SF Compute builds a compute exchange with physical delivery

SF Compute founder Evan Conrad described building a marketplace for trading future GPU time, modeled on commodity exchanges. Trades settle with actual compute delivery: about 5% of capacity sits on the order book, and a customer who buys $100 of capacity recovers $25 on average by reselling it.

• About 5% of SF Compute capacity is on the order book, but liquidity is high
• A customer buying $100 of capacity recovers $25 on average via resale
• Performance gap between two VMs from one operator reaches 40%
• A B300 earns about $40/hour on inference versus $5–6/hour cost