Evan Conrad

Evan Conrad is a cofounder of San Francisco Compute, where his work centers on giving AI researchers and companies access to large GPU clusters without requiring them to use every hour of a long reservation. Latent Space introduced him as CEO in April 2025. In September 2026, he and Eric Park coauthored an explanation of the company's model: operate the clusters customers lease and let customers resell unused capacity.[3][1][6]

Conrad recalled that SF Compute began with plans to train music and audio models. The team wanted a month of compute rather than a large fundraising round and a long hardware commitment, but providers required contracts lasting a year or more. Its workaround was to buy a year, use a month and sublease the remaining time. In his April 2025 account, the company owed roughly $500,000 every month and had roughly that amount in the bank: failing to sell its capacity threatened bankruptcy. Conrad and Park's September 2026 retrospective named the earlier lab Junelark.[1][6]

Finding sublessees became a brokerage business. Conrad described combining customers with different reservation lengths into deals that satisfied a vendor's longer commitment, then operating the cluster for those customers. The accumulated buyers and sellers became the basis of a market with bids, asks and a trading engine. Resale supplied the flexibility the team had originally needed: a customer could buy a long reservation and sell portions back, creating shorter bursts for someone else.[1]

Conrad singled out graduate students and startups that could afford a large burst of compute but could not sustain a large cluster over time as customers he especially valued. He saw their position as resembling his team's original predicament. In April 2025, he also explained how buyers could automate successive hourly reservations with a maximum price: they would buy at the market price while it stayed below their limit, but could lose the next reservation if prices rose above it.[1]

In “On Commoditization, Pricing, and Assets”, published March 1, 2024, Conrad framed AI as “software with expenses.” He argued that commodity competition ultimately favors suppliers with cheaper capital, while expensive AI inputs make price more consequential than the product differentiation familiar to software founders. He also argued that asset purchases and loans constrain a company's strategic options: a founder can change software direction more readily than escape the obligations attached to physical assets.[3]

In April 2025, Conrad applied that argument to GPU clouds. In his view, model builders often seek the most compute they can obtain for a fixed budget, making them unusually sensitive to price. He warned against assuming that adding software to expensive GPU inventory would produce traditional cloud margins. His proposed separation was between financing hardware and building services that buy compute from a market. He distinguished this from the technical operation required to deliver usable clusters.[1]

Conrad described how SF Compute's engineering team delivered usable clusters through LINPACK burn-in, performance tests resembling real workloads, and active and passive checks for failures. Remote management through BMC access let its engineers reset and troubleshoot machines; a UEFI shim downloaded and booted customer images, helping standardize entire clusters. Automated refunds were only partly working in April 2025, and failures missed by the checks still required manual handling.[1]

Conrad explicitly called SF Compute an online spot market in April 2025. Cash-settled futures were a possible later goal: he argued that a spot-price index could support instruments allowing data centers to lock in prices and reduce financing risk. In their September 10, 2026 announcement, Conrad and Park emphasized the combination of technical control and resale. Their argument was that lenders need long customer commitments, while customers need a way to recover money when workloads change.[1][6]

Conrad also helped shape SF Compute's early design. He described its restrained presentation as a deliberate response to exaggerated product expectations and the pressure of the company's beginnings: he wanted customers to expect good prices for short bursts. He credited Airfoil, John Pham and COO Ethan Anderson alongside his own design involvement.[1]

Before the compute business, Conrad explored several software products. In his April 2025 recollection, an earlier startup began as the mental-health app Quirk, then became Room Service, a distributed-systems company closer to his technical background. Later experiments included an AI email product. He attributed leaving the email effort to burnout after years of trying products, rather than concluding that the idea itself was bad.[1]

His earlier essays show interests beyond infrastructure. In September 2022's “The end of software”, he predicted generative interfaces that would create and adapt software for individual users. He anticipated smaller software teams and possible opportunities for models and runtimes that could safely execute generated code. These were forecasts about computing's direction.[5]

In February 2023's “How to make LLMs say true things”, Conrad proposed a World Model: an embeddings database of declarative beliefs with confidence values updated through Bayes' theorem. Language-model outputs would supply priors and evidence likelihoods; dependencies between beliefs would allow recomputation when an underlying assumption changed. His illustrative spicy-tuna example moved confidence that someone liked spicy food from 7% to about 30%. He warned that language models supplied overconfident likelihoods and that source trust required context; the essay presented a method and its limitations, without an empirical reliability evaluation.[4]