The number comes from DeepSeek's own published deployment Q&A and technical report, but it's worth separating what's reported from what's independently verified.
How it's measured (as disclosed by DeepSeek): It's the aggregate service throughput of a single production cluster running DeepSeek-V4.1-Flash. The figure is 155 trillion tokens/day . Context matters: DeepSeek reports an average input length of 77K tokens per request, and a 96% input cache hit rate—meaning most of that "throughput" is cache-served rather than freshly computed . Per-GPU numbers are also disclosed: ~60K tokens/s without cache hits, ~1.5M tokens/s with cache hits .
The honest caveat: This is a self-reported production statistic, not a reproducible benchmark from the open repo. Independent analysis notes the 155T figure appears in DeepSeek's Zhihu post and deployment Q&A, but is not in the GitHub repository itself, and the KV-cache management system that enables the 96% hit rate is not part of the open-source release . So you can't fully reproduce the number from the code alone—you'd need DeepSeek's internal 3FS storage cluster and cache infrastructure .
Bottom line: It's not exaggerated marketing in the sense of being invented—it's a disclosed production metric tied to specific conditions (77K avg input, 96% cache hit, off-peak pricing for the $38/card-hr revenue figure). But it's also not a general-purpose claim you can independently verify from the open repo. A fair reply would acknowledge both: the number is real as a self-reported production stat under specific conditions, and the community's job is to see what smaller teams can achieve without that cache tier.
