SF Compute
GPU clusters you can reserve and resell
Last reviewed Sep 28, 2026
San Francisco Compute (SFC) builds data centers and GPU clusters and leases them through a market: customers reserve capacity for a delivery window and can resell the hours they do not use back to other tenants. Capacity is sold as 8-GPU nodes on VMs, bare metal, or managed Slurm.
SF Compute does not publish public hourly pricing.
Contact the provider directly for a quote.
Pros & Cons
Advantages
- Reserved capacity can be resold, reducing the cost of over-buying
- Delivery windows are chosen per order instead of fixed multi-year terms
- Public, anonymized order book with quotes, depth, and fills
- No ingress or egress fees
- 24/7 support over Slack, phone, and email
Limitations
- Rates are set by the market and are only visible after signing in, so there is no public list price to compare
- Sold in 8-GPU nodes rather than single GPUs
- InfiniBand is available on bare metal and managed Slurm; on VMs it was listed as planned for Q3 2026
- Market mechanics (bids, asks, delivery windows) add steps compared with a fixed-price on-demand launch
Key Features
Compute market
An order book where buyers place bids and sellers place asks for GPU nodes over a chosen delivery window; the trading engine matches compatible orders continuously
Resell unused capacity
Tenants can sell back reserved hours they are not using, so idle capacity can be recovered as revenue
VMs, bare metal, and managed Slurm
Virtual machines for flexibility, and bare metal or managed Slurm for lowest-level access
Elastic InfiniBand partitions
Programmatically created, isolated InfiniBand networks that instances can join as a job grows (public preview)
CLI and API
The sf CLI and a REST API to buy, sell, and inspect capacity and read the public order book
Refunds for failed nodes
Failed nodes are refunded and workloads can be moved to healthy hardware
Pricing Options
| Option | Details |
|---|---|
| Market orders | Buy orders at a limit price per node-hour for a chosen delivery window, filled against sellers' asks |
| Reserved contracts | Longer reservations with the right to resell unused capacity |
Availability & Support
Support
Shared Slack channels with engineers and on-call coverage for production clusters; enterprise plans include a dedicated solutions engineer
Company
Published by SF Compute to identify the company behind this listing.
- Headquarters
- 🇺🇸 San Francisco, United States
Getting Started
- 1
Create an account
Sign up at sfcompute.com and install the sf CLI
- 2
Create a pool
Set up a capacity pool to hold the nodes you buy
- 3
Buy compute
Place a buy order for a node count, start time, duration, and limit price
- 4
Start instances
Launch instances on the filled capacity, and sell back hours you no longer need
Frequently Asked Questions
What GPU types does SF Compute offer?
Check the pricing table above for SF Compute's current GPU availability and pricing.
How do I get started with SF Compute?
Create an account, Create a pool, Buy compute, Start instances
What are SF Compute's main advantages?
SF Compute's main advantages include: Reserved capacity can be resold, reducing the cost of over-buying, Delivery windows are chosen per order instead of fixed multi-year terms, Public, anonymized order book with quotes, depth, and fills, No ingress or egress fees, 24/7 support over Slack, phone, and email.
What are SF Compute's limitations?
SF Compute's main limitations include: Rates are set by the market and are only visible after signing in, so there is no public list price to compare, Sold in 8-GPU nodes rather than single GPUs, InfiniBand is available on bare metal and managed Slurm; on VMs it was listed as planned for Q3 2026, Market mechanics (bids, asks, delivery windows) add steps compared with a fixed-price on-demand launch.