SF Compute CEO outlines vision for programmable data centers and liquid compute markets

SF Compute, a company trading GPU compute hours, describes its expansion from subleasing clusters to operating what it calls a programmable data center, handling physical settlement so buyers receive usable compute rather than cash against an index.

Evan Conrad, CEO of SF Compute, argues that independent markets require neutrality because companies with large hardware or inference businesses face incentives that make it harder to run a neutral exchange, contrasting with vertically integrated cloud providers.

The company matches orders with different constraints—such as interruptibility, recovery from storage, GPU generation, networking, and software configuration—reflecting physical realities that affect both contract terms and pricing.

Conrad positions SF Compute’s approach as analogous to software-defined networking, which separated control and data planes to virtualize routing, suggesting a similar architectural shift for GPUs and supercomputing.

SF Compute’s market supports spot and forward pricing and distinguishes between batch and live inference leases, aiming to reflect the actual cost of acquiring compute rather than relying on abstract indices.

The company highlights engineering challenges such as visiting sites, designing clusters, running the backplane, and making infrastructure programmable, including the ability to subdivide InfiniBand fabric on the fly.