Huawei and Alibaba have unveiled new hardware solutions for AI, focusing on proprietary interconnects and cluster architectures to bypass limitations on Western technologies. Huawei introduced the Atlas 960E SuperPoD, which utilizes near-packaged optics (NPO) and the Hi-ONE optical module to connect thousands of NPU accelerators.
The Atlas 960E system supports up to 4096 NPUs per block, with performance reaching 8 EFLOPS for FP8 and 16 EFLOPS for FP4. Huawei also announced the open-sourcing of its CANN software stack and introduced the OceanStor M900 storage system to manage context memory for AI agents.
Alibaba’s T-Head division presented the Zhenwu V900 AI chip, which the company claims is three times faster than its predecessor, the M890. The V900 supports MXFP8 and MXFP4 formats and can scale up to 500,000 accelerators via ICN switches for models with 5–10 trillion parameters.
Alibaba plans to launch the Panjiu super-accelerator in Q1 2027, integrating the V900 chip with SmartNICs and ICN switches. The company is also transitioning its Yitian CPU line to the RISC-V architecture, aiming for a fully proprietary stack that includes accelerators, CPUs, and networking hardware.
Other Chinese market players are exploring alternative approaches. Moore Threads offers the MTT S5000 GPU with CUDA-compatible aspirations, while Dongfang Suanxin is developing the DF1000, which uses 3D-stacking technology instead of HBM memory to achieve high performance in AI tasks.
