SF Compute's Evan Conrad: credit, not chips, is AI's bottleneck — the only physically settled compute market
Key points
Key takeaways from Evan Conrad, co-founder and CEO of San Francisco Compute, on The Upstarts Podcast (October 2026).
SF Compute describes itself as the only physically settled compute market. It does not own the hardware; it operates GPU clusters for owners as a "property manager for supercomputers" — doing all the functions of a cloud while not being a cloud — so buyers know exactly what they are getting and that it will be deliverable.
Credit, not land or power, is AI's true bottleneck. AI labs cannot borrow against GPU clusters, so they must raise at ever-higher valuations to pay massive deposits upfront; a financially settled market lets lenders underwrite the asset's future value rather than the lab's credit risk.
Founded out of necessity. SF Compute began as the AI lab Junark, which faced a roughly $400,000 monthly compute bill against about $500,000 in the bank; subleasing its own over-bought cluster revealed how broken GPU markets were, pushing the team down the stack into operating clusters.
$245m contract is "a drop in the bucket". Even a mid-sized deal in compute is bigger than many venture funds' entire assets under management, and SF Compute almost did not announce the contract because it was unremarkable relative to the market.
Arming the rebels is deliberate. Conrad wants compute to stay liquid so new labs can start, rather than letting capacity be locked up by a few frontier buyers; without that, he argues, a future startup like Junark would need a billion-dollar valuation out of the gate just to secure a deposit.
Why incumbents won't do this. A GPU cloud maximises profit by selling all capacity to the largest labs on long contracts, so making a fair market is a "fool's errand" for them — SF Compute built the capacity to become a cloud and then chose not to, which is why Conrad says anyone competing with him is "basically stupid".
Who works with SF Compute. Customers include venture firms and nonprofits that want reserved capacity for their portfolio or grantees, and investors seeking pure financial return on GPU assets; the company's CTO is a co-founder of the multi-billion-dollar GPU cloud Voltage Park.
Real-world chaos. Conrad recounts clusters installed in the wrong data centre, broken InfiniBand fabrics, data centres that were "literally chicken coops" and a site visit hit by a monsoon — the operating failure modes that make physical settlement valuable.
Read more: The Upstarts Podcast · YouTube
