Methodology & data sources

Transparency is the core of our E-E-A-T. This page documents exactly where each GPU rental price comes from, when it was captured, and the formulas behind every derived figure. We track 13 GPU models across 14 providers.

Snapshot date & volatility

All prices on this site are a dated snapshot (June 2026). Cloud GPU pricing is exceptionally volatile — list prices fell roughly 60-75% year over year through 2025-2026 as supply expanded — and varies by region, instance type, currency and discounts. Prices change weekly; verify on the provider's pricing page before renting. We do not fabricate figures: where a value could not be verified on the vendor's page, we render "—" with a reason (for example "not offered" or "contact sales").

Per-provider sources

Each provider's figures are taken from its own published pricing page:

How we derive figures

• On-demand & spot prices are read directly off each provider's pricing page (per GPU per hour). Where a provider quotes per-instance, we note the source.
• Hyperscaler per-GPU price (AWS, GCP, Azure) = published 8-GPU instance price / 8. These figures include the bundled vCPU, RAM and support that come with the instance, which is part of why hyperscalers run 2-4x the specialist neoclouds. Single-GPU instances (L4, L40S, A10, H100 NVL, V100, A100 PCIe) are taken as-is.
• Marketplace (Vast.ai) figures are representative medians of host-set prices, not fixed list prices; individual hosts can be cheaper or pricier, and the interruptible tier is ~50% lower.
• Cheapest-by-model & rankings sort providers by their published on-demand (or spot) rate for that GPU; ties break alphabetically.
• 8xH100 node = published H100 SXM per-GPU rate x 8; the monthly figure x 730 hours (continuous use).
• Calculator total = price per GPU per hour x number of GPUs x hours. It runs entirely in your browser; we don't store inputs. The optional training-hours helper sets hours = GPU-hours / GPUs.

Data sources (summary)

What's excluded

Our headline prices and the calculator focus on the per-GPU compute rate. They do not include storage (volumes), data egress, idle time, inter-node networking (InfiniBand), minimum billing increments, or committed-use/reserved discounts — all of which can materially change your real bill. For multi-node distributed training, fabric quality and node count matter as much as the headline per-GPU price.

Notes & caveats

• Verda is the former DataCrunch (rebranded in 2026; datacrunch.io redirects to verda.com).
• Hyperscaler flagship GPUs (H100/H200/B200/A100 SXM) sell only as whole 8-GPU instances.
• Some newest SKUs (B200/GB200/MI355X at several providers) are contact-sales only and shown as "—".

Limitations & accuracy

Figures are estimates for general information and may lag the underlying source or contain errors. This is an informational comparison, not a quote. Always verify against the provider's primary pricing page before relying on it. See our disclaimer.

Last updated: 2026-06-21