SF Compute

SF Compute is building the missing market structure for the largest new commodity of the decade. GPU procurement today is where energy was before liquid spot markets: dominated by rigid one-to-three-year bilateral contracts that founders sign under scarcity panic and regret within months — a pain Evan Conrad and Alex Gajewski lived personally before founding the company in 2023. Their answer, a marketplace where H100/H200/B300 capacity clears at market prices and any buyer can resell unused commitment back into the book, converts compute from a liability into a tradable asset. The model is beautifully asset-light: SF Compute owns no GPUs yet manages over $100M of hardware, earning exchange economics on flow rather than carrying depreciation. A $40M Series A at $300M (DCVC, Wing, late 2025) is a restrained mark by AI-infra standards, and the team is scaling deliberately — 54 people as of May 2026, with ex-Voltage Park CEO Eric Park hired as CTO to harden the operational layer. If GPU compute follows every prior commodity toward standardized contracts, spot clearing, and eventually forwards, the venue that established liquidity first becomes the NYMEX of compute — a fee-taking monopoly position worth far more than any fleet.

Marketplaces for commodities only capture value when they own liquidity, and SF Compute's liquidity is borrowed from a scarcity regime that may not persist. The resale proposition — exit your contract by selling capacity back — has value precisely when GPUs are hard to get; in a supply glut, resale prices collapse below contract cost, sellers eat losses, and the marketplace's core promise becomes a mechanism for crystallizing them. Meanwhile every competitor is attacking the same lock-in pain directly: hyperscalers and neoclouds now offer short-term and spot GPU terms, Together and Lambda sell flexible clusters, and aggregators like Shadeform and Prime Intellect broker across providers — flexibility is becoming a feature everywhere rather than a venue. The exchange model also carries hidden balance-sheet risk: SF Compute arranges long-term contracts financed by outside investors, which means counterparty and credit exposure if AI startup demand churns. Against CoreWeave-scale players, a $52M-funded company has no pricing power over supply, and data-center operators can disintermediate the market by listing directly on multiple venues. Financial-market analogies flatter the story, but compute is heterogeneous (interconnect, location, reliability), which resists the standardization that real commodity exchanges require. The likeliest outcome is a useful niche broker, not a monopoly venue.