Methodology
Primary sources, one link per number, a date on every row.
Where the data comes from
We read each provider’s own pricing pages, documentation, status pages, filings and press releases, plus named industry reports. Every record lists its sources with the date we accessed them, and every number in the table links to the source it came from.
How we verify
Each row shows the date we last checked it. A field is filled only when a source states it; when a provider does not publish a figure, the cell stays empty. List prices are the public, on-demand prices; negotiated and committed-use discounts are not shown.
What counts as a price
The H100 column is the published per-GPU-hour price of one H100. The basis sits beside it on every provider page: on-demand, a “starting at” price, or an 8-GPU instance price divided by 8 (CoreWeave, Vultr, Azure, IBM Cloud). Hyperscaler prices are read from the provider’s price API or pricing page for one named region. Providers that sell by quote, such as Rescale and Cudo Compute, show no price.
What counts as a fabric
The fabric column holds the node-to-node network the provider names for its cluster products: InfiniBand, AWS EFA, RoCE, Slingshot, or RDMA where a provider names cluster networking without the transport (Oracle, Alibaba). An empty cell means the provider’s pages do not name one.
Corrections
Vendors and readers can send corrections to hello@hpcproviders.com with a public source. We update the row and its verified date. Featured placement is paid and labelled; it never changes a row’s data or its position in the sorted table.
Reuse
The dataset is free to reuse under CC BY 4.0 with a link to hpcproviders.com. Download it as CSV or JSON.
Field definitions
21 fields- Type
- What kind of provider it is: hyperscaler, managed HPC service, GPU cloud, sovereign or EU-hosted cloud, publicly funded supercomputer, or simulation-software cloud.
- Fabric
- Primary node-to-node network the provider states for its HPC or GPU clusters. RDMA means the provider names RDMA cluster networking without naming the transport.
- Interconnect detail
- Speeds and instance families as the provider states them.
- Schedulers
- Job schedulers and orchestrators the provider offers or documents.
- Slurm
- Offers or documents Slurm (managed, or as a supported deployment).
- Kubernetes
- Offers managed Kubernetes for its clusters.
- CPUs
- CPU models the provider names for its HPC or GPU hosts.
- GPUs / accelerators
- Accelerators the provider lists, as listed on its own pages.
- H100
- Lists NVIDIA H100.
- Blackwell
- Lists at least one NVIDIA Blackwell data-centre GPU (B200, B300, GB200, GB300).
- AMD Instinct
- Lists at least one AMD Instinct GPU.
- Bare metal
- Sells non-virtualised bare-metal nodes.
- H100 $/GPU-hr $/GPU-hr
- Published per-GPU-hour price of an H100. Basis (on-demand, capacity block, derived from an 8-GPU price) is in the next column.
- H100 price basis
- What the H100 price is: purchase mode, instance size and any derivation.
- Price checked
- Date the price was read from the provider page.
- Pricing model
- How the provider charges, as it states it.
- Regions
- Locations the provider names on its own pages.
- EU-hosted
- Names data-centre capacity in an EU member state.
- Sovereignty claim
- The provider's own sovereignty or compliance statement, quoted or closely paraphrased.
- Software catalogue
- Pre-configured application catalogue a platform provider advertises.
- Runs on
- Clouds and hardware partners a platform provider lists.