SF Compute

SF Compute, also called San Francisco Compute Company, operates GPU clusters and lets customers resell unused reserved capacity. Its model combines long-term compute commitments with a market for shorter windows: customers can sell anything from a single hour to the remainder of a reservation. On September 10, 2026, the company announced $245 million in compute agreements with an unnamed AI company across two dedicated NVIDIA Blackwell B300 clusters, with capacity scheduled to come online through Q4 2026 and into 2027.[2][6]

The resale model grew out of an earlier AI lab's difficulty obtaining short-term GPU access. In its September 2026 announcement, SF Compute identified that predecessor as Junelark, a small lab working out of a Victorian in San Francisco.[6]

In an April 2025 interview, CEO Evan Conrad recalled that the company initially intended to train music and general audio models. The team wanted about a month of access to a large GPU cluster, but providers required contracts lasting a year or longer. Rather than raise a large initial round to cover training, it took a year-long commitment, intending to use one month and lease out the other eleven.[1]

That decision made selling compute an immediate survival requirement. Conrad said the company owed approximately $500,000 every 30 days and had approximately $500,000 in the bank; failing to sell out the cluster could bankrupt it. After spending its first year selling capacity, the team began arranging bespoke brokerage deals, combining customers who wanted shorter periods into a vendor's longer contract and typically operating the cluster as bare metal. The resulting book of buyers and sellers became the foundation of SF Compute's market.[1]

PYMNTS reported on July 16, 2024, relaying Bloomberg, that Evan Conrad and Alex Gajewski co-founded San Francisco Compute Co. and that it had raised $12 million, led by Alt Capital, to build a computing-power trading platform. Doubling staff to 30 was a reported plan. Data Center Dynamics reported on December 2, 2025 that the company had secured $40 million in Series A equity financing to expand its AI compute marketplace.[4][5]

By April 2025, Conrad described a market with bids, asks and a matching engine. Long-term buyers could sell capacity back, supplying shorter bursts to other users. He said customers could obtain thousands of H100s for an hour—a stated purchasing capability, rather than a fleet count. He cited Standard Intelligence, Phind and Schmidt Futures grantees as customers, and emphasized researchers who could afford a large burst of compute but could not sustain a long-term commitment.[1]

The market's short reservations allowed customers to construct their own purchasing policy. In Conrad's explanation, a customer could repeatedly reserve the next hour while setting a maximum price: purchases continued at market prices until that ceiling was exceeded. This exposed customers to changing prices and the possibility that a subsequent hour would not be purchased. As unused compute approached its delivery time, sellers could lower their asking prices because expired compute time could no longer be sold.[1]

SF Compute also operated the clusters behind the transactions. Conrad described LINPACK burn-in, performance tests approximating real workloads, passive checks during customer workloads and active checks during idle periods. Problematic hardware could be taken offline. In almost all cases, the company had BMC access for remote machine control; a UEFI-layer mechanism downloaded and booted customer images, allowing the team to standardize cluster software and automate testing. Automated refunds were only partly implemented in April 2025, and failures that checks missed could still require manual handling.[1]

Conrad explicitly called SF Compute an online spot market, not a derivatives market, in April 2025. He described an eventual goal of producing an index price that could support a cash-settled future, allowing data centers to hedge compute prices. That was a proposed extension of the business.[1]

At the time of the interview, Conrad identified Eric Park, a Voltage Park cofounder, as CTO. He credited design firm Airfoil, John Pham and COO Ethan Anderson, who had a RISD design background, for the company's visual design. Conrad said the restrained presentation was intended to set expectations around useful short-term compute and pricing rather than an exaggerated product experience.[1]

SF Compute's changelog records the Large Scale Inference (LSI) batch-inference dashboard becoming generally available on January 1, 2026, followed by B200 and B300 fleet availability on January 12. On March 2, the company sunsetted LSI batch inference, stating that it wanted to better support inference providers running on SFC. It also removed the legacy sf buy, sf sell and sf extend commands in CLI v0.30.0.[8]

On April 7, 2026, SF Compute introduced a rewritten API and Rust CLI in preview, with a model that separated compute allocation from running machines and added resale support. Deployments maintained a target number of running nodes; procurements bought and sold automatically within price bounds. The v2 API and CLI remained in public preview on May 12, when the company clarified its terminology: a node was a physical machine purchased by an order, while an instance was a VM used for workloads. June 9 migration tooling converted legacy v1 nodes to v2 instances without interrupting running VMs.[8]

A September 9, 2026 technical post described givemeanode, SF Compute's autoresearch product, as a market maker on SFC. Its capacity manager, called the treasury, bought wholesale GPU time for customer work, extended occupied capacity, sold spare windows and bid for future delivery. Capacity planning accounted for active and queued work, buffers, external supply and placement constraints. Low GPU utilization alone did not justify resale: disks and queued work could still require a node.[7]

The treasury coordinated trading with scheduling. It blocked new work before offering spare time, saved order intent before submission and reused stable idempotency keys when retrying requests. An open bid provided no capacity until it filled, and a resale listing earned nothing until it filled. Because cancellation could race with a fill, the treasury reconciled final sold intervals before reusing capacity or replacing an order. The post's prices and budgets were illustrative examples.[7]

On September 9, 2026, Conrad reported that on-platform customers received an average 25% realized discount through resale: a customer buying $100,000 of compute recovered $25,000 on average by selling unused capacity. This was a company-reported recovery of spending, rather than a reduction in posted GPU prices; the post did not specify the measurement period, sample size or weighting.[7]

The next day's announcement by Evan Conrad and Eric Park described two 36-month agreements worth $122.5 million each, totaling $245 million across two dedicated B300 clusters. The announced figure was contract value, with deployment extending into 2027. Customers could resell unused reservations; the announcement also described prospective bidding for future capacity using limit orders specifying quantity, delivery time and maximum price.[6]