How DeepSeek's open-sourced DeepInfer turns one rented GPU into $38/hour
cneuralnetwork · x · 2026-10-23
A developer published an interactive 3D blog breaking down DeepSeek's newly open-sourced inference engine DeepInfer:
- The economics: rent a GPU at $6.89/hour, run DeepInfer and sell tokens at DeepSeek's off-peak prices — roughly $38/hour in revenue per DeepSeek's own numbers, leaving an 82% margin on a GPU you don't own.
- Why cheap and profitable coexist: DeepSeek V4.1 Flash is 20x cheaper than frontier models like Claude Opus 5.5, driven by engine efficiency — a DeepInfer cluster serves 155 trillion tokens daily.
- How it works: one request traced end to end — text becomes tokens, prefill writes intermediate notes, decode generates the answer, and notes are persisted so similar prompts can skip prefill entirely.
- The author also showcases Flapjot, his own platform for embedding 3D diagrams in blogs.
More from Infra
- The Jensen effect: MicroDucks sells 4,000+ units in 24h after GTC showcase — andimarafioti · 2026-10-23
- Terry Tang co-founds TensorScale AI to rebuild inference infra for world models — TerryTangYuan · 2026-10-23
- Quantizing a 9B model: 70% of damage sits in the first 8 layers, and perplexity hides the tail — DinnerPrestigious953 · 2026-10-23
- HFT giants now fight for compute: Jane Street-backed Thinking Machines hits $40B valuation — dinabass · 2026-10-23
- Bloomberg: Tencent, Legend Capital and CPE circle Huawei Ascend chips, same names near DeepSeek — YvesMulkers · 2026-10-23
- How Many AI Agents Can a $6/Month 1GB VM Run? They Measured 200 at Once — swarmagent · 2026-10-23