NVIDIA and Cisco announce USFP, a new optical module built to double AI network bandwidth per lane - Tech Critter

USFP is not a product refresh—it is a standards lock-in maneuver. By co-defining a per-lane bandwidth doubling spec, NVIDIA and Cisco force a ground-up redesign of the AI interconnect stack. The immediate technical shock hits the DSP and laser layers: at 200G-per-lane, conventional InP EML modulation bandwidth becomes the binding constraint, and the CPO transition window compresses from 2027 to 2026. The strategic subtext matters more. NVIDIA's Spectrum-X and Cisco's Nexus, jointly specifying USFP, create a de facto ecosystem that sidelines third-party module vendors from the spec-setting table. For Innolight, Eoptolink, and Coherent (Taiwan, China), the volume upside from AI cluster buildout is real, but margin power shifts decisively to the platform owners. Within 12–24 months: 1.6T module mass production accelerates to Q4 2025, breaking the industry's 2026 timeline; CPO moves from optional to mandatory as discrete optics hit thermal and density walls at 200G/lane; and OIF/IEEE standardization gets forced into fast-forward by a vendor-defined spec the industry must either adopt or fork. Mid-tier optical module players face a structurally shrinking addressable market.