Current outcome
DeepSeek open-sourced its DeepInfer inference framework and disclosed that its cluster handles about 155 trillion tokens per day, equivalent to roughly $38 in theoretical billable revenue per GPU-hour, though this figure does not account for utilization, paid-request share, or hardware and power costs.
Progress timeline
2 material updates- #01
DeepSeek开源内部推理引擎DeepInfer：支持昇腾950，但还不是完整的在线服务
DeepSeek open-sourced DeepInfer, its Apache 2.0 inference framework used for DeepSeek V4.1 Flash, supporting Nvidia SM100 and Huawei Ascend 950 but currently limited to offline inference with key serving components still missing.
Source evidence: theblockbeats · ggcr
- #02
DeepSeek披露单GPU每小时可创造38美元计费收入，AI推理盈利能力引发关注
DeepSeek open-sourced its DeepInfer inference framework and disclosed that its cluster handles about 155 trillion tokens per day, equivalent to roughly $38 in theoretical billable revenue per GPU-hour, though this figure does not account for utilization, paid-request share, or hardware and power costs.
Source evidence: theblockbeats