The Programmable Data Center — Evan Conrad, SF Compute

SF Compute CEO Evan Conrad explains how the company moved from subleasing a GPU cluster to building much of a cloud's operating capacity so it can physically settle the compute traded on its market, delivering usable compute rather than a cash payment against an index. Rather than waiting for GPU clusters to become interchangeable, SF Compute matches orders with differing constraints—interruptibility, storage recovery speed, GPU generation, networking and software configuration—which makes trading possible while technology changes but spreads liquidity across more variants. Conrad also discusses spot and forward pricing, batch versus live inference leases, why an index can drift from the real cost of buying compute, and why he argues a neutral market operator cannot be run by companies with large hardware or inference businesses.