What's San Francisco Compute?

San Francisco Compute sells time on large-scale GPU clusters (thousands of GPUs) on contracts you can get out of. Normally, when you buy a large cluster, you're forced into a long-term agreement that puts your business at risk. We let you sign long-term contracts you can "exit", by selling back to someone else on our liquid market.

Unlike a reseller, we operate the clusters like a cloud provider, with our own virtualization and bare-metal stack.

The way we work with customers

SFC onboards only a few customers at a time, by signing an NDA, sitting down directly with your engineering team, proving out a pilot, aligning our roadmap to you so we can bring down your costs & keep you out of long-term contracts.

We've had a lot of success with this model and today operate thousands of chips that our customers can scale up and down with ~1hr average lead time.

The mission

We're going to secure the financial risk of the largest infrastructure build-out in the history of the world.

When people finance clusters, the data centers that house them, and the power that powers them, they need "offtake". Offtake just means someone who's signed a contract to buy the thing you're building for some period of time. It's pretty hard to write a loan against a cluster that isn't sold to someone. The lender doesn't want to take on the risk that the cluster developer can't sell it, and the cluster developer really does not take on that risk. So instead, they ship it to the customer, by forcing folks to buy risky long-term contracts. Since margins are thin & volumes are large, if you mess up that purchase, that's game over. A 20% utilization gap might mean the difference between profit & bankruptcy.

If you don't solve that risk, there's a bubble. Application layer companies sign multi-year contracts for compute and inference, but sell to their customers on monthly subscriptions. This isn't SaaS anymore. A minor shift in your growth rate could cause the revenue to fail to come in the door, the compute bill to still come due, at the same time the venture capital slows down.

As AI scales, the folks who can more effectively take on that risk increasingly become the largest and largest players. After all, a 2-person startup in a San Francisco Victorian can't realistically sign a 5-year take or pay contract on $100m supercomputer. But they may be able to buy the month of liquidity that someone else sold back.

So that's what we make: a liquid market for GPU offtake.