.svg)
I’ve been asked several times if we will need to re-architect networks for AI. Recently I came across a couple of articles from SDXCentral and Cisco (see below) on this topic and it prompted me to share my thoughts with a wider audience. Generally, to understand the impact that any workload has on the network, a good question to ask is “where is the data coming from and where is it being processed”, so I have considered the question on the LAN for Chatbots, Physical AI, and Agentic AI, as well as the data centre for all three.
AI in the Data Centre: For data centre networks the answer is obvious. Along with AI clusters, and AI factories, data centres do the data collection and processing for Chatbots, Agentic AI and Physical AI. These are used for training and inference, and as a result, have very specific network needs to achieve high bandwidth, zero packet loss, and low latency. These include port speeds up to 800Gbs, on network fabrics that are non-blocking, non-drop, and support for RDMA (Remote Direct Memory Access). Such capabilities are provided natively with Infiniband or with U-Ethernet (Ultra-Ethernet) using ECN (Explicit Congestion Notification), PFC (Priority Flow Control) and RoCE (RDMA over Converged Ethernet).
Chatbots over the LAN: On the branch, campus, or enterprise LAN for Chatbots such as ChatGTP, Gemini, and Co-Pilot, the answer is equally obvious but quite the opposite. The local device and network neither captures the data points nor carry out any kind of processing. This is effectively just browser traffic that will lead to a brief spike. So, for Chatbots there is no impact to the LAN.
Physical AI on the LAN: The next use case, physical AI, is again quite clear. Physical AI such as autonomous vehicles and robots need to capture and process data in real time. In some cases, this will be done within the physical AI itself or with local edge computing, driving the need for significant east to west traffic flows that require ultra-low latency. To support physical AI and the required edge computing, re-architecting to high-speed network fabrics and standards such as Wi-Fi 7, will be necessary but in many cases will already be in place.
Agentic AI on the LAN: Finally, the area that I am not so sure on is Agentic AI. The agents can be local or remote in the cloud or the providers data centre, so processing can be local or remote. However, I think it is likely that there will be a lot of agent-to-agent communication, which without human intervention will happen quickly and will result in a more sustained traffic flow than with Chatbots. My instinct is that a full re-architecture outside of normal lifecycle management won’t be necessary for modern campus and branch environments. However, with agents running in the cloud, Internet may be a limiting factor.
Further Reading...