Your cookie choices

We use necessary cookies to make this site work. With your permission, analytics cookies help us understand usage and marketing cookies support advertising.

You can reject non-essential cookies and still use the site. Change your choices anytime in Cookie settings.

Cookie settingsCookie policy

SF Compute review

Updated September 28, 2026 · DeployCue Editorial Team · Facts checked September 26, 2026

We earn a commission when you sign up through links on this site. Partner status can affect where a brand appears in our rankings. Partners with a live offer in your country come first, then every other brand by score.

Our verdict

SF Compute suits small research teams running short training and inference jobs on bare metal with InfiniBand partitions and per-minute billing. GPU hours start at $3 for H100, H200 and A100, a narrow but well-connected choice for distributed work rather than a broad hardware catalogue.

Key facts

Best for
InfiniBand bare metal on H100, H200 and A100, billed per minute
GPU models
H100, H200, A100↗
GPU hour from
$3↗
Billing granularity
Per minute↗
Egress fee
None (no ingress or egress fees)↗
Interconnect
InfiniBand RDMA partitions (bare metal and managed Slurm today, VMs planned)↗
Legal entity
San Francisco Compute Trading Company LLC↗
Headquarters
United States↗
Support channels
Email, Chat, Phone, Community↗
Pricing model
Hourly billing, Free tier
Official website
sfcompute.com

Facts checked September 26, 2026, 5 sources

Apps and profiles

Elsewhere
XLinkedIn

Pros and cons

Pros

  • GPU hours start at $3; data movement charges are listed as None (no ingress or egress fees)
  • Interconnect is InfiniBand RDMA, with partitions on bare metal and managed Slurm today
  • Support runs on email, chat and phone
  • Workloads can start on hourly and free tier options

Cons

  • The catalogue is limited to H100, H200 and A100, so newer accelerators are absent
  • Interconnect partitions cover bare metal and managed Slurm today, with VMs planned
  • The counterparty is San Francisco Compute Trading Company LLC in the United States, so buyers elsewhere sign with a US entity
SF Compute

Ready to try SF Compute?

Scored 8.6 / 10 by DeployCue.

See alternatives

Who it suits

  • Small research teams that need InfiniBand fabric for distributed training
  • Developers running inference endpoints on H100 or A100 capacity
  • Teams that want to begin on hourly and free tier options before a longer run

Who should look elsewhere

  • SberCloud, whose catalogue runs to B300, H100 and A100 for buyers who need a broader chip range
  • ComputePrices.com, which tracks H100 SXM, A100 SXM and H200 and suits shoppers comparing many accelerators
  • Salad, which serves community workloads on NVIDIA RTX 3060, NVIDIA RTX 3070 and NVIDIA RTX 3070 Ti rather than dedicated bare metal

Plans, egress and billing

GPU hours start at $3, so the entry point is a single named figure rather than a ladder of card tiers. Data movement is settled separately from compute: None (no ingress or egress fees). A team pulling large datasets and pushing checkpoints therefore keeps the total tied to runtime alone, and can size a run against the hourly rate before it starts.

The company behind it

The operator is San Francisco Compute Trading Company LLC, with its home country listed as US, so the contract sits with one named entity rather than a regional subsidiary. That matters for buyers who want a US counterparty and US billing. Support runs on email, chat and phone, putting phone within reach alongside email, chat and community channels.

How SF Compute compares

Where SF Compute is available

Available
Everywhere except Belarus, Cuba, Iran, North Korea, Russia, Syria, Venezuela

Countries from the brand (Source, checked September 26, 2026)

Frequently asked questions

What does an SF Compute GPU hour cost?+

GPU hours start at $3, which places entry pricing above the €0.63 site median. A longer run on bare metal will not hold that figure, so budget the hourly rate for the whole job rather than the opening number.

How is usage billed?+

Billing is per minute, so short jobs stop charging once they finish, and you can begin on hourly and free tier options before committing a full node. Workloads that idle between steps can be stopped and picked up later without paying through the gap.

Does SF Compute charge for data transfer?+

SF Compute lists None (no ingress or egress fees) for data movement, so datasets going in and checkpoints coming out do not add a separate line to the bill. Total spend stays tied to GPU runtime rather than to the volume of data moved.

Which GPUs can I rent?+

The catalogue covers H100, H200 and A100. Buyers who need a wider chip range can check SberCloud, which lists B300, H100 and A100, or ComputePrices.com, which tracks H100 SXM, A100 SXM and H200 across many suppliers.

What hardware sits behind the instances?+

Interconnect is InfiniBand RDMA, offered as bare metal and managed Slurm partitions, with VMs planned. Training jobs that depend on tight node-to-node bandwidth fit this fabric, and GPU hours start at $3 on it.

How do I reach support?+

Support runs on email, chat and phone, so phone sits within reach alongside email, chat and community channels. The operator is San Francisco Compute Trading Company LLC in the United States, and the contract is with that entity rather than a regional subsidiary.

Sources

Compiled from 5 source pages, checked September 26, 2026. We did not open an account.

SF Compute

SF Compute

InfiniBand bare metal on H100, H200 and A100, billed per minute

8.6/10See alternatives