Skip to content
Main Site News Console

The Programmable Datacenter — Evan Conrad, SF Compute

· Latent Space Translated
播客深度访谈

From GPU Trading to Physical Settlement, Independent Markets, and Programmable Compute

Image 3

Latent.Space

October 20 · 37 min

Recorded live during AIE NYC’s brand-new Compute Trading track. We look forward to seeing everyone again at AIE CODE, our next flagship event for the world’s top agent engineers.


In the early days of cloud computing, past guest Martin Casado built a technology so disruptive that Cisco felt threatened and offered $600M to acquire it, only for VMware to swoop in with a winning $1.3B bid:

View post on X ↗

That concept was software-defined networking (SDN), which virtualized routing networks by decoupling the control plane from the data plane. Today, that revolution is coming to GPUs. Where building a supercomputing cluster once required a $50M–$100M upfront commitment and wrestling with the entire stack from hardware and firmware down to the kernel, SF Compute now makes it possible not only to trade current and future GPU hours, but also to programmatically control resources from networking all the way down to memory.

Last time, we spoke with Evan about commoditizing compute. At AI Engineer NYC, we went deeper: What does it take to physically settle a GPU contract? What conditions must be met before a compute market can actually have liquidity?

Video 1

From Renting Rooms to Running a Hotel

Evan uses a hotel analogy to describe SF Compute’s evolution. Customers start out just wanting a room, but soon ask for housekeeping, room service, and security. In the end, you realize that to deliver the room, you have to run the entire hotel. SF Compute initially tried subleasing clusters; since then, to deliver the compute traded on its marketplace, the company has gradually built out full operational capabilities rivaling those of a cloud provider.

This is physical settlement: buyers receive actual, usable compute, not just cash settlements benchmarked against an index. This brings the cluster’s physical constraints directly into the trades themselves. Can workloads be interrupted? How fast can they recover from storage? Which GPU generation, network configuration, and software environment are required? All these details directly shape contract terms and pricing.

Making Compute Tradable

Rather than waiting for every GPU cluster to become completely fungible, SF Compute matches orders across varying constraints. Evan argues that this allows trading to happen even as technology rapidly evolves, with the trade-off that liquidity gets fragmented across more distinct specifications. We also discussed spot versus forward pricing, batch versus real-time inference, and why index prices can become detached from the real-world cost of procuring compute.

Independence is Evan’s central argument for a dedicated market operator. In his view, companies with massive hardware or inference footprints have conflicting incentives, making it difficult to operate a truly neutral exchange. We also dove into the gritty, critical engineering details: datacenter site visits, cluster design, operating backplanes, and making infrastructure programmable. Near the end of the conversation, he illustrated this with their ability to dynamically partition InfiniBand fabrics.

Guest: Evan Conrad, CEO of SF Compute.

Show Notes

Timestamps

  • 0:00 — Why Nvidia can’t do this

  • 0:25 — Welcome back to Latent Space

  • 2:23 — From renting rooms to running a hotel

  • 3:47 — Physical settlement: Actually delivering compute

  • 6:00 — Liquidity and GPU order books

  • 8:42 — Trading non-standardized compute

  • 10:00 — Interrupting and migrating workloads

  • 12:20 — Why hardware cohorts matter

  • 14:32 — How long can compute contracts run?

  • 16:19 — Batch vs. real-time inference leases

  • 17:58 — The shape of compute…

#播客#深度访谈#4Allapi.com

Published by the 4All API team

Original link:https://latent.space/p/sfcompute2/

Need a mainstream LLM API? 4All API gives you one key to call OpenAI, Anthropic, Google Gemini, Qwen, DeepSeek, and dozens more — at official-pass-through pricing, with enterprise-grade reliability, integrated in 5 minutes.

Sign up for the 4All API console →