A supercomputer built for startups
Hello! We're the San Francisco Compute Company, and as far as we can tell, we now sell the cheapest H100 training clusters in the world.
For example, say you want 1024 H100s interconnected with 3.2tb/s InfiniBand, and petabyte-scale storage. At that scale, every other provider will force you to sign for at least year, which is about $25m.
We'll sell you just for a month, for about $2.5m. 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 a company-defining purchase.
Wait, so what's the price?
About market rate, potentially even more expensive per-hour than if you bought on a long-term deal. The "cheapest" part is because we'll actually sell you a tiny sliver of time on it. If you only need 2 weeks, we'll actually sell you those 2 weeks.
Is it online now?
We have one cluster online now, called Angel Island. The next one, Bay Bridge, comes online in the next few months.
Do I just get SSH access?
Do you have capacity?
Well that depends on when you're reading this, but as of writing this article, yes, lots. Since we sell bursts, we tend to have a shorter waitlist than other folks. We just have more slots for folks to fit into.
What's the catch?
If you have an predictable constant amount of compute needed for inference, we are probably not for you. Our clusters have very large high speed disks and InfiniBand networking, which is great for training but it is unused for inference.
But if you want to train a scaled model before you have tens of millions of dollars in the bank, we are the right solution for you. Other cloud providers won't even talk to you without a couple months notice and a couple months of down payment.
No, I mean, what's the catch for SF Compute?
Oh, if you're asking why other folks don't do it, it's probably because it would make their lives harder. If you have H100 allocation, your incentive is to move up market and sell long-term deals. Long-term deals give you better deals on financing (loans) and so your expenses are lower. Your cashflow is better, since people are often paying you upfront, so you need less financing. Most importantly, long-term deals shelters you from the "what happens if the market shifts really hard" risk.
So why do it?
So, like, we could say our business model "is a really big bet on people training big models in the future" or something. These seem like good things to say. They sound like smart business words.
But more honestly, we started this because no one would sell to us at prices we could afford, on the terms we wanted. We were trying to make "Midjourney for music", everyone told us the minimum purchase was all the money in our bank account plus at least two of our limbs.
We made SF Compute to sell to ourselves. The rationale behind SF Compute is much closer to "if no one sells us compute without 3 year contracts our startup will die."
We're excited to see what you build.