GPU cloud prices

Public on-demand and listed GPU rental prices from 25 cloud providers (H100, H200, B200, B300, GB200, GB300, A100, L40S, MI300X, MI355X), with a source URL and check date on every row. Normalized to USD per GPU-hour.

This repo is the open data behind the Hundredth Run compute desk. Live tables, provider pages and a GPU memory and cost calculator are on the site. A JSON API is at /api/v1/prices.json.

Also on Hugging Face and Kaggle, with a quick cheapest-price browser as a Hugging Face Space. To call the JSON API without writing code, fork the public Postman collection. To query it with SQL in the browser, open the versioned DoltHub database. To chart the live API in a notebook, fork the Observable notebook. To try the endpoints in a browser API console, use the free RapidAPI Hub listing.

Use it from an AI agent (MCP)

A free remote MCP server answers live price questions in Claude, Cursor or any MCP client. No key.

{"mcpServers": {"gpu-prices": {"type": "http", "url": "https://hundredthrun-compute-desk.pages.dev/mcp"}}}

Tools: list_gpus, cheapest_gpu (for example H100), and get_prices (filter by GPU, provider, on_demand, spot or reserved). Every answer carries the source URL and date checked. Docs: https://hundredthrun-compute-desk.pages.dev/open-data

Cline

Remote server, nothing to install or clone, no API key. Add this to cline_mcp_settings.json (Cline > MCP Servers > Configure):

{
  "mcpServers": {
    "gpu-cloud-prices": {
      "type": "streamableHttp",
      "url": "https://hundredthrun-compute-desk.pages.dev/mcp",
      "disabled": false,
      "autoApprove": []
    }
  }
}

Then ask something like "what is the cheapest H100 on demand right now?". All three tools are read-only.

Listed on the official MCP Registry as io.github.hundredthrun/gpu-cloud-prices: https://registry.modelcontextprotocol.io/v0/servers?search=io.github.hundredthrun/gpu-cloud-prices

Listed on Glama: https://glama.ai/mcp/connectors/dev.pages.hundredthrun-compute-desk/hundredth-run-gpu-cloud-prices

Listed on Cursor Directory (one-click install in Cursor): https://cursor.directory/plugins/gpu-cloud-prices-hundredth-run

Files

File	What it is
data/gpu_prices.csv	One row per provider, GPU and price type. Columns include price_usd_per_gpu_hr, price_type, term, region, price_as_listed, source_url, date_checked, fetch_method, confidence, notes
data/gpu_prices.json	The same rows plus meta (generated time, per-provider fetch status)
data/lead_times.csv	Availability and lead-time notes as stated by providers
data/models.csv	Open model parameter counts used by the calculator, with sources
data/history/<date>.json	Daily snapshots (since 2026-11-04) of median price per GPU, for tracking changes

Starter notebook

notebooks/gpu-cloud-prices-starter.ipynb answers five buyer questions with pandas: cheapest on-demand price per GPU, provider spread, spot vs on-demand gap, daily median history, and provider price changes. It reads the files from the Hugging Face mirror, so it runs anywhere. Open in Colab.

Sources and method

• Prices are read from each provider's public pricing page or official price API. Every row links to the page it came from (source_url) and records the date it was read (date_checked) and how (fetch_method: official API, static HTML, or read by hand).
• Multi-GPU node prices are divided by the GPU count to give a per-GPU-hour figure. The original wording is kept in price_as_listed.
• confidence is high for API or clearly itemized prices, medium for "starting at" style prices, low when the only price is in an FAQ or similar.
• If a value is not published (region, lead time, a GPU price behind "contact sales"), the field stays blank or unknown. Nothing is estimated or filled in.

How often it is updated

Daily. The refresh re-reads every source, saves a new dated history snapshot and pushes here. The Kaggle and Hugging Face mirrors are republished weekly. If a provider page cannot be read, its old rows are kept and marked stale in meta.provider_status.

Limitations

• These are public list prices, mostly on-demand. Negotiated, reserved, spot and committed-use prices are often much lower and are mostly not covered.
• Prices change often. A row can be a day or more old if a page failed to load. Check date_checked and the source before you buy.
• Availability is not guaranteed by a listed price.
• Unknown stays blank. Missing data does not mean free or unavailable.

Updating (maintainers)

python3 scripts/sync_from_desk.py --source /path/to/compute-desk    # copy, commit, push
python3 scripts/sync_from_desk.py --source /path/to/compute-desk --dry-run
python3 -m unittest discover tests

The script checks the data has rows and the key columns, copies only changed files, and commits only when something changed.

License

Data: CC BY 4.0. Credit "Hundredth Run compute desk" with a link to https://hundredthrun-compute-desk.pages.dev. Code in scripts/ and tests/: MIT.

Corrections are welcome as issues or pull requests, with a link to the provider page.