DeepSeek Fully Open-Sources Inference Framework: NVIDIA & Huawei Ascend Compatible, 155T Tokens Daily Single-Cluster Throughput, $38 Per Card Per Hour Revenue - 36Kr

DeepSeek's full open-sourcing of its inference stack is a structural decoupling of the software layer from silicon. By natively supporting both NVIDIA and Huawei Ascend, it simultaneously neutralizes CUDA's decade-long moat and patches Ascend's chronic software gap in a single move. The 155-trillion-token daily single-cluster throughput and the $38/card/hour revenue metric are not benchmarks—they are a pricing anchor. Once inference unit economics become publicly quantifiable, the margin cushion cloud providers and chip vendors have enjoyed since 2023 compresses systematically, echoing how open-source PostgreSQL commoditized the database query layer in the 2000s. For Huawei, this is the single most consequential ecosystem validation since the 2019 export restrictions. Ascend's silicon specs have been competitive for years, yet the software 'last mile' remained the decisive no in enterprise procurement. A cross-architecture open-source inference stack rewrites that default assumption at the procurement level. NVIDIA's likely counter—accelerating NIM and TensorRT-LLM into a closed ecosystem—faces a structural problem: open-source frameworks do not need to convince, they only need to be compatible. Over the next 18 months, the inference scheduling layer will replay the browser wars of 2000–2005. Whoever owns the framework layer owns the compute pricing power.