Where Is CPO Used Today, and What Does It Replace?

The core idea behind co-packaged optics (CPO) is to move optical-to-electrical and electrical-to-optical conversion from pluggable modules at the equipment's front panel to a location close to the switch ASIC. The conventional path is “switch ASIC → high-speed electrical traces on the PCB → pluggable optical module → fiber.” CPO shortens the electrical connection between the ASIC and the optical engine, then uses fiber to carry the data. This reduces the signal loss and compensation power associated with long electrical traces.

Known applications and the products they replace

1) Ethernet switches in AI data centers, connecting large numbers of servers and racks. CPO replaces switch-side pluggable optical transceivers and part of the high-speed electrical connection between the ASIC and those modules. Public disclosures show that Broadcom has delivered its 51.2 Tb/s Bailly CPO switch, while partner Micas has announced volume production of switches. NVIDIA says Spectrum-X Ethernet Photonics has entered production.

2) InfiniBand networks for AI and high-performance computing. CPO replaces pluggable optical modules at the InfiniBand switch end. NVIDIA has introduced Quantum-X Photonics, and cloud provider Lambda has publicly shown a Q3450-LD engineering sample and described its testing plans. Testing an engineering sample does not establish broad deployment.

3) Larger-scale accelerator interconnects. Depending on the design, CPO may reduce the need for some copper cables, board-level electrical connections, or pluggable optical modules. Broadcom positions Tomahawk 6–Davisson for large AI clusters and has announced early customer sampling. Optical I/O on compute chips themselves is at an earlier stage: Intel has demonstrated a prototype co-packaged with a CPU and transmitting live data.

The clearest replacement today is the switch-side pluggable optical transceiver, such as an OSFP module that previously occupied a front-panel port. The CPO switch shown by Lambda instead uses fiber-array connections at the front panel. However, the other end of the link may still use a pluggable optical module. CPO adoption therefore needs to be assessed at each end of a connection; it does not automatically mean that optical modules disappear from the entire link.

Some signal-compensation components may also become less necessary. Bringing the optical engine closer to the switch ASIC can change the need for retimers, DSPs, or equalization circuits previously used to overcome the losses of long, high-speed electrical traces. Exactly which components can be removed still depends on the product architecture.

CPO does not eliminate fiber, PCBs, or all copper connections. Some CPO designs retain replaceable external lasers at the front panel. NVIDIA's rack-scale networking designs also combine short copper cables with optical connections for scale-out networking. Longer-distance data center interconnects still include solutions based on pluggable coherent optical modules.

For the PCB and copper-clad laminate (CCL) industries, the point to watch is where demand moves. Some ultra-high-speed traces between the switch ASIC and front-panel optical modules become shorter, while package substrates, fiber coupling, thermal management, and the remaining high-speed electrical circuits receive greater design attention. This is an industry inference based on disclosed architectures. The adoption of CPO alone is not enough to determine whether the overall consumption of a particular CCL grade will rise or fall.