Skip to content

docs: describe GPU Cloud API v3, spot instances and the new GPU cloud console section - #146

Merged
justprosh merged 8 commits into
mainfrom
feat/gpu-cloud-v3
Nov 13, 2026
Merged

justprosh merged 8 commits into
mainfrom
feat/gpu-cloud-v3

Conversation

@justprosh

@justprosh justprosh commented Nov 13, 2026 •

Copy link
Copy Markdown
Member

What changed

  • GPU Cloud API (build/api/gpu_cloud.md): rewritten for /v3/gpu. Covers scopes, the offer endpoints (offers, groups, facets, picks), quotes, creating instances with Idempotency-Key, the instance lifecycle, rename and container updates, logs and events, deletion, and errors. The legacy /gpu API moves to a short section at the end. The old /gpu/api-docs reference link returns 404, so it is removed.
  • GPU Cloud concepts (build/gpu_cloud/overview.md): adds offers and providers, on-demand and spot instances, and the /v3/gpu statuses. Billing now uses the offer's billing step and minimum charge in place of hourly charges.
  • Rent a GPU instance and Manage GPU instances: rewritten for the console's GPU cloud section (Marketplace, Create GPU instance, Instances).
  • API overview: the versions table adds GPU v3 and legacy GPU. The scopes table adds gpu:read and gpu:write.
  • Billing wording: when billing starts and stops, the minimum charge for a short-lived instance, and when charged.total is final, in the same words on the concepts, manage, and API pages.
  • Payment and balance management: adds the GPU Cloud usage block of the Billing page.
  • Legacy GPU API: marked deprecated on the GPU Cloud API page and in the API versions table. It still works; the pages ask to move to /v3/gpu.
  • Overview, Registration, x402: GPU billing wording, the console sidebar entry, and a note that GPU Cloud is not open to every account.
  • GPU Cloud API, full v3 coverage: adds the offer filters and sort options, the offer fields (offerKind, hostTier, availability, preemption, capabilities, provisioning, osImages), rejected and largestOnSale of offer picks, the quote response, instance fields (statusDetail, isBilling, placement, charged, actions), list filters, fleet_too_large and other common error codes, X-Request-Id, and GET /v3/gpu/openapi.json.
  • Volumes: a short section says that volumes are not available yet and that the volume API may change before it opens. The volume endpoints are not documented in detail.
  • The GPU tools of the MCP server are left out on purpose. They are not available in production yet.

How each fact was checked

  • Endpoints, fields, statuses, error codes, and the idempotency and update rules: read from the published spec at https://api.fluence.dev/v3/gpu/openapi.json (version 3).
  • Price fragment, providers, billing steps and minimums: observed in live anonymous GET /v3/gpu/offers responses on 2026-11-13.
    • Most offers bill in 60 s steps with a 1200 s (20 min) minimum.
    • 12 offers bill in 3600 s steps with a 3600 s minimum.
    • Spot offers exist, all with noticeSeconds: 0 and dataSurvives: false.
  • Which endpoints answer without credentials: observed on 2026-11-13. GET /v3/gpu/offers, GET /v3/gpu/offers/{id}, and GET /v3/gpu/openapi.json return 200. offer-facets, offer-picks, default-images, instances, and volume-offers return 401.
  • No offer takes a volume: observed. All 344 offers listed on 2026-11-13 have capabilities.volumes: 0.
  • Offer kinds: observed. Akash offers are market_profile, Lium offers are host_offer, other providers are provider_sku.
  • Error body shape: observed on a live 401 from GET /v3/gpu/instances.
  • Legacy /gpu still answers: observed. GET /gpu/plans/ returns 401 without a key, and GET /gpu/default-images/ returns 200.
  • Console labels and flow: read from the source of the console version currently deployed to production. They were not checked on screen; see "Not verified".

Not verified

  • Console screens were not seen live. All console labels come from the source of the deployed version.
  • Legacy scopes on /v3/gpu: that akash:write and spheron:write keys work on /v3/gpu comes from the service team and the backend code. No live call was made.
  • Billing when the balance runs out: that instances end within minutes and their disks are erased comes from the service team and the console text. It was not observed.
  • Legacy API billing: that /gpu instances use the same billing step and minimum comes from the service team.
  • New accounts and GPU Cloud: that new accounts may not have GPU Cloud access comes from the service team and the console's "GPU Cloud is temporarily unavailable" screen. It was not checked with a new account.
  • Spot instances stopped from inside: that a stop from inside ends a spot instance is confirmed for some providers, but not for every provider that sells spot.
  • Create request example: built from the spec, with placeholders for the offer ID and the OS image. It was not executed.
  • Em dashes: two bold UI labels on the Manage page keep them because the console renders them that way.
  • Clusters: the spec has a nodes filter and requirement, but production refuses nodes with invalid_input. Clusters are left out of the page.
  • Volumes closed: that the volume offer list is empty with credentials comes from the spec and the service team. It was not called with a key.
  • Instances that do not become ready: that a launch that is not ready in time ends as failed, and that a running instance that stops being ready stays billed until deleted, comes from the service team. It was not observed.
  • Unknown GPU model: observed. A list filter with an unknown gpuModel returns an empty list. That a quote or a create request returns invalid_input comes from the service team.
  • GPU Cloud usage block: the label and its columns come from the service team. The block was not seen on screen.

@vercel

vercel Bot commented Nov 13, 2026 •

Copy link
Copy Markdown

The latest updates on your projects. Learn more about Vercel for GitHub.

Project Deployment Actions Updated
fluence-dev-docs Ready Ready Preview Nov 13, 2026 7:26pm UTC

Request Review

@justprosh
justprosh merged commit 20d803c into main Nov 13, 2026
3 checks passed
@justprosh
justprosh deleted the feat/gpu-cloud-v3 branch November 13, 2026 19:27

This branch was successfully deployed

1 active deployment
Preview — b4373f5a Deployed Nov 13, 2026 by vercel[bot]
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant