Summary
- Semtech has joined the XPO MSA as a technology partner supplying 200G-per-lane driver and TIA technology.
- The first XPO generation targets 12.8T per module and incorporates liquid cooling.
- The roadmap extends towards 400G-per-lane electrical interfaces as AI network bandwidth rises.
Semtech has joined the eXtra-dense Pluggable Optics multi-source agreement, contributing 200G-per-lane modulator-driver and transimpedance-amplifier technology for a proposed 12.8T optical module aimed at AI infrastructure.
The XPO group is developing a higher-density pluggable transceiver intended to increase the amount of optical bandwidth available through a network-switch faceplate while retaining the serviceability of a removable module.
The current generation targets 12.8Tbps per module using 200G-per-lane electrical interfaces. The specification also incorporates liquid cooling.
That cooling requirement shows how the thermal consequences of AI infrastructure are spreading beyond GPU servers. Switch ASICs and optical components are increasing in density and power consumption, pushing parts of the network layer towards the same mechanical constraints already affecting compute racks.
Semtech is contributing linear modulator drivers and transimpedance-amplifier ICs intended to support the XPO architecture. The company is demonstrating its high-speed interconnect technology during ECOC 2026 in Málaga.
The bandwidth requirement comes from the way large AI clusters operate. Thousands of accelerators have to exchange model parameters and data quickly enough that the network does not leave expensive compute idle.
Increasing switch capacity creates physical constraints. A front panel has finite space, and faster electrical signals become harder to route efficiently over longer distances inside the switch.
Higher-density optical modules are one response. Co-packaged optics is another. XPO attempts to retain a pluggable architecture while increasing bandwidth density substantially.
The multi-source format is designed to provide a common mechanical and electrical framework rather than leave each supplier to develop a proprietary module. That can broaden the component ecosystem if module makers, photonics companies, and switch suppliers adopt the specification.
XPO has also outlined a roadmap towards a future generation using 400G-per-lane electrical interfaces. That represents a further increase in signal speed and will place additional pressure on packaging, power, thermal management, and manufacturing.
Semtech cited an analyst estimate that the AI datacom optics market could reach $37 billion in 2026, up from $9 billion in 2024. That is a forecast supplied by Cignal AI rather than measured market revenue and should be treated accordingly.
The specification is still moving towards volume commercial use. MSA membership and demonstrations do not establish large-scale operator adoption, and customers will judge the format on reliability, cost, power consumption, serviceability, and compatibility with their switching platforms.
The inclusion of liquid cooling is nevertheless a useful marker. As AI infrastructure becomes denser, optical bandwidth can no longer be separated neatly from the mechanical engineering required to remove the heat generated by the network equipment carrying it.
