A liquid-cooled multi-terabit pluggable class that reaches co-packaged densities without giving up field replaceability.
What XPO Is
XPO — eXtra-dense Pluggable Optics — is a module class defined to answer the strongest objection to co-packaged optics: operators do not want to lose the ability to pull a failed optic out of a running switch. It keeps a pluggable, field-replaceable, lever-latched module, and buys density by making that module very much larger and very much more capable than a faceplate cage has ever held.
An MSA was announced on 11 March 2026 around an Arista design, with 45 founding members, passing a hundred ecosystem companies by early April 2026. The first-generation module carries 12.8 Tb/s.
How It Reaches 12.8 Tb/s
The module holds two 32-channel paddle cards mounted belly to belly, giving 64 electrical lanes at 200G per lane. Eight 1.6T optical interfaces exit on MPO-16 connectors at the front. The envelope is roughly 60.8 by 111.8 by 21.3 millimetres, a different order of object from an OSFP but still an object with a handle on it.
Power comes from the rack bus at 46 to 53 volts rather than from the host board's 3.3 V rail, converted locally on each paddle card, and the module is designed for 400 watts and beyond. Cooling is liquid: an integrated cold plate between the two cards, fed through blind-mate quick disconnects at roughly a quarter to seven tenths of a litre per minute, with the drivers, TIAs and DSPs facing inward toward it.
What It Trades Against CPO and Pluggables
Against ordinary pluggables the argument is density and power delivery. One XPO module replaces eight 1.6T OSFP-class modules, so the front panel stops being the constraint and a rack unit can present far more capacity than a row of cages allows. The short, controlled electrical channel inside the module also relieves some of the DSP burden that a long host trace imposes — the same problem LPO attacks from the other direction.
Against CPO the argument is the one operators keep making: a failed co-packaged engine is a switch-level repair, with a spare chassis, a maintenance window and a fabric designed to lose a node. An XPO module is unlatched and swapped. That is the whole pitch, and it makes XPO a competitor to CPO rather than a compromise with it. The four-way comparison places it against the others.
What to Watch Before Committing
Liquid cooling is the honest cost. A rack that carries coolant to every optic needs manifolds, leak detection, a CDU and a service procedure that most enterprise data halls do not have. That is a facilities project, not a procurement decision, and it should be evaluated as one.
The second question is fibre. Eight MPO-16 connectors per module concentrates an enormous number of fibres into a small space, so cleaning discipline, bend management and patching layout all matter more than they did. The third is maturity: the MSA is new, first silicon is first silicon, and the wider 1.6T architecture question is not settled.
XPO as defined by the MSA
| Name | eXtra-dense Pluggable Optics |
|---|---|
| Announced | 11 March 2026, formed around an Arista design |
| Capacity per module | 12.8 Tb/s — the equivalent of eight 1.6T optics |
| Electrical lanes | 64 at 200G, on two 32-channel paddle cards, belly to belly |
| Optical interface | MPO-16 connectors |
| Module envelope | roughly 60.8 × 111.8 × 21.3 mm |
| Power | 400 W and above, taken from a 46 to 53 V rack bus |
| Cooling | integrated liquid cold plate, roughly 0.35 to 0.7 litres per minute |
| Roadmap | 25.6 Tb/s at 400G per lane |
Related Terms
NPO is the near-packaged middle ground, CPO is the destination XPO argues against, and the MSA entry explains what an announcement like this one does and does not guarantee.