rates.

DigitalOcean now offers MI300X GPU Droplets and TensorWave has built a business on Instinct capacity. Public per-hour rates are still less consistently surfaced than Nvidia. We plan to add a dedicated MI300X page when the rate-card coverage justifies it.