Scale-Across Network

Overview

The scale-across network is a new network tier beyond traditional scale-out (intra-datacenter cluster fabric) that interconnects AI GPU clusters spanning multiple data centers. It is driven by the power constraint on AI training: training GPT-5 requires 50,000–100,000 GPUs, which already exceeds what any single data center can power, forcing clusters to span multiple facilities. Connecting these distributed clusters requires a WAN-capable, high-bandwidth fabric — the scale-across tier.

Cisco's analysis at OCP 2025 established the scale-across requirements and contrasted them with existing DCI:

Modular chassis systems cannot meet the scale-across bandwidth density requirement. The recommended architecture is fixed pizza-box switches in a leaf-spine topology, using routing-capable ASICs (like the Silicon One P200) that combine router-level features (e.g., deep buffers, full routing protocol stack, security) with switch-level power efficiency.