Cisco Unveils Scale-Across AI Architecture at OCP 2026 Alongside NVIDIA Optical Pact and Hyperscale Silicon One Deployments
SAN JOSE — Networking enterprise Cisco has introduced a comprehensive suite of silicon, optical architectures, and open networking systems at the Open Compute Project (OCP) Global Summit 2026, addressing the physical requirements of hyperscale artificial intelligence clusters expanding from rack scale densities to distributed multi facility data center fabrics.
Central to the announcements is Cisco’s Scale Across architecture, an operational paradigm engineered to connect geographically dispersed artificial intelligence data centers into a single coherent compute fabric. While traditional data center interconnect systems link facilities for asynchronous data transfer, coordinated generative training workloads require low latency packet synchronization, deep packet buffering, and rapid link recovery to prevent GPU underutilization. Cisco addresses these requirements by unifying its Silicon One P200 processor, deep buffered switching systems, and coherent pluggable optics into a single packet to optical network layer.
Industry adoption highlights immediate production readiness. Cisco confirmed that leading hyperscalers have deployed its 51.2 Terabit per second (Tbps) P200 systems in live production environments. Microsoft has officially transitioned production cloud and high performance AI workloads onto P200 powered systems, citing enhanced memory buffering and throughput scalability. Concurrently, Alibaba Cloud Intelligence Group has integrated the P200 to power its proprietary Quantum Cat backbone routers across its next generation eCORE infrastructure, delivering 51.2 Tbps of routing capacity paired with custom systems and the open source SONiC operating system.
For intra data center networks scaling toward the 1.6 Terabit era, Cisco confirmed that its 102.4 Tbps Silicon One G300 silicon is currently shipping to enterprise customers. The vendor also unveiled the Silicon One G302, an industry first 51.2 Tbps switching processor built with 200G SerDes tailored specifically for top of rack deployments. Engineered to link high density server clusters into larger spine fabrics, the G302 incorporates direct to chip liquid cooling and native 1.6T optical interfaces, with customer shipments commencing before the close of 2026.
Expanding optical physical layer innovation, Cisco revealed a joint development initiative with NVIDIA to create the Ultra Dense Small Form Factor Pluggable (USFP) optical module. Operating at 400 Gigabits per second per lane, the USFP specification is designed to deliver four times the bandwidth density and 1.3 times the power efficiency of conventional 200G optics while maintaining the operational simplicity and field serviceability of pluggable transceivers. Both companies plan to release the architecture to the broader industry via an open multi source agreement.
Cisco further expanded enterprise software resilience across open AI fabrics with the introduction of Live Protect for Cisco validated SONiC on Nexus One architectures, offering in flight runtime mitigation against software vulnerabilities ahead of patch rollout. Additionally, the vendor launched the Cisco 8001, an out of band management appliance driven by the Silicon One A100 processor that standardizes remote infrastructure recovery and automation under unified SONiC operating controls.
