Every switch in an AI data center has to turn electrical signals into light. The question the industry has been working on is where that happens. The closer the optics sit to the switch chip, the less power it takes to get the signal there.
Linear pluggable optics (LPO), near-packaged optics (NPO) and co-packaged optics (CPO) are three steps along that path. This guide walks through each one, how they compare, and why CPO is the end point.
The one idea behind all three
In a traditional switch, the electrical signal leaves the switch chip (ASIC), travels across centimeters of circuit board to a transceiver in the front panel, and gets cleaned up by a digital signal processor (DSP) before it becomes light. That trip, and the DSP at the end of it, cost a lot of power at 51.2 Tb/s.
LPO, NPO and CPO each attack that cost, in different ways and to different degrees.
What is LPO?
LPO keeps the form factor data centers already know. Modules still plug into the front panel and can be swapped by hand. Dropping the DSP saves power and latency in each module.
- What it keeps: the board trace between the chip and the front panel. The chip's SerDes has to be strong enough to drive that distance without help.
- Where it fits: shorter-reach links where operators want lower power without changing how they buy and service optics.
What is NPO?
NPO shortens the electrical path from centimeters of board trace to a short hop beside the package. That cuts more power than LPO, and it frees the front panel for fiber connections. The optics are no longer hand-swappable modules, so servicing moves toward the system level.
What is CPO?
This is the shortest possible path. No board trace, no DSPs, and no transceivers to buy for each port. Fiber connects at the front panel and runs almost directly to the chip package. For the full explainer, see What Is a CPO Switch?
LPO vs. NPO vs. CPO at a glance
| LPO | NPO | CPO | |
|---|---|---|---|
| Where optics live | Front-panel pluggable modules | On the board, next to the chip package | Inside the chip package |
| Electrical path | Centimeters of board trace | A short hop beside the package | Millimeters, on the substrate |
| DSPs | Removed | Not needed | Not needed |
| Servicing optics | Swap a module by hand | System level | System level |
| Power saving | Good | Better | Best (over 50% in Micas testing) |
Why CPO is the end point
Each step removes more of the distance between the chip and the light, and CPO removes all of it. In Micas testing at 40 °C ambient with full line-rate bidirectional traffic, a Tomahawk 5 switch with 400G DR4 pluggables drew 3,267 W. The Tomahawk 5 CPO switch drew 1,626 W.
LPO and NPO are useful steps, and many networks will run a mix for years. But at 51.2T and beyond, the operators who need the most bandwidth per watt are moving to CPO.
Hear it live at OPC 2026
We're walking through the full LPO → NPO → CPO path in a 20-minute talk, with time to see where each one fits in real AI clusters.
Frequently asked questions
Is LPO the same as CPO?
Do I have to go through NPO to get to CPO?
Does CPO need liquid cooling?
What happens if an optical engine fails in CPO?
Ready to skip ahead to CPO?
See the Micas 51.2T CPO switch in person, or talk to our team about where co-packaged optics fits in your network.
Talk to our team →