Find the cheapest GPUs you can actually trust.
Tell us what you need. We rank real open-market listings by total cost for your job, not the sticker price, and show where every number comes from.
What do you need? · Results update as you choose
Tell us what you need. We rank real open-market listings by total cost for your job, not the sticker price, and show where every number comes from.
Query the same provenance-aware GPU catalog people see on Compute Pulse, follow every observation to its source, and use a declared contract instead of scraping the interface.
GET /api/indexes exposes independent H100, H200, B200, and A100 public ask benchmarks. Each fixes memory, 8-GPU offer size, provider segment, geography, and on-demand term; the models are never normalized together. Thin contracts return a building state and null value. A dated SemiAnalysis public composite is returned as an attributed external cross-check where available; its different contract/source basis is explicit, and it never rescales the Compute Pulse value.
The marketplace endpoint exposes only fields already published on listing pages. It omits contacts, private ownership data, source leads, and internal workflow fields. Filter by intent, GPU, region, current availability, known price, minimum capacity, or interconnect.
https://compute-pulse.com/api/supply?intent=selling&gpu=h100&available_now=true&price_known=true&sort=price&limit=50Use GET /api/pages to discover every canonical static page and dynamic page family, then use the sitemap or declared data endpoint to enumerate records.
Filter by GPU, provider, region, billing type, observed availability, or price. Results are bounded to 200 rows per request and include explicit pagination links.
https://compute-pulse.com/api/listings?gpu=nvidia-h100&available=true&max_price=3&sort=price_asc&limit=50The response states whether it came from the live/cached catalog or degraded seed fallback. Null availability means unknown; it never means available.
Prices are observations in USD per GPU hour, not binding quotes. Each result retains its collection method, fetch time, optional upstream update time, and source URL. For a recommendation that could trigger a purchase, inspect the source and cite the canonical Compute Pulse GPU page alongside it.
Verification levels run from public observation (0), through provider-submitted (1) and human-checked (2), to a currently valid firm quote (3). The API re-evaluates that level at request time so expired quotes cannot retain the highest rung.
The OpenAPI contract also describes RFQ, capacity, and double-opt-in subscription endpoints. An agent should show the material payload and obtain explicit user approval immediately before any POST. Capacity submissions remain unpublished until moderation.
Account, authentication, admin, and cron routes are intentionally outside the public agent surface. Agent friendliness does not weaken those boundaries.
Read the terms, privacy policy, and affiliate disclosure. This interface does not grant a separate bulk-redistribution license. Report incorrect data to harry@compute-pulse.com.