AI drives east-west traffic surges

Data-center networking used to be mostly about traffic coming in from users and going back out again. That was the old north-south model — front door traffic, basically. AI breaks that model. Training and inference both create huge volumes of server-to-server chatter inside the facility, and that is forcing operators to redesign networks, power systems, and even where they build next. ### What does “east-west traffic” actually mean? It means data moving sideways inside the data center — between GPUs, storage, orchestration software, APIs, and other internal services — instead of crossing the perimeter to a user or the public internet. That matters because classic enterprise networks were built to defend and optimize the edge. AI workloads spend much more of their life deep inside the building. ### Why is AI so much worse for internal traffic? Because modern AI jobs are distributed by design. A big training run is split across many accelerators that constantly exchange model states, gradients, and data shards. Inference can be lighter per request, but once it scales across many users and services it still demands fast, predictable internal links. Akamai says these environments need low-latency GPU-to-GPU and GPU-to-storage paths that old three-tier designs struggle to deliver. ### How big is the shift? Big enough that the old perimeter is no longer where most of the action is. Akamai says more than 76% of data-center traffic now flows east-west. IDC’s framing is similar but more architectural: AI fabrics increasingly need 400–800 GbE switching and leaf-spine back ends, because the bottleneck is no longer the internet uplink — it is the internal fabric between compute nodes. ### Why does that change the physical build? Because networking is now tied directly to power density and cooling. JLL says facilities are being redesigned for 100 kW rack densities and liquid cooling, while global occupancy has reached 97% and new capacity is constrained by power. Once racks get that dense, the network is not just cabling anymore — it becomes part of the whole mechanical and electrical design. ### So is the real bottleneck power, not fiber? Mostly, yes — but the catch is that power and fiber now travel together. Deloitte says US AI data-center power demand could rise from 4 GW in 2024 to 123 GW by 2035, with some grid interconnection requests facing waits of up to seven years. JLL says site selection has become “power first,” and CBRE says power scarcity is pushing construction time worthless no matter how good the land looks. ### Why are secondary markets benefiting? Because the biggest hubs are getting crowded, expensive, and power-constrained. CBRE says limited power in core markets is creating openings in newer hotspots, while McKinsey says training and inference are starting to pull site strategy in different directions: giant high-density campuses for star hubs. ### Does inference make this even more important? Yes. Training drove the first wave, but inference changes the map. JLL expects inference workloads to overtake training in 2027, and McKinsey says inference could make up a little more than half of AI workloads by 2030. That pushes operators toward facilities closer to users, with strong interconnects and enough local power to keep latency low. ### Bottom line? AI is turning the data center inside out. The winning sites are no longer just the ones with cheap land or good tax breaks. They are the ones with dense power, fast internal fabrics, liquid-cooling readiness, and enough fiber to keep thousands of machines talking to each other without slowing down.