China Mobile unveils AI clusters scale-across solution GSE-DCI and 115.2T router prototype based on packet spraying

The new solution enables high-efficiency cross-regional AI computing through packet-spraying networking and ultra-high-capacity routing

Partner Content At MWC Barcelona 2026, China Mobile launched the GSE-DCI (Global Scheduling Ethernet – Data Center Interconnect) solution, designed to overcome the long-distance networking bottlenecks in cross-AI-data-center computing and storage resource pooling.

As AI technologies rapidly evolve, the parameter scale and application scenarios of AI foundation models continue to expand. A single intelligent computing center, constrained by local computational resources, energy supply and network bandwidth, cannot efficiently support the distributed training and real-time inference tasks of ultra-large-scale models. In this case, the evolution path of computing infrastructure is extending from traditional Scale-Up and Scale-Out to Scale-Across, enabling efficient cross-regional computing power coordination and global intelligent scheduling.

GSE (Global Scheduling Ethernet), first launched by China Mobile in 2023, is a scale-out intra-data-center solution with native packet spraying capability for high-performance AI cluster networking. GSE focuses on intra-cluster switching, while GSE-DCI enables inter-data-center interconnection. Together, they form a unified end-to-end network architecture that seamlessly supports large-scale distributed AI infrastructure.

Targeting four major challenges—imbalanced load across WDM multi-wavelength links, slow response in long-distance congestion control, high overhead and latency in security, and complex coordination between computing and networking.

The GSE-DCI system integrates four key innovations:

• Multi-wavelength Wide-Area Packet Spraying (Super Pipe): Balances load across wavelengths, creating N×800G ultra-wide channels
• Long-haul Fast Congestion Notification (Long-haul CNP): Enables intermediate nodes to proactively notify sources, solving long-distance congestion perception delays
• Ethernet Physical Layer Security (PHYSec): Employs bitstream encryption with nanosecond latency, achieving “zero-overhead” data protection
• Efficient Segment Routing (G-SRv6, compliant with RFC9800 CSID): Compresses SRv6 header overhead, enhancing integrated computing and network scheduling efficiency

The newly unveiled prototype supports up to 144 long-haul 800GE ports, 115.2T switching capacity, and transmission over 100 km, achieving over 98% of single-cluster computing efficiency in distributed AI training across data centers. It effectively enables large-scale cluster collaboration and fragmented computing resource pooling for both training and inference.

Contributed by ZTE.