GPU suppliers

Only major providers enter the comparison. A provider is major only when it is publicly traded and meets both criteria below, each shown with the source that proves it. The criteria, the market-cap floor and the list are kept in the registry and change only after review.

• Publicly traded, and reports a public-cloud or GPU-cloud business in its latest SEC annual report (10-K or 20-F)
• Market cap of at least $10B, from an exchange quote page with its URL, date and value

Marketplaces, resellers and private companies are never major, and fleet size or data-center power is not a path to major. The market-cap floor is $10B; it can be changed in the registry. There is no include override, so no provider can be added without a sourced market cap.

Major providers only. On-demand, USD per GPU-hour, one full-node configuration, read from each provider's own page or API within 72 hours, at one compute precision (FP16/BF16 Tensor Core, dense). A model is compared only when at least 3 major providers have a price for it. The headline takes one GPU of each model at each provider.

By GPU model, $ per PFLOP-hour

$ per GPU-hour by provider and model

Not compared, fewer than 3 major providers with a comparable price: GB300, L40S, MI300X

Their prices come from listings, aggregators or marketplace asks and can mix rental terms, so they are shown on their own and never averaged with the major providers.

Version 2026-10-28.2 · FP16/BF16 Tensor Core, dense

No figure, left out: H100 PCIe, GH200

• Compute is counted at one precision for every GPU: FP16/BF16 Tensor Core, dense, in TFLOPS per GPU, from the vendor's own specification page (linked in the table, with the retrieval date).
• Effective compute of a provider is the sum over its GPU models of the number of GPUs times the TFLOPS of that model. The comparison price is a price per PFLOP-hour: the price of one GPU-hour divided by the TFLOPS of that GPU, summed over the same models and providers.
• A provider or model without a comparable price is left out and flagged with the specific reason; nothing is estimated and no price is carried over from an old reading.
• Fleet counts are either a published fleet or, for a marketplace, the GPUs available to rent now (labelled on each page). Prices in the comparison come from the major provider's own page or API, never from an aggregator.
• Every datum on a profile has a source link and a date. Gaps stay empty. Sources checked without a result are listed on each profile.

Data integrity

• Same inputs for everyone: the same GPU model, the same rental type (on-demand, never spot or reserved), the same unit and currency, and the same compute precision (FP16/BF16 Tensor Core, dense).
• Unit: USD per GPU-hour, taken from a full-node price divided by its GPU count; single-GPU slices, small servers and other sizes are not mixed in.
• Configuration: 8-GPU nodes (4 GPUs for the GB200 and GB300 rack systems). Where several instances qualify, the lowest list price is used and the instance is shown.
• Region: regional price lists are read in one reference US region; providers with a single global price list are labelled as such.
• Recency: a price counts only if it was read from the provider's own page or API within 72 hours; older readings are left out.
• Coverage: a model is compared only when at least 3 major providers have a price for it.
• Tier: major status needs a publicly traded owner with a sourced market cap of at least $10B (quote page, date and value kept in the registry); marketplaces, resellers and private companies can never enter the comparison.
• Gaps: when a major provider does not publish a comparable price (contact sales, unit not stated, two prices for one model) it is left out for that model and the reason is shown.
• Every price is checked by code at build time against these rules, and automated tests cover mixed units, mixed precision, stale data and a non-major provider trying to enter.