Data center networking is the set of hardware, software, and services that connects servers, storage, and accelerators to each other inside a data center, and connects those data centers to each other and to the outside world. The category includes switches, routers, network interface cards (NICs), DPUs, cabling, transceivers, and the fabrics that tie them together — predominantly Ethernet, with InfiniBand holding a share of AI clusters. We think the clearest way to understand the category today is to split it into two networks: the scale-out network that connects many servers into a cluster, and the scale-up network that connects many accelerators into what behaves like one computer.
The scale-out network is the traditional data center network: leaf-and-spine Ethernet switches moving traffic between server racks, north-south and east-west. Its economics are well understood, its vendors — Cisco, Arista, Juniper, NVIDIA, and the white box ecosystem built on merchant silicon — are established, and its speeds have stepped from 25G to 100G to 400G ports over the past decade. The scale-up network is newer in economic terms: the fabric that binds 72, 576, or thousands of accelerators into a single training domain, where bandwidth per GPU has climbed past 7 Tbps in systems like the GB200 NVL72 and where copper, optics, and increasingly circuit switching compete for every link.
The reason this category matters now is size and growth. As 650 Group has published, total data center networking is on a path to approach and potentially exceed $200B a year by the end of the decade, driven by AI. The AI portion of that spend is the fastest-growing: our research puts data center AI networking at nearly $20B in 2025, spanning the switch, NIC, and interconnect layers attributed to AI deployments. Two rulers apply here — total data center networking and the AI-attributed slice — and mixing them produces bad sizing. We publish both, quarterly, broken out by segment and vendor.
The components matter to buyers in different ways. Switches are the largest single line: fixed and modular Ethernet systems at 400G, 800G, and now 1.6T port speeds, from Cisco, Arista, NVIDIA, and the merchant-silicon white box ecosystem. NICs and DPUs are the second line — the endpoints that move data and increasingly offload compute tasks (our SmartNIC and DPU research covers that market). Optics and cabling are the third: transceivers, AOCs, and DACs whose mix shifts with reach, power, and speed, and whose economics increasingly decide fabric architecture. Software — the network operating systems and fabric controllers — is the layer that ties the first three together.
We believe the next five years will reorder this market more than the last ten did: Ethernet displacing InfiniBand across an expanding set of AI clusters, scale-up architectures pulling bandwidth per system higher than scale-out ever demanded, and optics moving closer to the accelerator package. Price per port, power per port, and software quality will decide who wins each layer, and the mix between copper and optics will shift with every speed step. We track the category quarterly and publish our data center networking analysis here.
Tags: #Data Center Networking #Ethernet #InfiniBand #Switch #NIC #DPU #Optics #AI #650 Group
650 Group researches several of the technologies we covered in this explainer, including Data Center, Co-Location & Cloud and Enterprise, and more.