As data centers scale to 800G and 1.6T for AI workloads, traditional pluggable optics face rising power and thermal challenges. This drives three new packaging and interconnect architectures.
CPO (Co-Packaged Optics) – Moves optical engines close to the switch ASIC, drastically shortening electrical paths. This reduces or eliminates high-power DSP/retimer functions, delivering superior power efficiency and density. Drawbacks: high manufacturing complexity, major board redesign, and poor field repairability (faulty optics require full board replacement).
NPO (Near-Packaged Optics) – A balanced intermediate: optics are separate from the ASIC but placed nearby on the same board. Offers meaningful power savings with less complexity than CPO, while retaining reasonable maintainability. Requires custom motherboard and enhanced cooling, but no full package integration.
XPO (eXtra-dense Pluggable Optics) – Retains front panel pluggable form factor, scaling per-module throughput to 12.8Tbps. Uses DSP free linear optics (like LPO) to improve efficiency and latency. Enables high density with upgrade flexibility, but demands advanced host SerDes and liquid cooling for higher thermal density.
Key Comparison
OCS (Optical Circuit Switching) – Complements these by using MEMS mirrors for all-optical routing, bypassing electronic switching. Reduces latency and power, with port densities reaching 320×320 (OFC 2026 roadmap). When combined with CPO/NPO, OCS enables fuller all-optical paths, promising for high-density AI clusters.
Outlook – NPO is currently preferred for 1.6T upgrades due to balanced power/speed. LPO and XPO find niche roles where pluggable flexibility is valued. CPO is poised for long-term growth as silicon photonics matures, especially for ultra-dense, power-constrained AI supercomputers, while NPO remains a reliable mid-range solution. Eventually, pluggable optics will hit physical limits, and CPO is expected to become central in high-speed interconnects.
