DeepSeek reports 155T tokens/day from one cluster with 96% input cache hits

DeepSeek reports that one of its GPU clusters processes 155 trillion tokens per day, at $38 revenue per GPU-hour, according to a Latent Space post quoted by X user swyx.

The figures are cited in a post about GPU value and stock market effects, and the report does not include the 96% input cache hit rate named in the headline or any further detail on how the figures were measured.

swyx's own post, a joke that every GPU sold today gained about 30% in lifetime value and would lift the stock market about $600B, relies on those cluster figures only as background.