Evan Conrad Warns Most AI Startups May Be Unable to Buy Compute

Evan Conrad warned in an essay published Thursday, Oct. 22, that most AI startups may be unable to buy computing capacity within six months as access to new GPU clusters hinges on credit.

Conrad described global computing capacity as sold out, saying companies seeking more supply must secure contracts for new clusters. Startups are not investment-grade entities, he said, and lenders and cloud providers are requiring larger prepayments to limit credit exposure. He cited contracts that require full payment before customers receive capacity.

The warning echoes earlier concerns about private-credit risks tied to AI infrastructure borrowing. Conrad said investment-grade GPU tenants were scarce, with leverage and resales of tokens to non-investment-grade buyers contributing to rating downgrades.

He expects investment-grade GPU buyers to require token customers to meet the same credit standard or make large deposits. At the time of writing, Conrad cited B300 capacity available for $6 an hour and inference sales at about $30 an hour. He anticipated higher GPU prices would narrow inference margins and make tokens harder for customers with weaker credit to obtain.

Conrad said a physically settled compute-forward market called SF Compute could let lenders use GPU hours as collateral instead of relying on one provider’s cluster.