The Programmable Data Center — Evan Conrad, SF Compute
- ID
- 37539
- Status
- summarized
- Published
- 21 Oct 2026, 3:31 AM
- Fetched
- 21 Oct 2026, 3:50 AM
- Provider
- Latent Space
- Category
- developer-ai
- Original URL
- https://www.latent.space/p/sfcompute2
- Source URL
- https://www.latent.space/feed
Summary
- Score
- 6.0
- Created
- 21 Oct 2026, 3:50 AM
- Tags
- Audience
- developersai_ml_learnerssaas_founders
What happened
SF Compute's Evan Conrad explains the company's shift from subleasing a GPU cluster to operating much of a cloud so it can physically settle traded GPU-hour contracts, not just cash-settle against an index. The episode details how contracts encode constraints like interruptibility, storage-recovery speed, GPU generation, networking, and software configuration, and why SF Compute matches orders across variants rather than waiting for fully interchangeable clusters. It was recorded at AI Engineer NYC's new Compute Trading track.
Why it matters
If you rent GPU capacity, the practical question is what the contract guarantees: a cheap GPU-hour may come with interruption, slow storage recovery, or a specific GPU/network configuration that changes your real job cost. Builders evaluating reserved capacity should ask for those settlement terms explicitly, because SF Compute says its market spreads liquidity across variants while clusters remain non-interchangeable.
Discussion angle
When you book GPU hours, would you accept a lower price for interruptible capacity with a defined recovery time, or do you need guaranteed non-interruptible settlement? What constraint would you want spelled out in a contract?