Co-Packaged Optics (CPO)

What Is Co-Packaged Optics?

In a pluggable switch, the ASIC sits in the middle of the board and the optics sit at the faceplate. Between them is ten to thirty centimetres of printed circuit board carrying signals at 100 or 200 gigabits per lane. That channel is lossy, and the electronics at both ends burn power fighting it.

Co-packaged optics removes the channel. The optical engines are mounted on the same package substrate as the switch ASIC, millimetres away. Fibre, not copper, carries the signal off the package. What used to be an electrical problem becomes an optical one at the earliest possible point.

The framing that helps most: CPO is not a better transceiver. It is the same optics moved to a place where the hard part of the problem does not exist.

How It Works

The engines are almost always silicon photonics, because silicon photonic circuits can be built at the densities and yields co-packaging requires. Silicon photonics puts waveguides, modulators and germanium detectors on a die the size of a fingernail, with eight or sixteen channels where discrete assemblies would need eight or sixteen separate parts.

Lasers are usually external — mounted separately and fed in by fibre — because lasers are the least reliable and most temperature-sensitive element in the system and you do not want them bonded to a chip that runs hot. An external laser source also means a laser failure does not condemn an engine.

Fibre reaches the package through a dense connector array. This is where new very-small-form-factor connectors matter: an ordinary MPO is far too bulky to land the fibre count a CPO package needs in the space available. See VSFF connectors.

Why It Is Being Taken Seriously Now

Two pressures arrived together. Power is the obvious one: at 51.2T a switch full of pluggables is dissipating an uncomfortable amount in the optics alone, and the faceplate has run out of room to cool them.

The less discussed one is signal integrity. At 224G per lane the symbol rate is around 112 gigabaud, and at those frequencies a long circuit-board channel is genuinely difficult — not expensive, difficult. Shortening it from centimetres to millimetres is the most direct fix available, and it happens to solve the power problem at the same time.

That is why the CPO conversation accelerated precisely as 1.6T became real. It is not fashion; it is the board running out of margin.

The Objection That Has Not Gone Away

A pluggable that fails is swapped in seconds by whoever is on shift. A failed co-packaged engine is a switch-level repair — which in practice means a spare switch, a maintenance window, and a fabric designed to lose a whole node gracefully.

Operators do not dismiss this lightly, and they are right not to. It changes the spares model, the repair skill set and the redundancy design all at once. Any CPO evaluation that does not have an answer for ‘what happens when one engine fails in year three’ is incomplete.

This is exactly the gap XPO is trying to fill: co-packaged density with a field-replaceable element.

Where You Meet It

CPO is aimed squarely at the largest switches — the 51.2T and 102.4T generation and beyond. It is a hyperscale and AI cluster technology first, and will reach the enterprise late if at all.

For anyone specifying infrastructure today, the practical consequence is on the cabling rather than the switch. A CPO switch still terminates in fibre, still needs structured plant, and still uses the same reach classes. What changes is connector density at the switch face — which is a patching and panel question, not a glass question.

Related Terms

NPO is the near-packaged compromise. LPO is the competing answer to the same power problem, and the four-way comparison sets them side by side.