SF Compute aims to make GPU compute markets programmable

At AIE NYC, Evan Conrad of SF Compute unveiled his vision for an independent trading market for GPU compute hours. According to the announcement, SF Compute now enables buying or selling GPU compute hours with programmable control over resources, including networking and memory. The company has evolved from simply subleasing clusters to managing a full cloud infrastructure, including physical delivery of compute (physical settlement), where buyers receive usable resources rather than just a cash settlement against an index.

The market relies on solving complex technical constraints: GPU generation compatibility, software configurations, workload interruption tolerance, and recovery times. SF Compute matches orders to the specific constraints of available clusters, allowing liquidity to persist despite diverse offerings. The company also addresses spot and forward pricing, batch versus real-time inference, and the risks of index-based pricing disconnecting from actual compute costs. A key focus is market independence, with Evan Conrad arguing that dominant hardware or inference players have conflicts of interest that make neutrality difficult.

This model could reshape how businesses access compute resources, offering flexibility akin to financial markets. However, success hinges on gradually standardizing infrastructures while maintaining sufficient liquidity for existing variants.