Cloud GPU Pricing

As of October 2026, renting an NVIDIA H100 (80GB) GPU costs a median of $5.00/hour on-demand. Rates vary significantly by provider tier: specialist clouds (RunPod, Crusoe, Lambda, Nebius, Together and CoreWeave) charge $3.49 to $6.16/GPU-hour, hyperscalers (AWS, Oracle, Google Cloud and Azure) charge $6.88 to $12.29/GPU-hour, and unreserved marketplaces drop as low as $2.67/GPU-hour.

Price variance is driven by bundled platform services, network topology, and compliance guarantees. Hyperscalers charge a premium because GPU nodes include high-availability SLAs, global compliance certifications, and deep integration with enterprise IAM and storage ecosystems. Specialist GPU clouds minimize non-GPU software overhead to offer lower unit rates, while marketplaces trade availability guarantees for minimal pricing.

The cheapest on-demand rates are found on GPU marketplace platforms (Vast.ai and RunPod Community), with H100s from ~$2.67/hr, though marketplace capacity comes without availability guarantees. For production-grade workloads requiring high bandwidth and interconnect guarantees, the lowest reliable rates come from specialist clouds: RunPod, Crusoe and Lambda at $3.49 to $3.99/hr for an H100, and Crusoe, RunPod and Nebius at $4.29 to $5.40/hr for an H200.

On-demand rates provide pay-as-you-go hourly compute with guaranteed availability and no minimum commitment. Spot (or preemptible) rates offer 40–60% discounts but can be interrupted with short notice when host capacity is reclaimed. Reserved rates offer 15–35% discounts in exchange for 1-year to 3-year term commitments.

All figures on this comparison page are strictly normalised to USD per individual GPU per hour. If a provider quotes an 8×H100 server node at $24.00/hour, our pipeline normalises this to $3.00/GPU-hour for accurate apples-to-apples evaluation.

Renting by the hour is the right choice for intermittent jobs, variable training runs and early R&D, where paying only for the hours you use beats holding capacity that sits idle. The balance shifts towards committed or fixed-cost capacity as sustained utilisation rises, because a metered hour costs the same whether the GPU is saturated or idle. There is no single crossover percentage: it depends on the GPU, the provider tier you would otherwise rent from, and how much data you move. Our GPU pricing calculator works the threshold out for your own usage from the same rates tracked on this page.