ChinaDeepSeek says one cluster serves 155 trillion tokens a day.
Revenue works out to about $38 per card per hour at off-peak prices, DeepSeek said.
It open-sourced DeepInfer, the serving framework behind V4.1 Flash, which runs on Nvidia and Huawei Ascend chips.
DeepSeek did not release its cache-management and load-balancing code, which is tied to internal infrastructure.
DeepSeek says inference code for its new models will keep landing in DeepInfer.
How each outlet framed it
- Zhidx leans favorable
- converts DeepSeek's $38/card-hour inference revenue into a per-card monthly yuan figure, emphasizing serving economics over the release
Sources: Zhidx