b stack radar
SF Compute logo

SF Compute

5.8/10

SF Compute is a marketplace where AI teams buy NVIDIA H100 and H200 GPU cluster time by the hour and can resell reserved capacity they do not use.

Data last updated November 2026 · Our methodology

Overview

SF Compute, from The San Francisco Compute Company, sells GPU cluster capacity through a market rather than a fixed price list. Buyers reserve NVIDIA H100 nodes, with H200 and B200 also offered, for set periods, often through the sf command-line tool, where they set the most they will pay per GPU-hour. Capacity they reserve but do not use can be sold back through the scheduler. The company says it does not own the GPUs. It holds long-term contracts with data center operators and takes a cut of about 10 percent on each trade. Capacity comes as VMs, bare metal, managed Slurm, or Kubernetes nodes with 3.2 Tb/s InfiniBand, and Elastic InfiniBand Partitions are in public preview. Pricing is hourly and set by the market, varies by zone, and needs no contract. A March 2026 comparison put H100 prices at about $1.64 per GPU-hour, with no egress fees. The product suits AI startups and research labs that need large multi-node training clusters for days or weeks and do not want a multi-year commitment. Its main limitation is the cluster-first design. The web buy page requires at least 4 nodes (32 GPUs) for 4 hours, the CLI requires at least one 8-GPU node, and single-GPU or general-purpose hosting is not practical. It also offers no managed databases, object storage, or other general cloud services.

Our Rating

5.8
out of 10

AI Software5.7/10

Integration
4
Analytics
3
Customization
6
Ease of Use
5
Support
5
Cost
8
Scalability
8
Security
6

Hosting & Infrastructure6/10

Integration
4
Support
5
Cost
8
Customization
6
Security
6
Reliability
6
Ease of Use
5
Analytics
3
Scalability
8

Who SF Compute is best for

Best software for Startups

Good fit

  • It is built so AI startups can use very large amounts of compute for short periods without long contracts.

  • The scriptable CLI is the main interface, which suits ML engineers but not beginners.

Less suited

  • Large cluster deals and cluster ownership are offered through sales, backed by a SOC 2 attestation.

  • The minimum orders of 8 to 32 GPUs make it impractical for general small-business hosting.

Pros & Cons

Pros

  • Unused reserved GPU time can be resold, which lowers the real cost of a commitment
  • H100 prices around 1.6 to 1.8 USD per GPU-hour (March 2026) are at the low end of the neocloud range, with no egress fees
  • Scriptable sf CLI with maximum-price limits allows automated buying
  • Multi-node clusters with 3.2 Tb/s InfiniBand for large training jobs
  • SOC 2 attestation and a published SLA and DPA

Cons

  • Minimum orders of 8 to 32 GPUs rule out single-GPU work
  • Narrow hardware range with no A100, L40S, or consumer GPUs
  • Prices change with the market and vary a lot by zone, which makes budgeting harder
  • Standard VMs lacked InfiniBand as of early 2026
  • Few independent user reviews are available

Key Features

API and CLI
Uptime SLA
Managed Kubernetes
Dedicated Servers
Cloud Compute

Not included

  • Managed WordPress
  • Shared Hosting

SF Compute Pricing

Enterprise clusters / Own a cluster

Contact for pricing
  • Contact sales
  • B200 on request
  • Kubernetes nodes with InfiniBand on request
Get Started

H100 on-demand market (hourly)

Contact for pricing
  • No egress fees
  • No contract required
  • CLI minimum 1 node (8 GPUs)
  • Unused reserved time can be resold
  • Web buy page minimum 4 nodes (32 GPUs) for 4 hours
  • Market price by zone, from about 1.64 USD per GPU-hour as of March 2026
Get Started
Open interactive pricing calculatorGet a tailored cost breakdown for your team size, billing cycle, and add‑ons.

Frequently Asked Questions

Prices are set by the market, hourly, per GPU, and vary by zone. A March 2026 comparison listed H100 from about 1.64 USD per GPU-hour, and the platform takes roughly a 10 percent fee on trades.

The web buy page requires 4 nodes (32 GPUs) for at least 4 hours. The CLI allows a single 8-GPU node, subject to availability.

Yes. Reserved capacity you do not use can be sold back to other buyers through the scheduler.

No contract is required for market purchases. Larger cluster deals and cluster ownership go through sales.

The site points to a SOC 2 attestation in its Trust Center and publishes an SLA, a DPA, and an acceptable use policy.

H100 SXM5 is the main offering. H200 is listed, and B200 is available on request.

Ranked near SF Compute in AI Software

See SF Compute in the list

SF Compute is #627 in AI Software by our score.

  1. #624
    IFTTT logo
    IFTTT
  2. #625
    Revenue.io logo
    Revenue.io
  3. #626
    VelocityEHS logo
    VelocityEHS
  4. #628
    Wondershare EdrawMind logo
    Wondershare EdrawMind
  5. #629
    Oracle CX Sales logo
    Oracle CX Sales
  6. #630
    LiDAR360 logo
    LiDAR360