Tom Hollingsworth, CCIE #29213, is The Networking Nerd. He has spent the last twenty-five years implementing and understanding IT infrastructure, specializing in data center and campus networking, wireless and mobility solutions, and cybersecurity. He has extensive experience designing and implementing complex architectures and explaining their benefits to stakeholders and practitioners alike. Tom has hosted numerous Tech Field Day events focused on educating the wider enterprise IT community about solutions and products across the spectrum of offerings. He has participated in roundtable discussions and moderated panels on current and future technology outlooks. His advice is sought after by community members and company stakeholders at all levels. Tom has also hosted a weekly technology news podcast since 2018.
Predictions
East-West Traffic Dominates Data Centers
AI usage has changed traffic patterns in the enterprise data center. Traditional user-focused flows to servers (north-south) have given way to server-to-server traffic (east-west). By the end of 2026, east-west traffic will account for 90% of all data center traffic flows.
Why This Is Trending
Three primary factors have caused the shift:
The AI Multiplier: AI training workloads, such as LLMs or generative media, transform traffic patterns. Training these models requires GPUs to use all-reduce gradients constantly. This creates traffic patterns that are 24 to 32 times more intensive than cloud applications. When combined with the AI Factory model being championed by providers, the majority of traffic is server-to-server.
Microservices and Talkative Apps: Current software design utilizes microservices much more than previous iterations. A user request triggers hundreds of east-west server calls behind the scenes between databases, authentication servers, and other systems. Every year brings more of these applications, leading to a wider range of calls and more east-west traffic for every AI agent request.
Disaggregated Storage: AI clusters have minimal storage on the servers themselves. High-speed storage lives in its own space connected via fabric. Server DPUs handle I/O requests, making off-system storage usage transparent to the CPU and GPU. These requests still consume network bandwidth, turning every disk read into east-west traffic.
Vendor Impact
The massive traffic increase has forced organizations to rethink their architecture and develop new use cases that optimize hardware to best serve the systems that require priority.
No More Three-Tier: The historical three-tier model (Access -> Distribution -> Core) will only serve legacy clients and enterprises. The traffic route through the core to reach other servers cannot meet the needs of AI clusters. Oversubscription of distribution links creates expensive bottlenecks that can’t be eliminated with old thinking.
Standardizing on Leaf-Spine: The hyperscale leaf-spine (Clos) architecture provides the most predictable east-west traffic path in the data center. The spine switches’ connectivity means servers are only two hops away from their destinations, making load-balancing traffic much easier. Newer switch designs can scale to 128 ports at 800 GbE to support more GPUs connected to a network fabric while still delivering high-speed throughput.
The Rise of Rail Optimized Solutions: 2026 will see increased use of rail-optimized topologies, specifically designed to address GPUs that use all-reduce gradients. The GPUs in a server are each connected to a different switch in the leaf, which means each GPU has a dedicated path to its partner device. This path design eliminates port congestion and ensures that GPUs are never idle because of the network.