Huawei Unveils Liquid-Cooled AI Switch
At MWC Barcelona 2026, Huawei unveiled its Xinghe AI Fabric 2.0 solution, featuring the industry's first commercial 51.2T liquid-cooled fixed switch. The new hardware is designed to boost performance and efficiency in large-scale AI data centers.
The introduction of a 51.2T switch addresses the massive data transmission needs of AI training, where thousands of GPUs must synchronize and exchange petabytes of data. Any network latency can leave expensive GPU resources idle, severely degrading overall efficiency. High-throughput, low-latency networking is therefore as critical as raw GPU computational power for scaling AI models effectively. Huawei's CloudEngine XH9230-128DQ-LC liquid-cooled switch provides a throughput of 51.2 terabits per second, configured as 128 ports of 400GE. This level of performance is crucial for the large-scale, east-west data traffic characteristic of AI training clusters. Competitors in the 51.2T switch market include Marvell with its Teralynx 10 and Broadcom with the Tomahawk 5, both targeting the intense bandwidth and low-latency demands of AI/ML workloads. The switch's liquid cooling system offers twice the heat dissipation efficiency of typical industry solutions by providing 100% liquid cooling for the optical modules. This improved thermal management allows for greater density, enabling the deployment of up to eight switches per cabinet and doubling cabinet utilization. Efficient cooling is a critical component of data center design, often accounting for 30-40% of total energy usage. Shifting from traditional air cooling to liquid cooling can significantly improve a data center's Power Usage Effectiveness (PUE), a key metric for energy efficiency. While a typical air-cooled facility might have a PUE of 1.5-1.7, liquid-cooled centers can achieve 1.2 or lower, with some reaching below 1.05. Studies have shown that a 75% liquid-cooled environment can reduce cooling power usage by 27% and total site power by over 15%. The Xinghe AI Fabric 2.0 solution is a broader three-layer architecture encompassing "AI Brain" for automated operations, "AI Connectivity" for reliability, and "AI Network Elements" like the new switch. Key software features include the StarryWing Digital Map 2.0 for automation and Xinghuan AI Turbo 2.0, which handles load balancing to optimize network performance.
