Open networking has settled the technical argument, so the advantage has moved from openness itself to operational maturity: how well the network runs, and who owns the integration work that gets it there. That work used to require a platform team of your own. It is now something an operator can either keep in house or buy as a complete system.
The networks going into AI data centers and modernized service-provider cores are among the largest infrastructure investments an operator will make. The operating model chosen for them moves the three numbers that decide the business case: capital cost, time to deploy, and operational risk. Operators that have made the move report network capital cost cuts as large as 60 percent, and new networks live in months rather than quarters.
The technical debate is effectively over. OcNOS carries production traffic in more than 600 operator networks across 60 countries, on more than 40 validated hardware platforms, and 650 Group has forecast the disaggregated router market growing more than 50 percent a year through 2027. The question stopped being whether open networking works.
The first decade of open networking was about separating hardware from software. The next is about putting them back together, without returning to lock-in.
So the advantage is no longer openness itself but operational maturity: how well the network runs, and who should own the integration to get there.
Where the cost actually is
That shift reflects where the real costs have always been. Buying the open hardware was never the hard part. The operational work was: sourcing the boxes, matching the optics, testing the configurations, coordinating the vendors, troubleshooting the failures at three in the morning. For operators without a platform team, that work often outweighed the savings going open promised.
That burden scales with the traffic, and nothing is harder right now than an AI fabric. Ethernet has won the transport argument there, taking about two-thirds of AI-cluster switch sales by the first quarter of 2026 according to Dell’Oro Group, and it settles less than it sounds like: performance turns on adaptive load balancing, lossless RoCEv2 with DCQCN, a PFC deadlock watchdog, and before long Ultra Ethernet.
A GPU waiting on the network is a GPU you are paying for and not using. Tuning all of that, proving it on your own silicon, and keeping it correct through every firmware revision is what integration means now. The AI fabric page and the AI Fabric Design Suite carry the detail.
It is real engineering, and someone has to do it. Until recently that someone had to work for you. Now it is something you can order: vendors ship complete open systems, not only software licenses. IP Infusion’s OcNOS Systems is validated open hardware with the software pre-installed, qualified optics, one purchase order and one support agreement. Operators that would rather source and integrate their own validated hardware keep licensing OcNOS on its own. What changed is that the integration project became a choice rather than the entry fee.
| Build it yourself (license OcNOS as software) |
Buy the system (OcNOS Systems, pre-integrated) |
|
|---|---|---|
| Hardware | You source it from the qualified list | Sourced and pre-loaded, priced into the system |
| Integration | Your platform team | Done before it ships |
| Support | IP Infusion SLA on the software, hardware and optics from their vendors | One purchase order, one SLA across software, hardware and optics |
| Strength | Maximum control, lowest cost at scale | Fast to deploy, low operational overhead |
| Trade-off | You own the labour and the lifecycle | Single-source by design, hardware choice narrows to the validated list |
| Best fit | Teams with platform engineering to spare | Teams without it |
Both are open networking. The difference is who does the integration.
What operators report
These are individual deployments, not a survey.
NWPS built a greenfield BGP multihoming network for a Swiss hospital on OcNOS and UfiSpace routers, holding more than 1.27 million routes in hardware. It reports roughly 60 percent lower capital cost than a comparable proprietary offer, no recurring software fees, and no recorded outages across ten months of production.
Vyve Broadband did not do its own integration. RocNet Supply stood up the lab, validated the configuration and standardized the deployment plan, and the rollout began within three months. RocNet’s chief executive put that at about half the time to revenue they typically see. Vyve is consolidating three merged networks onto OcNOS, swapping a different network operating system on Edgecore white box hardware it already ran, on a migration tracking toward 200 nodes and the displacement of Cisco and Juniper gear.
Uniti Fiber cut energy consumption on its upgraded routing layer to roughly one sixth of its legacy proprietary footprint, and the NTT IOWN 400G solution, built with IP Infusion among eight partners, targets a 50 percent cut in equipment cost and a 40 percent cut in power.
So who should still build it?
Self-integration remains the right answer for teams with the engineering capacity for it, from hyperscalers to regional operators with real platform skills. That path is still ours. License OcNOS on its own, source the hardware from the validated list, and keep the integration in house. It is the same software and the same roadmap either way. What changes is who does the assembly and how many suppliers you carry.
For everyone else, a pre-integrated system removes the project without giving up the open hardware or the pricing that comes with it. None of this makes open networking effortless. Someone still has to qualify the next firmware, keep the fabric from wedging under load, and answer the phone at three in the morning. Ten months without a recorded outage is an operational result, not a procurement one. What changed is who is on the hook for it.
If you want to go further:
- See what ships in an OcNOS System, or compare OcNOS platform by platform with Cisco and Juniper.
- Download the free OcNOS VM to evaluate it yourself, see how to buy, or talk to a specialist.
Market figures are drawn from public reporting by Dell’Oro Group (first quarter 2026) and 650 Group (2023 forecast). The Dell’Oro figure counts Ethernet switch sales in AI back-end networks regardless of vendor or business model, so it measures the transport rather than the open networking model. Customer figures come from IP Infusion case studies and public announcements for NWPS, Vyve Broadband, and Uniti Fiber; the Vyve migration is in progress, and the NTT IOWN figures are stated targets for that solution rather than measured results. The 90 percent fabric-utilization figure is an IP Infusion design target for dynamic load balancing on Broadcom Tomahawk 4 and 5, not a published benchmark. AI networking capabilities are based on IP Infusion product documentation. Cisco, Juniper, Nokia, NVIDIA, and Broadcom are trademarks of their respective owners. IP Infusion is an independent company and is not affiliated with them.