Read every rate in context

What is included

This pass contains 351 configurations across 22 providers. Rows cover exact public-cloud SKUs, named GPU presets, complete GPU nodes and bare-metal servers. Published CPU/RAM variants are distinct configurations. Every row has at least three sourced hardware fields beyond name, URL and summary.

Provider coverage and gaps lists partial catalogues and excluded offers. Fractional GPUs, inference API endpoints, regional duplicates, custom attachments and non-GPU accelerators are outside scope. A specification establishes hardware; it does not establish live inventory, quota or a purchasing offer.

How current hourly prices are compared

The main price is USD per complete instance-hour. A provider’s GPU-hour rate is multiplied by GPU count. The table, exports and detail pages also calculate the equivalent GPU-hour from the instance rate. Both input sources remain attached to the calculation. Rounding affects display only; JSON keeps the captured value.

AWS prices use its public Linux on-demand feed for N. Virginia, with no preinstalled software. Azure uses its Retail Prices API for East US Consumption base/Linux VM meters, excluding Windows, Spot and Low Priority. Google uses the selected Iowa region and the standard hourly consumption column. CoreWeave uses the North America full-instance rate. Crusoe uses its current public GPU-hour list, with the December price change kept separate. Nebius uses its on-demand GPU-hour list and the documented preset region. Oracle uses the public USD PAY_AS_YOU_GO GPU-hour catalogue entry for each GPU model, multiplied by the shape’s GPU count. IBM uses its Global Catalog USD instance-hour price for us-east.

Disk and network charges, software, taxes and minimum billing periods affect the final bill. Google’s rate includes bundled Local SSD where the configuration includes it. Each row’s Price terms records the relevant scope. Spot, Flex-start, Calendar, inference-only and committed rates are separate purchase models and do not fill an empty current on-demand cell.

RunPod uses the Secure Cloud one-GPU Pod cards, checked against the page’s structured Offers. Together uses GPU Clusters tariffs for whole nodes, rather than Dedicated Inference prices. DigitalOcean and Paperspace are separate catalogues. Hyperstack quotes per-GPU on-demand tariffs; its allocated CPU cores are kept separate from vCPUs. OVHcloud uses the public USD hourly list, without inferring a deployment region. Vultr’s captured catalogue plan has no listed deployable location.

Hetzner’s EUR hourly rates have setup fees and monthly caps. Scaleway’s captured L4 prices are in EUR before tax. Native-currency fields retain these offers. No exchange rate is invented to fill the USD comparison.

Why a price stays empty

Google lists no standard on-demand price for A4; it sells B200 capacity through Spot, Flex-start and reservations. Other empty cells mark shapes whose provider publishes no public USD rate for that exact configuration. An empty cell is a missing public tariff, never a zero price.

Lambda’s public list takes effect on 16 November 2026 at 00:00 UTC and replaces login-only pricing. Lambda rows use that list and say so in their rate scope.

GPU memory and interconnect

Memory per GPU preserves the provider’s nominal device capacity and its GB or GiB label. The numeric VRAM field carries the original unit in a separate field; use the combined memory display when comparing providers. When a source gives only aggregate GPU memory, divide it by its GPU count. Total nominal memory is a calculation, not a promise that a model fits on one device. AWS nominal marketing GB and its manual’s visible GiB can differ; the family product page supplies nominal capacity. Host memory preserves the source’s units. Google’s host JEDEC GB are GiB, as defined on its pricing page.

NVLink and NVSwitch describe communication between GPUs; EFA, InfiniBand and RDMA describe node networking. Network limits retain baseline, burst, per-port and aggregate qualifiers. Oracle’s OCPUs, Hyperstack’s CPU cores and Hetzner’s physical CPU specifications are not silently converted to vCPUs. Their vCPU cells stay empty.

Region availability

Published regions and zones are captured from provider deployment tables where verified. Some tables apply at the family level, rather than guaranteeing every size in every zone. The Rate scope field describes a tariff’s geography and is kept separate from deployment regions. Neither a catalogue entry nor a region listing guarantees stock or account eligibility.

Evidence and updates

Every filled field maps to a dated first-party source in the JSON. Provider product pages establish hardware, and pricing pages or public catalogue APIs establish tariffs. An accessed date records when the source was read; it is not a hardware launch date or price effective date. A record’s verified date changes when its evidence is checked, rather than on every build.

Send corrections with a public source to hello@cloudgpulist.com. Paid featured placement is labelled and has no effect on data or table sorting.

Reuse the comparison

The dataset is available under CC BY 4.0. Download CSV or JSON with source mappings and access dates, and credit Cloud GPU List with a link to the canonical page.