SF Compute founder explains why standardized compute contracts remain far away

SF Compute founder Evan Conrad says a market for AI computing needs more than buyers and sellers: it needs someone running the infrastructure. In an interview published by Latent Space on October 20, he said standardized compute contracts remain far away. Instead, the company matches orders across different operating requirements rather than treating all capacity as interchangeable.

Conrad described a business that started by trying to sublease its own cluster of computing equipment. Over several years, SF Compute took on the broader operating responsibilities of a cloud provider so it could reliably deliver the capacity customers bought.

He calls that work physical settlement: actually delivering the computing capacity traded through the market. Borrowing from markets where goods such as oil or soybeans are delivered, SF Compute designed its delivery operations and order book—the offers to buy and sell capacity—together.

Conrad offered one example of why standard contracts are difficult: one virtual-machine configuration can suffer a 40% performance drop relative to another from the same operator. He pointed to delays in starting operations on graphics processing units, or GPUs. Those starts happen frequently during model training and inference—the work of running a trained model—reducing effective chip use.

Rather than wait for common standards, SF Compute lets buyers attach multiple requirements to an order, including when capacity must be available. Its matching engine finds sell orders with those properties. Conrad said this allows trading while technology changes, without first deciding which configurations should count as equivalent.

The tradeoff is a wider gap between what buyers bid and sellers ask, Conrad said. Varied requirements spread trading activity across different kinds of capacity. He argued that customers able to use many types could benefit from those price differences.

Conrad characterized the compute market as immature, while describing concentrated trading interest around SF Compute’s small order book. He said customers recover an average of $25 for every $100 spent by reselling capacity. The company published the same average in a September 10 statement; it is a company-reported measure.

Timing creates another distinction. Selling capacity that starts a month from now does not require interrupting work already running, Conrad explained. Reselling capacity in use can require that interruption, making the way a workload stops and restarts part of the market’s operating problem.

Storage and connections between clusters determine how practical those moves are, he said. A workload may stop on one cluster and resume elsewhere if its data can move quickly enough. That requires suitable storage and a fast network path between sites—not simply another available machine.