SONiC separates the software controlling a network from the switches beneath it, giving operators a common operating system across hardware suppliers. At the Open Compute Project (OCP) Global Summit, Cisco, Microsoft, and Nokia backed its role in Ethernet networks serving AI workloads. Their contributions brought a practical question into focus: how does that open software model hold up when a fabric must manage congestion and recover from routing changes?
Route recovery puts the software model to work
SONiC is a Linux-based operating system that supports hundreds of switches and ASICs across vendors. Its networking features include Border Gateway Protocol (BGP), remote direct memory access (RDMA), QoS, and Ethernet/IP. Although it can serve conventional enterprise and hyperscale deployments, its growing use for AI networking is concentrated largely among hyperscalers.
Hardware choice is only part of that story. Microsoft and Nokia used a summit technical session to discuss community work on route programming, the process that converts routing decisions into forwarding state usable by switch hardware. Deepak Singhai, an engineering manager with Microsoft, explained that improvements become relevant with large route tables, frequent changes, or a need for rapid failure recovery.
The operational question is direct: how quickly can the hardware act on a changed route? Singhai said the community effort aims to accelerate route programming and recovery in large cloud and AI fabrics. He also linked longer recovery from network churn to greater customer impact, describing recovery as especially critical in AI data center scale-up scenarios.
That work gives SONiC’s AI positioning a concrete engineering basis. An open operating system offers flexibility, but the usefulness of that flexibility also depends on how promptly its routing decisions reach the forwarding hardware. The session described work toward that goal, rather than a completed performance result.
Cisco ties hardware choice to support continuity
Cisco’s announcements approached the same operational challenge from the deployment side. The company said Cisco-validated SONiC is available on its N9000 Series using Cisco Cloud Scale silicon for dense 100G and 400G leaf-spine deployments in enterprise and web-scale environments. It also described availability on the Cisco N9100 Series using Nvidia Spectrum-X ASICs for backend AI fabrics sensitive to packet loss, and an extension to the Cisco Silicon One ASIC–powered N9000 Series.
Cisco already supports SONiC on fixed and modular 8000 Series routers. Surbhi Paul, director, product management, for data center AI networking at Cisco, wrote that the expansion would complete integration across the company’s multi-silicon Nexus portfolio. Cisco framed that coverage as a way to connect WAN and data center operations through common FRR routing tools and a single Technical Assistance Center (TAC) contract.
The distinction is operational: common software does not, by itself, settle support responsibility. Cisco’s proposition combines SONiC’s open model with a vendor-supported path across its own hardware families. For teams weighing software choice against troubleshooting complexity, that is a more specific offer than compatibility alone.
The announcements expose several separate deployment considerations:
- Forwarding updates: Microsoft and Nokia’s community work targets faster route programming and recovery.
- Congestion controls: Cisco says architects can tune RoCEv2 PFC/ECN algorithms within SONiC on the Nexus platforms it described.
- Software placement: Paul said organizations can retain Cisco NX-OS or ACI in the core while using SONiC on identical Nexus hardware for specialized data pods.
Paul presented that mixed-software approach as protection for hardware investment as operational requirements change. It also suggests a selective adoption path, with SONiC assigned to particular parts of the network rather than requiring an immediate replacement of the existing core software. The practical decision becomes where a shared operating system helps, and where established application policies still determine the choice.
Community growth broadens the deployment work
The SONiC Foundation reported three additions to a membership already exceeding 43: Lenovo, NADDOD and Zhanwang (NEXAI). Their planned contributions reach into specific network functions. Lenovo is focusing on automation, telemetry, BGP EVPN and RoCE; NADDOD plans platform and SAI work alongside interoperability testing and deployment guidance; NEXAI intends contributions covering RDMA, RoCE, Priority Flow Control, SRv6, telemetry and AI network observability.
These are plans, not delivered improvements. Still, their scope places deployment experience and interoperability alongside protocol support in the community’s agenda. For an operating system spanning many hardware platforms, that is useful context.
The 650 Group forecasts that revenue from data center switching based on SONiC will exceed $5 billion in 2026, with hyperscaler demand and growth among neoclouds and enterprises contributing. The summit’s engineering discussions offer a grounded way to read that forecast. Broader adoption will need to be supported by work on forwarding updates, congestion tuning and support arrangements that operators can use in running fabrics.
