Village Global Podcast Worldbuilders: The Largest Infrastructure Project in History with Evan Conrad (SF Compute)
Mar 13, 2026
Evan Conrad, founder and CEO of SF Compute, who turned an accidental GPU cloud into a supercomputing marketplace. He describes surviving huge monthly cluster costs and building a manual order book for GPU time. He explains why GPUs need multiyear contracts, what offtake means, and the Marriott-like financing model for building and operating large GPU clusters.
AI Snips
Chapters
Transcript
Episode notes
GPU Contracts Require Multi Year Commitments
- GPU clouds require multi-year leases because providers finance clusters with loans that need multi-year payback schedules.
- That structure forces customers to lock in long contracts or leave clouds exposed to catastrophic churn risk.
Pre Sell Compute Products Instead Of Assuming Markups
- Pre-sell to customers and build what they will actually pay for rather than assuming they'll pay extra for software on top of GPUs.
- Talk to customers, validate demand, then design product pricing to match founders who already committed capital to GPUs.
From Spreadsheets To An Order Book For GPUs
- With no cash, SF Compute became an accidental broker, building a manual order book and matching buyers and sellers of multi-month GPU blocks.
- That spreadsheet-and-phone approach later automated into a financial-style order book for compute time.