Nvidia and Cisco develop USFP form factor: up to 12.8 Tb/s capacity - Techzine Global

The USFP joint development by Nvidia and Cisco is not a spec update—it is a standards lock-in play on AI interconnect. 12.8 Tb/s implies an 8x1.6T lane architecture, doubling DSP compute requirements and pushing Marvell and Broadcom's 800G DSPs toward their thermal ceiling. Silicon photonics shifts from optional to mandatory, as EML lasers hit physical power limits at 1.6T per lane. The deeper game is supply-chain sovereignty. Over 70% of global transceiver capacity sits in mainland China. If USFP enters the OCP standards framework, it becomes a de facto market-access barrier: uncertified modules simply cannot ship into North American hyperscale data centers. For Innolight, Coherent, and peers, an 18-month certification window separates sustained revenue from a cliff. CPO's commercial timeline gets compressed. When pluggable modules reliably deliver 12.8T, Nvidia and Cisco lose urgency to co-package optics with switch silicon before 2026, and Broadcom's CPO roadmap likely slips 12-18 months. The 12-24 month tail: silicon-photonics PDK services at Tower and GlobalFoundries become the new bottleneck; Chinese vendors will accelerate proprietary packaging standards to hedge against lock-in; and Nvidia is extending its moat from compute into the interconnect layer. That is the real strategic intent.