Why AI Startups Will Disappear Without Compute Markets — Evan Conrad, San Francisco Compute

Evan Conrad explains how San Francisco Compute, which began as an AI lab that had to sublease its own GPU cluster to stay alive, ended up both operating clusters and running a market for them. He walks through the chain of problems a new lab or inference company hits: everyone is sold out, the deposit eats the bank balance, clusters arrive late and break, and demand is nearly impossible to predict. He argues these are all one risk problem, that AI companies run into credit limits before power limits, and that without a working compute market the price of entry keeps climbing until no new startups can get in. Chapters: - 00:00 Intro - 00:42 Physically settled compute - 01:35 Supercomputing as a service - 02:08 How a struggling AI lab became a GPU broker - 03:32 Two miracle problems - 04:26 Sold out, deposits and delays - 06:20 It's all one risk problem - 08:13 Credit, not power, is the bottleneck - 10:13 What a compute market has to solve - 11:37 Why you have to operate the cluster - 13:05 Why nobody does physical settlement - 13:59 From $10M to $1B deposits - 15:26 Why startups need this Evan Conrad, San Francisco Compute: https://sfcompute.com Join us in SF for the second annual AIE CODE summit: https://aie.codes our last flagship summit of 2026. Applications live now!