About the San Francisco Compute Company
Hi, we're the San Francisco Compute Company.
We make supercomputers. We're in San Francisco. Unlike other providers, who force you into year-long or three-year-long deals, we almost exclusively sell short-term bursts on very-large supercomputers for very-big training runs.
If you want 512 H100s, it would normally cost you in the $12m range, because you gotta buy it out for the whole year. If a provider can sell someone the whole year, then they'll just do that and print money. We'll sell it to you for two weeks (around $500k).
Originally, SF Compute started as Alex Gajewski and Evan Conrad trying to make a generative music model. We flew to LA, Alex dyed his hair (it was great), and played around with some different architectures. We liked this auto-regressive VAE concept. When we went to scale up our model, we reached out to every provider at the time, who informed us that the minimum purchase was a year, and would cost at least $1m. We did not have $1m.
No one wants to sell you just a month for a pretty good reason: you make less money doing so. If you have a big cluster, you're better off selling it on a 1 to 3 year contract and getting paid all up-front. It's way less risky, so you get cheaper cost of capital, and make higher margins. Why sell to Junelark, the 2-person music startup, who has almost nothing in the bank and could die before the contract ends?
Unfortunately for us, we were Junelark, the 2-person music startup. So were most of our friends. So, we got fed up, took the risk, and started a cloud provider.