DeepSeek Open-Sources DeepInfer: 155T Tokens/Day and Up to 93% Inference Gross Margin
teortaxesTex · x · 2026-10-21
DeepSeek released DeepInfer, a fully open-source large-scale inference framework powering DeepSeek V4.1 Flash, and unusually disclosed production stats that triggered community margin analysis.
- Tech: high-performance kernels, deep reasoning scheduling, speculative decoding, SWA Bounded Replay, structured output; runs on NVIDIA and Huawei Ascend, with production-grade prefill/decode components (deepseek-recipe).
- Numbers: 155 trillion tokens/day per cluster, 23ms latency, 71.3% cache hit rate; 1M context at 200+ TPS with 72-hour cache retention.
- Pricing: $0.003 / $0.15 / $0.6 per 1M cache-hit / cache-miss / output tokens (off-peak).
- Third-party math: $78.2/hr revenue per GPU pair matches the stated $38/GPU-h; gross margins estimated at 93.4% on owned GB200 ($2.50/hr) down to 72% on CoreWeave rental; breakeven is $38/GPU-h. Zephyr estimates cluster revenue near $2.2M/day and $200B/GW-yr inference revenue, arguing DeepSeek's pricing aims to commoditize inference while staying highly profitable.
Related event: DeepSeek Open-Sources DeepInfer Inference Framework and Token Codec(47 posts)→
More from Infra
- DeepSeek open-sources DeepInfer: 155T tokens/day, $38/GPU-hour, >80% margin — Latent Space · 2026-10-21
- Cerebras WSE-3 Turbo packs 4T transistors and 250 PFLOPS into a wafer-scale chip — GregoryDiamos · 2026-10-21
- Uncensored DeepSeek V4.1 Flash GGUFs drop refusal rate from 91.4% to 0% with minimal MMLU loss — kalyan_kpl · 2026-10-21
- First commercial 1024-card Huawei Ascend 950 SuperPoD delivered to StepFun — teortaxesTex · 2026-10-21
- Specter speeds single-GPU MoE inference with quantized multi-step speculative offloading — Wenfeng Wang · 2026-10-21
- SpaceX reportedly targeting 13 GWh of compute by EOY 2027; Grok bots for businesses in weeks — elonmusk · 2026-10-21