[AINews] DeepSeek DeepInfer: Open Source SOTA Inference running 155T toks/day/cluster making $22m/day at ~80% margin for a 81 day payback
DeepSeek open-sources DeepInfer inference stack running 155T tokens/day at ~80% margin
DeepSeek timed the release of DeepInfer, their production inference stack, to coincide with day one of PyTorch Conf 2026, open-sourcing a heavily customized vLLM implementation capable of up to 72-hour prompt caching—six times longer than OpenAI's newly announced 12-hour caching. The release reveals DeepSeek's economic scale: 155T tokens/day/cluster generating an estimated $22M/day at roughly 80% margin with an 81-day hardware payback. Notably, their crown jewels—ultra-optimized KV cache, PD disaggregation, and load balancing—were withheld, making this a powerful but incomplete open-source signal.