Cisco unifies network for AI
- Cisco used two March 2026 posts to argue that AI data centers should stop splitting training and inference onto separate networks and instead run both on one “unified AI fabric.” - The clearest detail is economic: Cisco said unified-fabric operators can push gross margins from the mid-teens toward the high-30s, while dual-fabric rivals can leave “hundreds of millions” behind at scale. - The pitch extends Cisco’s broader AI stack, from Secure AI Factory with NVIDIA to Nexus Hyperfabric and 800-gigabit switches built for training, inference, and edge deployments. (cisco.com)
Artificial intelligence data centers are being built around two jobs: training models and serving answers, and Cisco is pushing customers to run both on one network. (blogs.cisco.com) In a March 31, 2026 blog post, Cisco said neocloud providers are choosing between “dual fabrics” that separate training and inference and “unified AI fabrics” that carry both workloads under one control plane. (blogs.cisco.com) Training is the heavy lifting phase, where thousands of graphics processors exchange data across a cluster. Inference is the production phase, where the model answers users in real time and has tighter latency commitments. (blogs.cisco.com 1) (blogs.cisco.com 2) Cisco’s argument is that separate networks raise capital and operating costs, make capacity harder to shift, and let higher-margin inference work move somewhere else after a customer finishes training. (blogs.cisco.com) The company put numbers behind that claim. It said dual-fabric operators tend to land in the mid-teens on gross margin, while unified-fabric operators can move toward the high-30s, with the profit gap reaching hundreds of millions of dollars at scale. (blogs.cisco.com) Cisco tied that case to products it has been rolling out this year. At NVIDIA GTC in March, it expanded Nexus Hyperfabric, a cloud-managed system for designing and operating AI infrastructure, and added the Nexus 9164E switch with up to 64 ports of 800G Ethernet. (blogs.cisco.com) That switch uses NVIDIA Spectrum-4 silicon and supports Remote Direct Memory Access over Converged Ethernet version 2, or RoCEv2, a way to move data between servers with less delay and less processor overhead. Cisco said that matters for distributed training, where idle graphics processors waste expensive compute time. (blogs.cisco.com) Cisco has also been packaging the network argument into larger systems. Its January 2026 AI POD design guide describes validated bundles of UCS servers, Nexus networking, Intersight management, and storage integrations with NetApp, Pure Storage, and VAST Data. (cisco.com) In March, Cisco and NVIDIA expanded Secure AI Factory to cover deployments from a central data center to local sites, linking the same infrastructure pitch to edge inference and agent systems that need faster responses near where data is created. (newsroom.cisco.com) Cisco is not announcing a single new standard here as much as pressing a design choice: buy one fabric that can juggle training and inference together, or keep paying for two. (blogs.cisco.com)
