The San Francisco Compute Company is selling the cheapest access ever to beastly H100 training clusters.²⁸ We're talking bleeding-edge setups with 1024 interconnected H100 GPUs and massive InfiniBand networking.

What's going on here?

SF Compute is renting AI startups time on an H100 cluster by the month or even just weeks.

What does this mean?

San Francisco Compute is offering short-term bursts on thousands of interconnected H100s with InfiniBand. They have one cluster online now, called Angel Island. The next one, Bay Bridge, comes online in December.

Other providers force you to buy a whole year upfront, which costs like $25 million for 1024 H100s. SF Compute is planning to sell the same beast rig for just the months you need, like $2.5 million for one month. It's more per hour than what most folks will do, which doesn't always make it a great choice for inference. But for training very large models, it's cheaper since you can say goodbye to multi-year lock-ins.

Why should I care?

If you're an AI shop wanting to play with the biggest models, you need access to hardware like this. But until now you had to pre-pay tens of millions a year to play. SF Compute also guarantees bare metal grade performance, ML infra support and refund if things go wrong.