L4 cloud pricing: from $0.43/hr across 12 providers (Oct 2026)

Prices are as printed on each provider's own pricing page or public pricing API, read between October 9, 2026 and October 17, 2026: a survey of the 336 GPU-selling operators in our Neocloud Directory, plus the hyperscaler quotes on our GPU price pages (3 hyperscalers: AWS Price List API, Google Cloud's pricing page and the Azure Retail Prices API). Prices are read from the provider itself, never from a third-party price-comparison site. 117 providers publish at least one price; 89 sell GPUs but quote only on request; 6 publish prices only for GPUs this comparison does not cover; the other 127 publish no GPU pricing.

Every price is per GPU-hour in US dollars. Instance prices are divided by the GPUs in the instance (so they include its CPUs, memory and storage); monthly prices by 730 hours, and 30-day prices by 720. Where a provider prints its own hourly equivalent, that figure is used. Prices tied to a region are shown with it; otherwise the region column shows the provider’s headquarters region, marked (HQ).

Prices listed in other currencies are converted at the European Central Bank's euro reference rates of October 16, 2026 (1 EUR = $1.1253, 1 INR = $0.0105, 1 CHF = $1.2137, 1 IDR = $0.000055, 1 JPY = $0.006328); the table shows the listed price below the dollar figure. Prices exclude VAT and sales tax where the provider says so.

Each provider counts once per GPU and commitment, at its lowest listed price. Lowest, median and highest are across those provider prices. "From" marks a provider’s cheapest configuration. Lowest reserved is the lowest price on a stated term (1 month to 5 years); reservations with no stated term are in the table only. A range whose ends the page does not tie to a term is shown as indicative.

Marketplaces and aggregators (the Akash Network, Vast.ai, Oneinfer, Prime Intellect and others), whose prices are set by hosts on the platform or resold from other clouds, are shown in the table as Marketplace and left out of the figures. So are H100 rows that do not say SXM or PCIe: they are listed on the H100 page, marked, but not counted as SXM. A100 rows with no memory size are left out.

Rows marked "check" are kept in the table but left out of the figures: prices below 45% of the median provider price for the same GPU and commitment class, pages that look out of date, prices shown in calculators rather than price lists, and per-GPU prices where the page does not say how many GPUs the instance has.

These are list prices. Large reservations are usually negotiated below them, and spot prices move by the hour. Confirm on the provider’s page before you buy.