SF Compute builds physical settlement layer for gpu trading

SF Compute, formerly known as Junark, has pivoted from subleasing clusters to operating the full cloud stack for the physical settlement of compute markets. The company states that 5% of its total capacity is currently on its order book, with high liquidity relative to volume. A key metric cited by the firm is that a customer spending $100 to buy capacity makes back an average of $25 by selling it back. SF Compute co-designed its order book to handle physical constraints such as preemption, storage recovery, and WAN transit paths between clusters. It supports contracts ranging from real-time to three to five years out, noting that clusters are typically financed over that period.

The company does not own hardware to maintain independence and prevent market manipulation, describing itself as 'Switzerland' in the ecosystem. It partners with data center owners to provide 'supercomputing as a service,' handling cluster architecture, backplane management, and supply chain consulting. SF Compute has closed several hundred million dollars in volume and recently enabled customers to subdivide InfiniBand fabrics on the fly. The firm argues that physical settlement is a prerequisite for standardized contracts, noting that early H100s and GB200s had firmware bugs and defects that required cohorting.

SF Compute observes a 'hump' in the compute forwards curve, where prices drop for capacity approaching immediate delivery because unsold inventory must clear, then rise to a peak. The company notes a significant arbitrage opportunity exists between the cost of buying V300s at $5 to $6 an hour and the potential revenue of running inference on B300s, which can generate approximately $40 an hour at 100% utilization. SF Compute posits that Nvidia, despite being the dominant player in AI hardware, cannot create an independent compute exchange because it would lack the necessary independence for traders to rely on it. The company argues that if Nvidia controlled the exchange, it would not be independent enough for people to trade on, which would reduce liquidity for lenders financing against the forwards market. This structural limitation also applies to hyperscalers and AI-native cloud providers like Crusoe, CoreWeave, Lambda, Base10, Fireworks, and Modal, as they have too much demand-side interest and would need to 'kill their business' to remain neutral. SF Compute warns that building financialization before physical settlement is standardized creates a market prone to manipulation and basis risk, as current indexes do not accurately reflect the cost of new capacity.