Co-packaged optics [/ˈko-pakkaged ˈɑptɪks/] n - Co-packaged optics (CPO) lifts the Silicon photonics engine from the switch faceplate and sets it within the same package as the switch ASIC or the accelerator, millimeters from the silicon rather than inches. The fiber then plugs straight into the package edge.

The purpose is almost wholly one of energy. To drive a high-speed electrical signal across a board toward a pluggable transceiver burns a meaningful share of power in the SerDes and retimers along the way. Shorten that electrical run to nigh zero and you cut both power and heat; estimates oft cited reckon a 30–50% reduction in interconnect energy at the switch. In a building where every watt spared at the chip is a watt never fed through the whole The Engines of the Hall chain nor drawn back out as heat, the saving compounds.

Why It Proves Hard

• Serviceability. [/ˈsɛrvɪkɪːbɪlɪti/] n A faceplate transceiver that dies is a two-minute swap. An optical engine fused into a switch package is not; one failed laser may condemn a very costly assembly. This remains a live argument against CPO, and a reason pluggable optics keep winning in many designs.
• Thermal coupling. [/ˈthɛrmæl ˈkæʊplɪng/] n The optics now share a package and a heat path with a chip dissipating hundreds of watts, and photonics despise temperature swings. This presses designs toward Direct-to-chip liquid cooling.
• Laser reliability. [/ˈlæsɛr ˈrɛlɪæbɪlɪti/] n External laser sources become the dominant failure mode, so many CPO designs place the lasers in replaceable modules at the edge.

Of Its Place

CPO stands among the clearest instances of the broader shift: as racks densify for AI, the old boundaries between “compute,” “network,” and “power and cooling” dissolve into a single problem of co-design. See AI factory power and cooling and Optical interconnects in AI clusters.

Part of The New Buildout.