Cisco, Nvidia push ultra-dense networking for Ethernet AI fabrics - Network World

This is not a partnership announcement—it is a power realignment in AI networking. Nvidia's pivot from InfiniBand path dependence to Ethernet signals that AI cluster scale has outgrown IB's architectural ceiling. Ultra-dense port density forces a generational jump in switch ASICs toward 256+ ports, directly benefiting Broadcom's Tomahawk 5 and Marvell's Teralynx 10. Optical transceiver demand shifts from IB-specific 800G to generic 1.6T Ethernet, requiring product-line recalibration across Coherent, Innolight, and peers. On compliance: Ethernet's openness reduces single-vendor lock-in, yet switch ASIC concentration remains dangerously narrow. If US export controls tighten further on AI infrastructure, Ethernet's universality paradoxically lowers the substitution barrier for Chinese vendors—Huawei CloudEngine and H3C already ship mature Ethernet switches, making replacement far cheaper than replicating a full InfiniBand stack. Competitive fallout: Arista Networks is the most exposed. Its "Ethernet for AI" narrative is now being defined by Cisco and Nvidia, compressing pricing power. AMD's MI300 faces structural disadvantage if cluster networking is locked into Nvidia's ecosystem. 12-to-24-month outlook: Ethernet AI fabric becomes the de facto standard for hyperscale data centers; InfiniBand retreats to HPC niches. The real battleground is not silicon—it is the protocol and traffic-scheduling layer. Whoever owns AI flow orchestration owns pricing power in the next networking era.