Optical modules · CPO · silicon photonics · AI interconnect · in depth

1 · Why CPO is the physical bridge from 800G to 1.6T

AI cluster scale doubles roughly every 18 months; network bandwidth follows at 1.5–2x. Front-panel pluggables (QSFP-DD / OSFP) at 800G are already near the SerDes-power wall — a single 800G module burns 15–18W today. On a B200 GPU with 8x 800G ports, network optics alone exceed 100W per GPU. CPO (co-packaged optics) moves the silicon-photonic engine inside the ASIC package, shrinking SerDes traces to <5mm and (per industry estimates) cutting 1.6T per-port power from 24W (pluggable) to 10–12W. In a GB200 NVL72 72-GPU rack, that drops optics power share from 18% to 9% — material for thermal budgets. Bottom line: CPO is not optional for 1.6T deployments in 2026–27; it is the only physics that closes the power gap. Broadcom CPO product page