specification: API Commons Plans
specificationVersion: '0.1'
schema: https://raw.githubusercontent.com/api-evangelist/interface-research/main/schema/api-commons.yml#/$defs/Plans
provider: SF Compute
providerId: sfcompute
created: '2026-06-21'
modified: '2026-06-21'
reconciled: false
tags:
- GPU
- Compute
- Marketplace
- H100
- Spot Pricing
- Plans
description: >-
  SF Compute is a market, not a fixed-rate cloud. There is no single price per
  GPU/hour - each block of compute is priced independently by the orderbook,
  and the rate a buyer pays is whatever clears against open sell orders for the
  requested instance type, quantity, start window, and duration. The /v0/quote
  endpoint returns an indicative price for a desired block; published marketing
  figures (e.g. "H100 nodes from ~$1.94/gpu/hr") are starting points, not
  contractual rates. Sellers can list unused contract time back to the market.
  Prices below are reconciled=false because they float with the market.
notes: >-
  Per-GPU-hour rates are market-driven and change continuously; the from-price
  is a marketing reference observed on sfcompute.com/prices and should be
  re-verified at reconciliation via /v0/quote. Billing is in USD cents.
sources:
- https://sfcompute.com/prices
- https://docs.sfcompute.com/docs/how-the-market-works
- https://docs.sfcompute.com/docs/using-the-api
plans:
- id: sfcompute-market
  name: GPU Compute Market
  type: usage
  description: >-
    Spot-priced market for blocks of GPU-hours. Buyers place buy orders at a
    price they are willing to pay; orders fill against the orderbook and resolve
    into contracts. Each block is priced independently - no flat per-GPU-hour
    rate.
  entries:
  - label: H100 (Kubernetes, h100i) per GPU-hour
    name: h100i_gpu_hour
    type: usage
    metric: gpu_hours
    limit: -1
    timeFrame: usage
    geo: global
    unit: 1
    price: market-priced per block (from ~$1.94/gpu/hr reference); see /v0/quote
    userMultiplied: false
  - label: H100 (Virtual Machine, h100v) per GPU-hour
    name: h100v_gpu_hour
    type: usage
    metric: gpu_hours
    limit: -1
    timeFrame: usage
    geo: global
    unit: 1
    price: market-priced per block; see /v0/quote
    userMultiplied: false
  - label: H200 (Kubernetes, h200ki) per GPU-hour
    name: h200ki_gpu_hour
    type: usage
    metric: gpu_hours
    limit: -1
    timeFrame: usage
    geo: global
    unit: 1
    price: market-priced per block; see /v0/quote
    userMultiplied: false
  - label: Reselling (sell unused contract time)
    name: resale
    type: usage
    metric: gpu_hours
    limit: -1
    timeFrame: usage
    geo: global
    unit: 1
    price: seller-set price clears against market buy orders
    userMultiplied: false
  elements:
  - name: Orders / Market
  - name: Contracts
  - name: Prices / Quote
  - name: Clusters / Nodes
  - name: Reselling
- id: sfcompute-inference
  name: Inference Fleet
  type: usage
  description: >-
    OpenAI-compatible inference served on SF Compute capacity. Pricing follows
    the underlying GPU capacity used to host the fleet.
  entries:
  - label: Inference Capacity
    name: inference_capacity
    type: usage
    metric: gpu_hours
    limit: -1
    timeFrame: usage
    geo: global
    unit: 1
    price: derived from underlying market GPU capacity
    userMultiplied: false
  elements:
  - name: Inference
- id: sfcompute-enterprise
  name: Enterprise / Reserved
  type: enterprise
  description: >-
    Larger reserved capacity, dedicated clusters, and negotiated terms for
    sustained pre-training workloads. Contact SF Compute.
  entries:
  - label: Reserved Capacity
    name: reserved
    type: flat
    metric: contract
    limit: -1
    timeFrame: usage
    geo: global
    unit: 1
    price: contact sales
    userMultiplied: false
  elements:
  - name: Reserved Clusters
  - name: Dedicated Support
maintainers:
- FN: Kin Lane
  email: kin@apievangelist.com