Snowflake CEO: AI labs risk a capacity bust, not a bank run (opens the source in a new tab)
- Snowflake CEO Sridhar Ramaswamy argues in a thread on X that AI labs commit to compute for years while selling usage customers can cut tomorrow. In his hypothetical, $100 of capacity against $150 of expected revenue leaves $50; a 30% revenue miss cuts that to $5.
- He draws lessons from banks, commodity traders, airlines, utilities and cloud giants, and warns that clouds which invest in labs and also sell them compute can make one promise look like several revenue streams.
- Citing Anthropic's S-1, he says compute cost fell to 55% of first-half sales of $16.3bn, from 201% a year earlier.
- $50 to $5
- spread after a 30% revenue miss
- 55%
- Anthropic compute cost to sales, H1, as cited
- $265.6bn
- future purchase commitments, as cited
Why it matters As AI firms head for public markets, investors must judge whether long compute contracts are a moat or a trap; his three gauges, paid GPU Graphics processing unit: the chip, mostly from Nvidia, that trains and runs AI models. use, unit prices and firm obligations, make a practical checklist.
Other side He is not a neutral observer: Snowflake sells data and AI services. And contract-backed cash flows that lenders treat as safe are only as strong as the labs signing them.