DeepSeek's Cluster Serves 155T Tokens/Day at $38 Revenue per GPU-Hour
zephyr_z9 · x · 2026-10-21
DeepSeek disclosed that a single cluster serves 155 trillion tokens per day, generating $38 in revenue per GPU-hour at off-peak pricing. At peak pricing, revenue per hour would double.
Notably, they open-sourced not just their inference engine (DeepInfer) but also their revenue numbers. The GPU model wasn't specified, though DeepInfer supports both NVIDIA Blackwell and Huawei Ascend.
Related event: DeepSeek Open-Sources DeepInfer Inference Framework with deepseek-recipe(38 posts)→
More from Infra
- OpenAI adds 12-hour prompt caching to API, cutting costs for persistent agents — nikunjhanda · 2026-10-21
- Zyphra ships second wave of MI355X inference optimizations for MoE and long-context — teortaxesTex · 2026-10-21
- DeepSeek V4.1 Flash large-scale inference deployment write-up exposes real frontier inference economics — teortaxesTex · 2026-10-21
- Together launches post-training service with $9/GPU-hour B200 pricing, lets you own the RL loop — togethercompute · 2026-10-21
- Analyst: AI revenue will cover 85% of fleet costs by 2030, against $9.6T datacenter spend — BenBajarin · 2026-10-21
- Google and Meta launch TorchTPU, a native PyTorch backend for TPUs — PyTorch · 2026-10-21