Evan Conrad

Evan Conrad is co-founder and chief executive of San Francisco Compute, which builds and operates GPU clusters that customers can reserve and sublease. His focus there is tradable GPU time: AI teams can secure capacity ahead of demand, then offer unused hours for resale when their plans change.

Before working on AI infrastructure, Conrad created Quirk, an open-source cognitive behavioral therapy app for iOS and Android, and built a Y Combinator-backed company around it with his brother. The subscription model exposed a conflict: some users felt better and canceled, while others kept paying without feeling better. Conrad concluded that retention could reward the wrong outcome. In 2020, the team shifted the same company to Room Service, infrastructure for real-time collaboration such as shared cursors and simultaneous editing. Quirk is no longer maintained, although its GPL-licensed code remains available.

Reflecting on the pivot in his 2021 essay “Moral Competence”, Conrad urged social-good founders to judge their work by its benefits, examine existing solutions, and put solving the problem ahead of their own involvement in it. He included his own company in the criticism: attachment to a mission could make it harder to admit that a product was not helping enough.

San Francisco Compute grew out of Junelark, a small AI lab that had committed to more GPU capacity than it needed and had to sublease it to stay in business. Conrad and his colleagues turned that problem into a cloud service with an integrated resale market. The company operates the clusters behind the market, taking responsibility for delivering the compute customers buy. Long contracts provide the customer commitments lenders need to finance new clusters. Customers can offer unused capacity for resale, from a single hour to the rest of a reservation, and recover money when another buyer purchases it.

In September 2026, Conrad and Eric Park announced a $245 million compute sale to one AI company, covering two NVIDIA Blackwell B300 clusters under separate 36-month agreements. The capacity was scheduled to come online through late 2026 and into 2027.

Conrad also writes about product design and the technical and financial constraints on AI:

• Product positioning and customer control. In his 2020 model of product positioning, developed from ideas he credits to Cheng Lou, Conrad calls the work customers must supply to get value from a product “scaffolding.” He uses a move from Heroku to AWS to illustrate why customers may accept more setup for greater control. A focused product can save work but become too restrictive as a customer's needs grow. He argues that developer products sometimes need to offer more options or require more code, and that positioning should make both the value and the required work clear.

• Revisable beliefs for language models. In 2023, he outlined a “world model” built around an external database of statements, supporting evidence, and confidence estimates. The design asks a language model how likely the evidence would be if a statement were true and how likely it would be if it were false, then uses Bayes’ theorem to update confidence. Dependencies between beliefs allow a correction to trigger revisions to conclusions built on it. Conrad also identified the limitations: models can assign unjustifiably high probabilities, and an assessment must account for the reliability of its sources.

• Pricing AI around physical costs. His 2024 analysis of commoditization argues that advice built around high-margin software often neglects equipment purchases, financing costs, depreciation, and the risk of unused capacity. When prices exceed a customer's budget, Conrad prioritizes reducing costs before polishing the product. He argues that commodity competition favors businesses with cheaper capital, while large equipment purchases and debt limit how easily a company can change direction.

• Credit and access to compute. In “The end of startups,” published in October 2026, Conrad warned that lenders’ preference for highly creditworthy customers could restrict startups’ access to compute or force much larger upfront payments. He argues for lending backed by tradable GPU-hours: if a borrower fails to pay, a lender could sell the remaining reserved capacity pledged as collateral. He sees that resale option as a way to make access depend less on a startup's balance sheet and leave small AI teams more room to compete.