San Francisco Compute has sold a leading AI company $245 million worth of compute across two dedicated NVIDIA Blackwell B300 clusters. The sale is covered by two 36-month agreements, each worth $122.5 million.
Many years ago, San Francisco Compute was Junelark, a small AI lab working out of a Victorian in SF. As Junelark, we were stuck in long-term contracts that forced us to sublease or go under.
That experience forced us to invent the concept of a compute market. A compute market and a neocloud cannot be independent concepts. Reducing the risk of AI requires the technical expertise to handle physical settlement. Customers must contract with technical experts who hold the keys to the cluster.
That's why SFC operates the clusters it leases to customers, like a neocloud, and reduces its customers' risk through resale, like a market.
No other supercomputing company is so doggedly dedicated to the single mission of reducing the risk of AI. We believe the world is going through a profound shift and the only path forward is by tackling the risks head-on. The problems of technical competence, security, credit, construction delays, and utilization are tightly linked. If you do not solve them all, these projects cannot be underwritten. That is why the San Francisco Compute Company vertically integrated to build the world's only physically settled compute market.
Long-term contracts are required to get supercomputers financed. But resale gives customers a way to recover money when they don't need all the compute they've committed to.
New compute clusters need more than day-to-day demand to get built. Lenders need firm commitments from customers who will pay for capacity over several years.
But no one knows exactly how much compute they'll need two years from now. A training run gets canceled. A model ships early. Traffic changes. Without a way to resell, the customer keeps paying for idle capacity.
SFC builds supercomputers customers can sublease. They can commit for the long term and resell what they don't use, from a single hour to the rest of their reservation. That helps cover the bill, improves their margins, and reduces the cost of gaps between workloads. Combining their own workloads with resale can bring reserved capacity close to 100% utilization.
On SFC today, customers recover an average of 25% of their compute spend by reselling unused capacity. A customer spending $100,000 gets $25,000 back on average. Here's how that works.
Other customers will be able to bid for future capacity by choosing how much compute they need, when they need it, and the most they want to pay. These bids are called limit orders.
That could mean every evening, Saturday mornings, weekday business hours, full weekends, a two-hour daily burst, or a three-day training run. When a matching offer fills the order, those hours are reserved.
Capacity is coming online through Q4 and into next year. If you're planning inference or training for 2027, or want first access to subleased B300 capacity, talk to our sales team.
