Reading the OpenCPX MSA, the socket that makes co-packaging serviceable

The least glamorous document at the show

Around OFC 2026, the industry announced three interconnect agreements:

• One defined the light between compute packages.
• One defined a 12.8-terabit liquid-cooled module.
• One defined, essentially, a socket.

Guess which one this site thinks matters most to whether co-packaged optics (CPO) actually happens.

The placement ladder explains the stakes. A move of the optics from the faceplate toward the package roughly halves energy per bit. The pluggable band is 15 to 20 pJ/bit, and co-packaged systems report 5 to 10. Every serious roadmap wants that saving.

Photonics is not the obstacle. The obstacle is a question from operations. Operations teams ask it in the first meeting, and vendors answer it last. The question is this: when a laser dies at two in the morning, what exactly do we do?

At the faceplate, the industry perfected the answer over twenty years. You pull the module and insert a spare, hot, in one minute. On a package, with the engine soldered down, the honest answer is a rework station or a replaced switch assembly. The failure domain grows from a thousand-dollar box to the most valuable board in the chassis. That answer fails procurement and fails operations. It also quietly limits how far down the ladder anyone dares to go.

OpenCPX exists to change that answer.

What a socket buys

The MSA’s scope is basic infrastructure. It covers these items for a socketed optical engine that serves both near-package and co-packaged designs:

• connector mechanicals
• thermal interfaces
• electrical pinout
• mechanical form factors
• the electrical, optical and management interfaces

No item in the list needs exotic physics. That is deliberate.

Each plain item buys something concrete:

• A defined mechanical socket makes a failed engine a part swap at the shelf. That is a planned outage instead of a rework. It holds serviceability at what the ladder essay calls a “shelf visit” instead of a “board rework”.
• A defined pinout and management interface lets an engine from one vendor replace an engine from another vendor in the same socket. That is the whole difference between a supply chain and a sole source. The supply-chain essay shows where the optics layer already has more than one source.
• A defined thermal interface makes the cold plate part of the contract rather than an afterthought. That matters because engines sit beside 100-kilowatt-class racks.

The membership makes sense against that list. The founding members are Ciena, Coherent, Marvell, Molex, Samtec and TeraHop. Early contributors are Accton, Intel, Nexthop, Qualcomm and TE Connectivity. These are connector companies, optics makers and DSP silicon vendors: the companies that must physically interoperate. Compare OCI’s roster of hyperscalers and compute silicon, and the division of labor across the three MSAs becomes clear.

Our stake, stated plainly

This section is a disclosure, not a neutral reading: our 6.4T near-package transceiver is built to this MSA. We chose it for exactly the reasons this essay gives. NPO (near-package optics) with a standardized socket keeps the optics a component. It stays separately manufactured, separately tested, separately replaceable and second-sourceable. It also captures most of what co-packaging exists to capture. When we argue that the socket is the difference between a demo and a deployable product, we explain our own engineering. We do not comment from a distance.

There are also honest open questions against the socket, and they are worth stating. A socket adds an electrical discontinuity exactly where signal integrity matters most. So the connector must be very good, and that is why the connector companies are in the room. Also, a standard is interoperability on paper until two vendors’ engines replace each other in one socket in public. Watch for that demonstration. It is the moment this MSA becomes real.

What to take away

1. OpenCPX standardizes the socketed optical engine for NPO and CPO alike. It covers mechanicals, thermals, pinout and management.
2. The problem is operational, not photonic. The socket turns a board rework back into a shelf-level part swap. Multi-vendor sockets turn a sole source back into a supply chain.
3. The member list shows the logic: connector companies, optics makers and DSP makers. That is the layer that must physically interoperate.
4. Watch for the first public two-vendor engine swap in one socket. That demonstration, not the document, is the milestone.
5. Our 6.4T NPO part is built to this MSA. The socket argument is our argument, stated in the open.