The San Francisco Compute Company
We run a market for large-scale, vetted H100 clusters. Need 800 H100s on a single InfiniBand fabric for an hour? You got it. 8 H100s dedicated for a year? Sure. Need to cancel? Just sell back what you don't need. You're never locked-in to a long-term contract.
Never pay to move your data. We run the clusters from UEFI on up and support you just like a traditional cloud.
*Prices are from the private beta and may change.
Average prices
Clusters starting
(Mar 7)| 8 | 16 | 32 | 64 | 128 | 256 | |
|---|---|---|---|---|---|---|
| 1 hour | $2.98 | $2.98 | $2.98 | $2.50 | $2.40 | $2.40 |
| 1 day | $2.50 | Sign up to buy | $2.40 | $2.50 | $2.40 | $2.40 |
| 1 week | $2.50 | $2.50 | $2.40 | $2.30 | $2.20 | $2.20 |
| 1 month | $2.30 | $2.30 | $2.25 | $2.18 | $2.12 | $2.10 |
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.
A simple but powerful CLI
$ sfc create -n "4 node, 32 A100s at a max price of $1.50/gpu/hr" $ id: 1234 $ status: provisioning $ eta: ~11m $ $ sfc ls $ id name status gpus price $ 1234 4 node, 32 A100s at a max price of $1.50... provisioning 32 $1.23 $ $ sfc ssh 1234 $ [sfcompute] > hostname -f $ ip-172-31-16-1.us-west-1.compute.internal
Used to train hundreds of large scale models
"Very competitive low prices, an abundant amount of H100s for our workloads, and great and responsive customer service."
"We used SF Compute to be our primary cloud for training and inference workloads. The hardware spec is great and the team is amazing to work with."
"SF Compute has made training compute radically easier and cheaper for us, with no compromises in pricing and availability of H100s. We are especially grateful for Compute team's fast and reliable technical support during training."
FAQ
Are nodes fully-interconnected with InfiniBand?
Yes. A 16-node cluster includes 8 H100s with 800Gb/s InfiniBand fabric.
How fast do clusters spin up?
About 2-5 minutes.
Is it bare metal? Containers? VMs? Kubernetes?
If you have a specific need, you'll get a kubernetes cluster. If you'd like a bare-metal cluster, or a VM setup, that can be set up via the console too.
Will the nodes go down?
Yes. Hardware failure rates are much higher on GPU clusters than on CPU clusters. We are designed for failure, so we're shipping you a cluster that can keep going. Unlike other providers, we stand by for failed nodes and will replace them quickly. All system failures are monitored and we'll reach out to you to grow with healthy hardware.
What about networking?
Large clusters get a dedicated Slurm network with our engineering team and emergency phone escalation. Our team has managed 8-9% of humans and hardware on the planet.
Job openings
No openings at the moment, but check back soon!
Who are we?
We have been written about in Bloomberg, The New York Times, and The Verge. We have a team of veterans from Anthropic. You can find us in person in San Francisco.