Cisco (Nasdaq: CSCO) has introduced new silicon, optics and management capabilities designed to make AI infrastructure easier to connect and operate across racks, data centers and locations. The announcement reflects a shift in enterprise infrastructure planning: AI capacity increasingly depends on the network that connects distributed compute, not simply on the accelerators inside an individual cluster.
At the Open Compute Project Global Summit, Cisco described a scale-across architecture for coordinated AI workloads running across facilities. Traditional data-center interconnect technology links sites, but synchronized AI jobs place more direct pressure on congestion, latency and recovery because multiple accelerators can be participating in the same workload. Cisco said it is combining Silicon One P200 routing systems, deep-buffered platforms and coherent optics to connect those fabrics as a unified cluster.
For enterprise infrastructure teams, the practical benefit is the prospect of extending usable AI capacity beyond one building or one locally constrained cluster. That does not eliminate the operational complexity of distributed workloads, but it can give organizations a path to use compute capacity across sites while retaining a common network architecture. Cisco’s approach also emphasizes programmability and embedded security as part of the infrastructure rather than separate overlays.
Inside the data center, Cisco is extending its Silicon One portfolio with the G302, a 51.2-terabit switching chip built with 200G SerDes for top-of-rack systems. Cisco said the G302 supports 1.6-terabit optics and direct-to-chip liquid cooling integration, and is planned to begin shipping before the end of 2026. Top-of-rack switches aggregate server traffic and connect it to the broader fabric, making their capacity, density and cooling characteristics consequential as AI racks become more bandwidth-intensive.
The company also said it plans to work with NVIDIA (Nasdaq: NVDA) on an open industry specification for Ultra-Dense Small Form Factor Pluggable optics. Cisco expects the 400G-per-lane design to deliver four times the bandwidth density and 1.3 times the power efficiency of 200G-per-lane optics. Those are company projections, and the specification is still planned rather than a generally available product, but they illustrate why optical components are becoming a strategic part of AI infrastructure economics.
Cisco also introduced the 8001, an out-of-band management system using Silicon One A100. Out-of-band networks provide remote access when the production network is unavailable, which becomes harder when an operator must manage thousands of GPU nodes through separate legacy tools and interfaces. Cisco said the 8001 brings SONiC and familiar APIs to that management layer, allowing operators to use automation practices that more closely match the production fabric.
The announcements are aimed primarily at hyperscalers, neoclouds and large enterprises building substantial AI capacity. Still, the enterprise lesson extends to smaller deployments: networking, optical capacity and recoverable management paths need to be planned alongside GPU procurement. Cisco is positioning its stack as an open, programmable alternative to fragmented AI fabrics, where the network’s ability to scale and recover increasingly defines the value of the compute it connects.
That represents a change from legacy data-center designs, which could treat inter-site connectivity and management as supporting services. In AI operations, they are becoming central production infrastructure—and vendors that can combine scale, serviceability and operational control will have an advantage over point products focused only on raw bandwidth.