All clusters are vetted and meet minimum technical requirements. Deploy on-demand or reserve instances with no long-term contracts.
NVIDIA H100 SXM5 80GB VRAM
24/7 support (Slack, phone, email)
2TB+ NVMe storage (high IOPs)
1TB+ of RAM
No ingress/egress fees
3.2Tb/s InfiniBand (fully interconnected)
100% uptime SLA
Pricing shown per gpu/hour based on selected duration and cluster size. Need a custom configuration? Use our CLI or go directly to the buy page.
Get competitive pricing on on-demand clusters up to 2k GPUs or reserved clusters up to 4k interconnected GPUs.
Choose between reserved and on-demand compute with market-driven pricing.
Reserved
Lock in fixed prices for 1 hour to multiple months. Never get preempted. Sell excess compute back to the market anytime.
On-demand
Get the lowest prices with 1-hour compute blocks that auto-renew at current market rates. Set a max price to control your risk—you'll only be preempted if market rates exceed your limit.
Try this simulation of on-demand with real 30-day pricing data
Buy on
Simulated orders
Auto-renewed 30 times at an average of $1.12/gpu/hr. Cluster remained active throughout the simulation.
Buy the cluster on-demand with the CLI.
sf buy -n 8 --duration 1h
Enable auto-renewals.
sf scale -n 8 --duration 1h --price 2.32
Yes. A 10-node cluster includes 3.2Tb/s InfiniBand fabric.
About 0.5 seconds.
If you buy on the website/CLI, you'll get a Kubernetes cluster. If you'd like a bare-metal cluster, or a VM-setup, reach out to us via the contact form.
Yes. Hardware failure rates are much higher on GPU clusters than on web servers. At certain scales, they're guaranteed, so we've designed for failure. We have strict hardware requirements and have seen just about everything that can go wrong. Unlike other providers, we refund for failed nodes and can repack your nodes to ones with healthy hardware.
Large clusters get a dedicated Slack channel with our engineering team and emergency phone escalation. Our team has managed $1B+ of hardware and worked on several of the world's supercomputers.