Ep. 034 - Engrams: How DeepSeek Offloads KV Cache to DRAM and SSD (Core Research) | Jordan Nanos, Cam Quilici, Alec Ibarra, Bryan Shan

…right? Meaning, yeah, obviously, you don't have users fully saturating the system at all times and maybe you have like some nodes down or whatever. [about serving Kimi-K3 on a router] So, at 60% utilization, with 30% license fee that you have to give to moonshot, you can literally make a almost 50% profit margin serving this on GB300. It's just a money printer, right? So this is assuming like the cash hit rate that we get on Agent X and then looking at open router data,…

Semianalysis does paint a rosy picture of serving Kimi. They do preface this with "Until prices return to sane values" (or something like that, not verbatim)