San Francisco Compute

San Francisco Compute builds and operates data centers, GPU clusters and a cloud for AI workloads, offering virtual machines, bare metal and managed Slurm. Its customers can reserve capacity and list unused portions for resale, including entire future reservations or gaps between jobs. That gives teams a way to recover some of their spending when plans change, though listing capacity does not guarantee a sale: customers keep the allocation until a buyer purchases it.

Evan Conrad and Alex Gajewski started the company, which grew out of an AI lab that had committed to an oversized GPU cluster and needed to sublease it. That experience shaped its business: SFC operates the clusters behind its market rather than simply brokering access to another provider’s machines. It also builds and operates clusters on behalf of outside owners, allowing it to expand capacity while retaining technical control.

How capacity changes hands

Abel, SFC’s settlement ledger, connects capacity trading to the machines that deliver it. Compute ownership is a schedule—how many nodes a customer holds, on which cluster, over which period—rather than a single balance. Each trade transfers payment and compute together, so a buyer cannot pay without receiving the allocation. The cloud control plane reads that schedule to enforce how many nodes the customer can run. A team whose training run finishes early can sell the remaining time without transferring the rest of its reservation.

The market connects customers with different capacity needs: Standard Intelligence reserves for peak demand and resells unused hours, while Roboflow buys shorter reservations for individual jobs. On September 10, 2026, SFC announced $245 million in compute contracts with one AI company, covering two dedicated NVIDIA Blackwell B300 clusters under two 36-month agreements. The announcement scheduled capacity to come online through late 2026 and into 2027.

Agent-driven experiments

SF Autoresearch adds an experimentation platform that buys compute through SFC’s order book. Developers can connect their own agents to a research harness that executes experiments and records results; the agent decides what to test next. Performance measurements help compare changes, while sandboxes capture disk and memory state so reinforcement-learning experiments can fork from a prepared environment instead of repeating setup for every attempt. Fully managed research agents, given source code, an objective and a budget, are in private preview.