Ask most people what limits AI infrastructure, and you’ll hear about the accelerator. How many GPUs you can get, how much HBM they carry, how much power they draw. That focus makes sense. Accelerators are the biggest line item in the rack. But no accelerator works alone. Training and inference increasingly depend on moving data between chips, between racks and, more and more, between datacenters. When the network falls behind, the GPU waits. And an idle GPU is one of the most expensive things in the datacenter.

So connectivity deserves more of the spotlight than it gets. At the Open Compute Project (OCP) Global Summit this week, Cisco made a strong case for it. The company introduced new silicon, systems, and optics, including a new optical module effort with NVIDIA. More importantly, it showed that Cisco’s open networking strategy now has production deployments behind it, including ones where the customer builds the system and chooses the software.

A Different Kind of Cisco

For most of its history, Cisco sold integrated systems. Its silicon, its hardware, its operating system. That model built the enterprise network. But this isn’t how hyperscalers like to buy. They want open software, disaggregated hardware, and the freedom to build their own systems.

Cisco has spent years adjusting to that reality. Silicon One, introduced in 2019, gave Cisco a single networking architecture it could use in its own systems or sell to others as merchant silicon. The Acacia acquisition brought coherent optics in-house. And Cisco became a major contributor to SONiC, the open-source network operating system behind much of the hyperscale world.

Its relationship with NVIDIA fits the same pattern. The two companies have been making Spectrum-X Ethernet and Cisco Silicon One-based networks work together. OCP 2026 adds another layer to that collaboration.

What Cisco Announced

• USFP optics with NVIDIA. The companies plan to introduce a new pluggable optical module, the Ultra-Dense Small Form Factor Pluggable (USFP), built for 400 Gb/s per lane. The first designs fit 16 lanes, or 6.4 Tb/s, into a module about the size of today’s OSFP, and the architecture is designed to scale to 32 lanes and 12.8 Tb/s. Cisco and NVIDIA claim 4x the bandwidth density and 1.3x the power efficiency of 200Gb/s-per-lane optics. An open, multi-source specification is planned, with TE Connectivity and Amphenol already involved.
• Silicon One P200 in hyperscale production. Microsoft and Alibaba are both running P200-based systems in production.
• G300 shipping, G302 announced. The 102.4T G300 is now sampling, and the first G300-based systems are shipping. The new G302 is a 51.2T top-of-rack chip with 200G SerDes, built to power systems with 1.6T optics and direct-to-chip liquid cooling. It’s planned to ship before the end of 2026.
• Cisco-validated SONiC. Customers can now adopt SONiC with Cisco’s Nexus One architecture, with Live Protect coverage coming in mid-2027.
• Cisco 8001. A Silicon One A100-based out-of-band management system that brings SONiC to the management network.

Production Is the Proof Point

The headline was the NVIDIA partnership. For me, though, the most important news is the customer list.

Just take a look at what Alibaba did. It built its own Quantum Cat router for its backbone, put Cisco silicon inside, and runs open-source software on top. It’s that open model working as promised, with Cisco as a willing part of it.

Microsoft’s move to production adds a second hyperscale reference.

Why does P200 matter so much? Because of where AI capacity may be heading. Power availability is increasingly shaping where AI capacity gets built, and no single site can always get enough of it. So operators are spreading capacity across multiple datacenters and asking those sites to act like one giant cluster. Cisco calls this scale-across. It’s a different problem than traditional datacenter interconnect. When GPUs in two buildings work on the same job, congestion or a slow recovery between them directly costs compute time. Deep buffers, programmability, and coherent optics are what that problem demands, and those are areas where Cisco has deep roots.

To be clear, the Microsoft and Alibaba deployments are cloud and backbone networks. They don’t prove out coordinated AI jobs across sites. But what they do show is P200 running at hyperscale in production, and distributed AI is the next opportunity I see.

Why USFP Matters

Here’s a constraint most people outside networking never think about: the front of the switch. There’s only so much room for optical ports. When a switch runs out of ports, expansion can require another tier of switches. That means more optics, more power, more latency, and more cost. So if you can squeeze more bandwidth into each port, you push that wall further out.

Co-packaged optics is one answer, and it will have an important place in AI fabrics. USFP gives operators another path. It keeps the pluggable model, so a failed module can still be swapped in the field. If you’ve ever run a datacenter, you know how much that matters. And because USFP is planned as an open, multi-source specification, customers should get supply chain choice, too.

One caution: USFP is a plan, not a product. I haven’t seen a sampling or availability date yet. Still, having Cisco’s optics expertise and NVIDIA’s AI networking scale behind one design gives it a strong start.

The Operations Story

Two smaller announcements caught my attention, because they solve real operational problems.

The first is the Cisco 8001. Out-of-band management is the network you rely on when everything else is down. At AI scale, it can be (and often is) a patchwork of older gear and separate tools. Bringing SONiC and the same automation used in the production fabric to the management layer is a welcome improvement.

The second is Live Protect for SONiC. Open networking is great, but many enterprises hit pause over support. Who stands behind it when a vulnerability drops? Starting in mid-2027, Live Protect will give SONiC customers on supported products and releases time to test and roll out the permanent fix without disrupting operations. That should make open networking an easier decision for organizations without a hyperscaler-sized engineering team.

A Market Big Enough for Several Winners

Ethernet has established itself as a major fabric for AI scale-out networking, and the market around it is busy. Broadcom’s Tomahawk and Jericho families power systems from many vendors. NVIDIA’s Spectrum-X pairs its switches with its own SuperNICs for a tightly integrated AI fabric, and Meta’s plan to adopt Spectrum-6, announced the same day, shows that momentum. Arista and a broad ecosystem of system builders give customers still more options. And the largest operators increasingly design their own systems around merchant silicon.

I don’t see this as a winner-take-all race. Most buyers want more than one source. In that market, Cisco’s position is distinct. It’s one of the very few companies that build the silicon, systems, optics, and software, and it gives customers multiple ways to buy into that portfolio. That flexibility fits how hyperscalers, neoclouds and enterprises now buy.

The NVIDIA partnership reinforces the point, and Cisco’s work on OCP’s Ethernet for Scale-Up Networking (ESUN) specification points the same way. The industry is collaborating on standards even as it competes on products.

What This Means for Enterprise IT

I see three takeaways.

Budget for the network as part of AI. An AI cluster is only as fast as the fabric connecting it. If your plans start and end with GPUs, you risk paying for accelerators that spend time waiting. Treat bandwidth, optics, and network power as first-class parts of the AI design.

Take another look at open networking. Running SONiC has often meant leaning heavily on your own engineering team. Vendor validation and security backing can reduce that operational burden. If open networking was off the table before, it’s worth revisiting.

Don’t overlook the management network. As AI environments grow, the out-of-band network becomes more critical, not less. Make sure it can scale and recover with the rest of the infrastructure.

The Bigger Picture

AI infrastructure conversations still start with the accelerator. They shouldn’t end there. As clusters grow and span sites, the network decides how much useful work every GPU actually delivers.

As an analyst who has been compute-first for most of my career, I find Cisco’s OCP news significant. It shows a company that has embraced openness and now has production deployments to back it up. That matters beyond any single product. It means Cisco can remain relevant in more infrastructure decisions, including those where it doesn’t supply the entire system. Its silicon can sit inside a customer’s own router. Its optics work can help shape an industry specification. And its SONiC support can make open networking practical for buyers who aren’t hyperscalers.

Keep an eye on the network, and on Cisco’s place in it. That’s where much of the next phase of AI economics will be decided.