Offering overview
There is no fixed rate card at this San Francisco company; it runs a two-sided market instead: buyers place orders with a maximum price per node-hour, sellers set a floor, and the scheduler matches them. Reserved capacity that goes idle can be sold back, which is the core pitch to teams that would otherwise overbuy for training runs. Capacity comes as 8-GPU H100 and A100 nodes delivered as VMs, bare metal or managed Slurm, with InfiniBand networking exposed through Elastic InfiniBand Partitions (public preview). An sf CLI handles buying, reselling and job submission from the terminal. SKU names in the docs point to sites around Amsterdam, Northern Virginia, Chicago, Seattle and Tokyo. Because prices float with supply and demand, there is no published starting price; the homepage uses an illustrative $3.00 reserved versus $4.29 realized per GPU-hour comparison. The service suits research labs and startups running large multi-node jobs who are comfortable with market-based procurement.
- Management
- Not stated
- Infrastructure
- Not stated
- Pricing model
- Not stated
- Support
- Not stated
- SLA
- Not stated
- Best for
- Model Training, Fine-Tuning, Research And Experimentation, Batch Processing, Model Inference
Technical specifications
- Containers
- Unknown
- Kubernetes
- Unknown
- Root access
- Unknown
- API access
- Unknown
- Autoscaling
- Unknown
- Persistent processes
- Unknown
AI capabilities
- Model serving
- Not stated
- AI frameworks
- Not stated
- Inference endpoints
- Unknown
- Fine-tuning
- Unknown
- Vector database
- Unknown
- Model catalog
- Unknown
- Monitoring
- Not stated
- Evaluation
- Not stated
Security & compliance
No controls are stated.
Geography
Provider headquarters are separate from offering infrastructure and availability.
Infrastructure regions
Netherlands, Europe · Japan, Asia · United States, North America
Data-residency jurisdictions
Not stated
Service availability
Not stated
Support coverage
Not stated
Available plans
Pricing can change. Promotional prices are shown separately from standard prices.
No active plans are recorded for this offering.
Questions & answers
- How does the SF Compute market work?
- Buyers place orders with a maximum price per node-hour, sellers set a floor, and a scheduler matches them. Prices vary with supply and demand rather than following a fixed rate card.
- Can users resell unused reserved GPU time?
- Yes. Reserved capacity that goes idle can be sold back through the market.
- What cluster deployment options are documented?
- Capacity consists of 8-GPU H100 and A100 nodes delivered as virtual machines, bare metal or managed Slurm. Elastic InfiniBand Partitions expose InfiniBand networking in public preview.
- What can users do with the sf CLI?
- The sf CLI supports buying compute time, reselling it and submitting jobs from the terminal.
- Who is SF Compute suited to?
- The service suits research labs and startups running large multi-node jobs that are comfortable with market-based procurement.
Data sources & provenance
Where this information came from, how it was checked and when it was last looked at. “Last updated” records a change to the record — it does not mean the record was verified.
- Data quality
- Data quality: Unverified
- Primary source
- sfcompute.com
- Sources cited
- 2
- Last reviewed
- 7 October 2026
- Last updated
- 18 October 2026