In a Latent Space episode recorded at AIE NYC, SF Compute CEO Evan Conrad describes how his company moved from subleasing GPU clusters to building much of the operating capacity of a cloud. He compares it to a hotel: a customer asks for a room, then cleaning, room service, and security, and eventually you discover that delivering the room means running the hotel. For SF Compute, that means handling physical settlement — giving buyers usable compute rather than a cash payment tied to an index.
Conrad argues that making compute tradeable does not require waiting for every GPU cluster to become interchangeable. Instead, SF Compute matches orders with different constraints — GPU generation, networking, software configuration, interruptability, and recovery from storage — and accepts that this spreads liquidity across more variants. The episode also covers spot and forward pricing, batch versus live inference leases, and why an index can become disconnected from the real cost of buying compute.
Independence is central to Conrad's case for a market operator. He says NVIDIA cannot build this market because it would not be independent enough for people to trade on it, and companies with large hardware or inference businesses have conflicting incentives. The episode is a single-source interview, so no contrasting views are included; Conrad's claims are presented as his own.