Cisco expands AI networking with new chips, SONiC, cross-site deployment

Cisco's move is not about adding chips—it's a power grab over AI network control. When 100K-GPU clusters hit bottlenecks in cross-site orchestration, thermal coordination, and operational consistency rather than raw bandwidth, whoever owns the network OS layer collects the "system tax" on AI infrastructure. Adopting SONiC signals Cisco is deliberately surrendering NX-OS's closed-stack premium in AI workloads. This is a direct flanking strike against NVIDIA's Spectrum-X and Broadcom's Thor, not a technical concession. Technical cascade: The open-source stack forces Arista and HPE Aruba to reprice their AI fabric offerings. Upstream ASIC design shifts from bandwidth stacking to low-latency cross-domain synchronization—a structural tailwind for MediaTek and Realtek in Taiwan, China. Market dynamics: NVIDIA will likely accelerate Spectrum-X's cross-datacenter variant, leveraging CUDA lock-in to counter Cisco's open narrative. Broadcom may bypass Cisco's channel entirely by targeting hyperscalers with Tomahawk 6 AI-optimized SKUs. 12-24 month call: The AI networking battleground migrates from chip performance to cross-site orchestration protocols. SONiC becomes the de facto standard-setter. Closed NOS share in AI scenarios drops below 30% by 2026.