Alibaba launches industrial-scale training infrastructure with 64-card SuperNode

Alibaba has unveiled the Zhenwu M890 SuperNode GP9A, a 64-GPU training node that ships in a rack ready to operate within an hour, deployed at its Ulanqab data center in Inner Mongolia. The system targets inference on models such as Qwen-3.8 Max and Kimi K3, and the company says it can scale to clusters of up to 122 thousand GPUs, signaling a major on-site capacity expansion.

Compared with the previous-generation 810E, the M890P jumps on every core metric: substantially more HBM per GPU, higher compute throughput, and 64 GPUs per SuperNode versus 16. It adds MXFP4 and FP16 precision support, while sixteen 64-core CPUs manage each node. Scale-up bandwidth leaps to 51.2 TB/s from 11.2 TB/s, driven by a shift from 400G to 800G interconnects.

Ulanqab, one of Alibaba Cloud's eight largest Chinese data centers, draws on especially cheap green power at 0.32 yuan per kWh, against 0.6–0.9 yuan (roughly 0.31–0.47 shekels) in southern and eastern regions. At a one-gigawatt facility, that spread translates to roughly 5 billion yuan in annual electricity savings alone.

The launch accompanies a plan to sharply raise output of Alibaba's proprietary chips. The ability to deliver 64-GPU racks with a one-hour lead time indicates a supply chain and inventory already geared for mass rollout, not an experimental prototype.