SF 's Evan Conrad joins the Latent Space podcast to discuss how compute is traded as a commodity but delivered as a complex physical service. Conrad explains why SF Compute built its own cloud operations to reliably deliver GPUs it brokers, rather than relying solely on supply from NVIDIA and major cloud providers.
The conversation covers how interruptions, storage, hardware generations, and workload types shape compute pricing and contract terms. Conrad discusses the difference between batch and live contracts, the risks of cash-settled trades disconnected from actual hardware, and why independent operators have different incentives than integrated vendors.
He also explores how to turn a into a programmable GPU cluster and why NVIDIA and the major clouds cannot serve the emerging market for flexible, independent compute capacity.
This is an AI-written summary. The original reporting belongs to @LatentSpacePod. How we write these · Spot a mistake?