Optical MSA Tracker

Last updated: 12 October 2026

Bandwidth in the AI data center doubles every generation — and what the next generation of optical modules actually looks like is increasingly decided not by an IEEE standard, but by a multi-source agreement (MSA) that a handful of vendors sign first. Four new MSAs appeared in the first half of 2026 alone, all aimed at AI scale-up.

This tracker covers the 15 MSAs, implementation agreements and standards bodies on the main line of optical and AI interconnect. For each one: the problem it set out to solve, who is behind it, the technical specification, and how far that spec has actually progressed.

OCI MSA — Copper can’t reach far enough for scale-up, and existing optical modules burn too much power.
XPO MSA — The faceplate is out of room for more pluggables, but the industry isn’t ready to commit to full CPO.
Open CPX MSA — Every CPO optical engine is proprietary, leaving buyers with no second source and no way to service them.
ACC-MSA — Passive copper runs out of reach; DSP-based active copper costs too much and draws too much power.
LPO MSA — The DSP inside an optical module accounts for most of its power and most of its cost.
RTLR / LRO — Fully linear is hard to close, yet the receive-side DSP is the one most worth removing.
OSFP MSA — Past 400G, modules run too hot for the existing QSFP mechanical envelope to handle.
QSFP-DD MSA — Eight lanes are needed, without forcing customers to replace every QSFP port they already own.
SFP-DD MSA — The server side needs twice the bandwidth of SFP, in the same mechanical footprint.
CW-WDM MSA — Silicon photonics and CPO need external multi-wavelength lasers, but nobody had defined where the wavelengths sit.
ELSFP — A laser sitting next to the optical engine runs hot and can’t be replaced.
OpenZR+ MSA — 400ZR covers DCI and no further; everything at metro reach and beyond stayed proprietary.
OCP ESUN — Scale-up wants to run on Ethernet, but standard Ethernet was never designed for it.
UEC — Outside InfiniBand, Ethernet lacks a complete stack built for AI and HPC.
UALink — NVLink is closed; everyone else needs an open scale-up interconnect.

Optical MSA Tracker

OCI MSA — Copper can’t reach far enough for scale-up, and existing optical modules burn too much power.

XPO MSA — The faceplate is out of room for more pluggables, but the industry isn’t ready to commit to full CPO.

Open CPX MSA — Every CPO optical engine is proprietary, leaving buyers with no second source and no way to service them.

ACC-MSA — Passive copper runs out of reach; DSP-based active copper costs too much and draws too much power.

LPO MSA — The DSP inside an optical module accounts for most of its power and most of its cost.

RTLR / LRO — Fully linear is hard to close, yet the receive-side DSP is the one most worth removing.

OSFP MSA — Past 400G, modules run too hot for the existing QSFP mechanical envelope to handle.

QSFP-DD MSA — Eight lanes are needed, without forcing customers to replace every QSFP port they already own.

SFP-DD MSA — The server side needs twice the bandwidth of SFP, in the same mechanical footprint.

CW-WDM MSA — Silicon photonics and CPO need external multi-wavelength lasers, but nobody had defined where the wavelengths sit.

ELSFP — A laser sitting next to the optical engine runs hot and can’t be replaced.

OpenZR+ MSA — 400ZR covers DCI and no further; everything at metro reach and beyond stayed proprietary.

OCP ESUN — Scale-up wants to run on Ethernet, but standard Ethernet was never designed for it.

UEC — Outside InfiniBand, Ethernet lacks a complete stack built for AI and HPC.

UALink — NVLink is closed; everyone else needs an open scale-up interconnect.