saved
The end of startups
Hraness wrote this summary from a saved copy of the source. Quotations are taken word for word from the source.
gist
San Francisco Compute's Evan Conrad argues that within six months most AI startups will be unable to buy compute, because new capacity depends on credit. Lenders and clouds want investment-grade tenants, so startups face upfront payments that have reached 100% of a contract. He says a B300 costs about $6 an hour and earns about $30 an hour selling inference, a margin he expects rising GPU prices to erode. His fix, which his company sells, is a physically settled forwards market that lets lenders hold the GPU hour, not the cluster, as collateral.
ideas
- Compute has become a credit product. The world's compute is sold out, capacity comes from new clusters, and those are increasingly hard to get online for buyers who are not investment grade.
- Investment-grade tenants are scarce and slipping. Conrad says the few that remain are being downgraded as they borrow more and resell tokens to buyers who are not investment grade themselves.
- Inference margins prop up credit for now. High inference margins can make providers look creditworthy, but Conrad expects others to arbitrage the gap away and GPU prices to rise.
- Credit compounds into product advantage. A Fortune 500 company that can deploy $1 billion in tokens in a few days beats a startup that must raise $1 billion before its product exists.
- The GPU hour should be the collateral. A liquid book of GPU hours avoids a non-fungible cluster and a servicer that might fail when lenders need it; the post closes with a hiring pitch for San Francisco Compute.
quotes
“Within the next 6 months, most AI startups will not be able to buy compute.”
“This is the credit bottleneck.”
“We’ve seen it go as high as 100% of the contract paid up front”