沐曦将在上海临港建设万卡GPU集群，采用曦云C550或C600
英文摘要
MetaX is advancing a 10,000-GPU cluster project at Shanghai Lingang, with the final choice between the existing C550 and the newer C600 GPU. The company has already delivered a full-scale domestic AI cluster and reported ¥743M in 2024 revenue, with over 25,000 C500 chips sold to major clients like China Telecom, Alibaba, and Tencent. A separate Wuxi-based 10,000-GPU cluster using C550 cards is now operational. The industry is shifting from training-centric deployments to inference, as China's inference demand is roughly 8× training demand. MetaX is also preparing for an H-share listing in Hong Kong, while supply chain and 7nm production capacity remain key constraints.
中文摘要
沐曦正在上海临港推进一个万卡GPU集群项目，最终将选择曦云C550或新一代C600方案。该公司此前已交付超万卡规模国产智算集群，2024年营收7.43亿元，C500芯片累计销量超2.5万颗，获中国电信、阿里、腾讯等头部企业批量采购。无锡另一采用C550的万卡集群已投入实际运营。行业重心正从训练转向推理，国内推理需求约为训练的8倍。沐曦日前宣布拟赴港上市，但供应链尤其是7纳米产能仍是主要挑战。
关键要点
MetaX to build a 10,000-GPU cluster in Shanghai Lingang, finalizing between C550 and C600 GPU models.
沐曦将在上海临港建设万卡GPU集群，最终方案将定于C550或C600。
A Wuxi-based 10,000-GPU cluster using MetaX C550 cards has already been lit up and is in live operation.
无锡沐曦国产GPU万卡集群采用C550设备已完成一期点亮，投入实际运营。
MetaX's 2024 revenue reached ¥743M, with over 25,000 C500 chips shipped to top customers, and it is pursuing an H-share IPO.
沐曦2024年营收7.43亿元，C500累计销量超2.5万颗，并已启动赴港H股上市。
The AI infrastructure focus is shifting from training to inference, with domestic inference demand roughly 8× higher, driving a need for more distributed deployments.
AI算力重心正从训练转向推理，国内推理需求约为训练需求的8倍，促使用户结构从集中式集群向分布式部署调整。
Key constraints include 7nm production capacity and reliance on traditional server partners, with token-level cost efficiency becoming the new competitive metric.
供应链尤其是7纳米产能受限，沐曦仍依赖浪潮、联想等传统服务器厂商，单位Token成本正成为新的竞争力衡量标准。