Listen: Latent Space episode and transcript. Watch: publisher-linked YouTube video. Published 22 October 2026; 86 minutes.
Who: Erik Bernhardsson, introduced as Modal’s CEO. The conversation discusses elastic GPU capacity, startup latency, workload variation and the financial exposure behind fixed compute commitments.
Useful takeaway: a monthly average workload can hide expensive peaks and quiet periods. The supplier’s ability to pool demand is part of its business model; it does not remove the need to check the buyer’s own bill, service limits and failure behaviour.
Buyer exercise — proposed: plot your demand by hour for a representative period. Compare the cost of peak-sized dedicated capacity with the actual variable service quote, including scale-up delays and required redundancy. Use observed traffic, not a smooth growth line invented to make utilisation attractive.
Question to take into procurement: “Who bears the cost and risk of idle capacity, bursts and falling compute prices under this contract?”