OcNOS-DC: the data center NOS.

OcNOS-DC is IP Infusion's data center variant of OcNOS, the network operating system in production with 600+ operators across 60+ countries. It is built for leaf-and-spine fabrics that carry EVPN-VXLAN overlays, lossless RoCEv2 traffic for AI and ML clusters, and 400G or 800G DCI links. The OS runs on open Broadcom-based switches from Edgecore and UfiSpace, giving operators a commercial alternative to Arista EOS, Cisco Nexus, and DIY SONiC builds.

OcNOS-DC implements EVPN-VXLAN with both Type-2 MAC/IP and Type-5 IP-prefix routes, symmetric IRB, distributed anycast gateway, and ARP suppression. EVPN multi-homing with all-active ESI LAG replaces legacy MLAG for dual-homed servers and storage. The control plane uses MP-BGP per RFC 7432 and RFC 8365, so it interoperates with mainstream EVPN implementations from other vendors.

OcNOS-DC ships with Priority Flow Control (PFC) and Explicit Congestion Notification (ECN) profiles pre-tuned for RoCEv2, so RDMA traffic classes can be paused on a per-priority basis to avoid GPU-to-GPU packet drops. The intent is a zero packet loss fabric for xCCL collectives such as NCCL and RCCL. Real production deployments to date have run on the order of 20 to 200 GPUs, and the design scales further as engagements grow.

OcNOS-DC is offered in three SKUs. DC-MGMT provides Layer 2 and Layer 3 features for out-of-band management switches. DC-IPBASE adds EVPN-VXLAN overlay and lossless Ethernet for AI fabrics, and DC-PLUS extends that with streaming telemetry and 400G ZR/ZR+ coherent optics for data center interconnect.

OcNOS-DC supports 18 open hardware platforms from Edgecore and UfiSpace, spanning 48x25G ToR leaves up to 64x800G spines. Supported Broadcom silicon includes Tomahawk 5 (51.2T, 800G), Tomahawk 4 (25.6T, 400G), Tomahawk 3, Trident 4, and Trident 3. The same NOS image runs across the portfolio, so operators can standardize one CLI and one automation toolchain across leaf, spine, and border roles.

OcNOS-DC exposes NETCONF, RESTCONF, and gNMI/gRPC northbound, with OpenConfig and native YANG models, so it slots into existing Ansible, Terraform, or in-house controller pipelines. Zero Touch Provisioning uses DHCP options 66/67 to fetch the image and day-zero config, which lets a new leaf or spine join the fabric without console access. Streaming telemetry is available in the DC-PLUS tier for closed-loop assurance.

These are three ways OcNOS-DC scales the same fabric. Scale-up grows a single pod's capacity by raising per-switch bandwidth and port density, for example moving from a 100G fabric to 400G and 800G, so the same leaf-spine-super-spine topology feeds larger, higher-bandwidth compute nodes without adding more pods. Scale-out grows capacity by adding more of the same building blocks, more spine and super-spine switches and leaf pairs, widening the Clos topology horizontally so more racks and leaves connect while every endpoint keeps any-to-any reachability. Scale-across extends the fabric beyond a single data center, using a global EVPN and BGP control plane over a long-haul coherent interconnect to link independent DC pods into one logically unified fabric. OcNOS-DC runs all three under one NOS and one control plane.