Inference is a beautiful business. deepseek reports 155T tokens/day from one inference cluster at $38/gpu-hr. I estimate their cluster at ~5 MW, implying ~$160M/MW/year in API revenue and 84–92% margins after $3–6/GPU-hour rent working assumptions: 96% input-cache hits, ~107 input token/output token ratio -> ~$14.35 revenue per billion tokens at off-peak prices. working backwards from 155T tok/day, 60,000 uncached input tok/sec/gpu prefill + 14,000 output tok/sec/gpu decode, each two-GPU pair handles approximately 1.5m input + 14k output tokens/sec -> 2400 gpus. on blackwell hardware I assume 2 kW/gpu including external networking, cache storage, etc. No wonder everyone wants be part of the token flow / bid up GPUs :)