DeepSeek's production data: $38 revenue per card-hour, 96% cache hit rate check out to within 3%
teortaxesTex · x · 2026-10-20
- Headline numbers: DeepSeek's actual online operating data reports 14,000 output tokens/s per decode card (230 tokens/s per user), roughly 1:1 prefill:decode card allocation, and $38 revenue per average card-hour at off-peak prices.
- Key metrics: 60,000 fresh input tokens/s per prefill card; 1.5M tokens/s including cache hits; 96% input-cache hit rate; 77,000 average input length; 400ms mean / 1,000ms P99 time to first token.
- Verification: Reconstructing revenue from official V4.1 Flash off-peak prices ($0.15/M uncached input, $0.003/M cached, $0.60/M output) yields $47.95/card-hour for prefill, $30.24 for decode, $39.10 averaged — within 3% of the disclosed $38, confirming the numbers are self-consistent.
- Community take: teortaxesTex argues all public baselines are garbage next to these internal stacks, inference is a money printer, and DeepInfer is being compared to vLLM/SGLang via SemiAnalysis InferenceMAX.
More from Infra
- Together, Fireworks, Prime Intellect and Modal are all converging on inference + RL training — NielsRogge · 2026-10-20
- NVIDIA GTC kicks off in Berlin with Jensen Huang keynote set for tomorrow — NVIDIAAI · 2026-10-20
- Hyperdrives CEO interview: cheaper power-dense electric motors via cooling breakthrough — MarvinTBaumann · 2026-10-20
- MiniMax H3 video generation runs on a single RTX 3060 12GB via GGUF quantization — Ok-Inevitable8391 · 2026-10-20
- ByteDance, Alibaba, Tencent place over $35B in Nvidia chip orders amid supply crunch — pstAsiatech · 2026-10-20
- DeepSeek-V4.1-Flash tops Agent Arena open models, ranks #11 overall at $0.20/task — vista8 · 2026-10-20