DeepSeek Says One DeepInfer Cluster Serves 155 Trillion Tokens a Day
@swyx surfaces DeepSeek’s Zhihu figures, including $38 per card per hour in revenue at off-peak prices.
The numbers behind DeepInfer were on Zhihu.
DeepSeek announced DeepInfer, its in-house inference framework, on X as fully open source. @swyx says the company “posted way more alpha on their zhihu page,” and that readers who only follow X get the “laowai approved stuff.”
He quotes DeepSeek’s Zhihu post: “Based on this framework, we have implemented a high-throughput, low-latency large-scale inference service for DeepSeek V4.1 Flash.”
| Measure | Figure |
|---|---|
| Throughput, single cluster | 155 trillion tokens/day |
| Overall revenue | $38 per card per hour |
| Price basis | Off-peak prices |
@swyx does the margin math.
@swyx puts the cost at roughly $5-10 per GPU hour. By his reading, DeepSeek is claiming an 80% serving margin at its own prices.
This morning’s edition covered the DeepInfer release itself: the engine behind DeepSeek’s production services, which for now runs on NVIDIA SM100 and Huawei Ascend 950.
Sources
- @swyx on X · Oct 20, 9:51 AM PT