Author: Mr. Lin Weizhi, Executive Vice President, Ji-Pu Industrial Trend Research Institute

When observing the evolution of AI data center networks, one thing is becoming increasingly clear: the industry is moving from the problem of "insufficient bandwidth" to the problem of "topology and physical limitations". As AI clusters move from scale-up in a single data center to scale-across data centers, the network architecture that used to rely on OEO (optical-electrical-optical) switching begins to experience structural stress. Latency is no longer determined by the number of switching nodes, but is gradually dominated by the distance and topology across data centers; bandwidth is no longer solved by just stacking more switches, as power consumption and SerDes evolution speeds have begun to lag behind AI computational power growth. As with previous research on OCS, I think it's not a coincidence that optical circuit switching (OCS) is being revisited, but rather an inevitability; the core value of OCS is that it offers a different way of thinking than packet switching: instead of letting the traffic compete for a spot in the network, it gives it a direct and definitive path. When GPU to GPU