据 Bloomberg 报道，DeepSeek 计划在内蒙古数据中心用华为昇腾 950DT 运行模型，部署规模取决于芯片供应，这些芯片目前不计划用于训练。华为据称已投入 $20B+ 囤积内存并自行完成封装，950DT 年产量估计在 200K 至 400K–500K 之间，阿里和字节跳动也在争取供应。阿里、腾讯和字节跳动还通过第三方租用 Nvidia B 系列芯片，反映出算力依然紧缺。
转述 Bloomberg 的报道呈现国产算力供应与部署选择的现状，读者可据此了解算力紧缺对模型运行的影响。
DeepSeek plans to run models on Huawei’s Ascend 950DT at its Inner Mongolia data center, Bloomberg reports. Deployment depends on supply; the chips are not currently planned for training.
Huawei reportedly spent $20B+ stockpiling memory and handles packaging itself. Annual 950DT estimates range from 200K to 400K–500K, with Alibaba and ByteDance also chasing supply.
Alibaba, Tencent and ByteDance also rent Nvidia B-series chips via third parties—another sign that compute remains scarce.
来源：@thexpin · x.com