On 20 October, DeepSeek open-sourced DeepInfer, its self-developed inference framework, and a detail buried inside the release is reshaping how China’s AI stack talks to silicon. Both Huawei’s Ascend 950 NPU and Nvidia’s SM100 Blackwell GPU appear on the officially supported hardware list, pulled from the same code tree, and switching between the two takes a single environment variable at inference time.
The timing was charged. One day earlier, on 19 October, the US Department of Homeland Security published an order requiring federal agencies to stop procuring DeepSeek products and services, a move that instantly intensified the debate over what the framework actually signals.
For most of the past two years, domestic chips inside Chinese models sat in a subordinate role, bolted on as a late-stage backend after the real work was done on Nvidia. DeepInfer changes the premise: Huawei and Nvidia are now equal build targets inside one software framework, and the migration cost between them is, in theory, flattened. Founders in developer groups say a new project no longer assumes Nvidia by default and then considers alternatives. It now picks hardware from a single framework based on supply, cost, compliance or geography.
The shift bites hardest in inference rather than training. CloudMinds chief executive Chen Ning has argued that training’s dependence on CUDA and inference’s dependence on it are not the same order of magnitude. Nvidia is undisputed master of the training era, where changing platforms means rebuilding from scratch, but in inference the game is cost, efficiency and system capability, and everyone starts on a new line. That opens room for a fresh software stack and ecosystem.
Industry voices agree that training-chip moats are extremely high while inference thresholds are far lower, and China’s dense deployment scenarios give domestic chips a real runway. Reuters reported on 20 October that Chinese internet giants are bulk-buying Nvidia cards, and Nvidia has struggled to fill the surge of orders even though those cards are not datacentre-specialised parts.
The caveats are real. DeepInfer’s public benchmarks so far run mostly on Nvidia GPUs, so there is no clean read yet on whether Ascend 950 matches Nvidia on throughput, latency and energy at the same model and load. DeepSeek stresses that what it opened is an offline inference framework, and production pieces such as KV-cache management and load balancing are not all public. Nvidia itself open-sourced its Dynamo inference framework in March 2025 and iterated it to production grade this March.
The deeper story is in training. The Information reports DeepSeek is co-designing a next-generation AI chip with Huawei: Huawei engineers are embedded in DeepSeek’s research effort, with priority supply of the next training chip planned so DeepSeek can train advanced models on domestic silicon and cut Nvidia dependence. From price cuts enabled by the domestic ecosystem, to equal inference compatibility, to a direct attempt to lower training reliance, the pace of model-and-chip coordination looks set to accelerate.
Editor’s note: This is an adapted translation of the original Sohu report. It has been trimmed and restructured for readability for an international business audience. The full original (in Chinese) is at https://www.sohu.com/a/1090643558_116132.