Listen/watch: Latent Space episode page with audio, transcript and video link. Published 20 October 2026; 37 minutes.
Who: Evan Conrad, introduced as SF Compute’s CEO. The episode discusses compute trading, physical delivery, interruption and the differences between capacity contracts.
Useful takeaway: GPU-hours are not automatically interchangeable if networking, storage, availability or interruption behaviour differ. This is particularly useful for a team comparing batch training with live inference, where the cost of losing a running job or delaying a request can be very different.
Buyer exercise — proposed: turn the infrastructure quote into a requirements checklist: hardware, usable memory, interconnect, storage, region, interruption rules, recovery time and support. Ask whether quoted capacity is usable now, reserved for later or conditional. Do not treat trading/resale anecdotes as a guaranteed recovery value.
Question to take into procurement: “Which contract and technical characteristics prevent us from treating these two compute quotes as equivalent?”