Cisco & NVIDIA Launch New Pluggable Optics for AI Factories - Data Centre Magazine

This is not a product launch—it is a vertical-integration power play. NVIDIA is extending its compute-fabric monopoly from GPU through NVLink and InfiniBand into the optical interconnect layer. Co-defining the pluggable spec with Cisco means writing the constitution of AI data-center networking, systematically compressing design leverage for upstream EML laser-chip suppliers and module makers in mainland China and Taiwan, China. The strategic calculus is clear: once the spec is locked by two system vendors, the commodity-optics market degrades into a qualified-supplier market. NVIDIA has every incentive to extend the pluggable lifecycle, protecting its InfiniBand and NVLink switching IP while deliberately slowing CPO adoption. On compliance, US export controls already restrict 800G-plus transceivers to China; a joint spec further entrenches the AI factory as a US-controlled, closed technical stack. Domestic alternatives face a moving target—specification iteration outpaces domestic catch-up cycles. Competitively, Broadcom's Tomahawk plus custom optics is the closest counter-position; Marvell differentiates via custom silicon photonics. The real losers are follow-on module makers without design IP. Over the next 12 to 24 months: 1.6T becomes the AI-cluster baseline; CPO mainstream deployment remains two to three years out; NVIDIA will likely embed optics into next-gen NVLink. The AI-factory framing is industrial policy in technical clothing—turnkey, US-spec, closed-ecosystem.