Lightmatter's 1.6 Tbps-per-fiber sampling targets interconnect density

Lightmatter said on March 11, 2026 that its Passage co-packaged-optics chiplet had reached sampling with 1.6 terabits per second of throughput per fibre. The announcement describes a 16-wavelength architecture and evaluation kits for lead customers.

This is an interconnect development: it concerns moving data between electronic systems. It does not report the speed or energy efficiency of a neural-network computation performed in light.

Keep direction and aggregation visible

The company's Passage EVK100 page specifies 1.6 Tbps in one direction, using separate transmit and receive fibres. It separately lists 3.2 Tbps of aggregate bandwidth. These numbers describe different counting boundaries and should not be interchanged.

The page also labels its 3.2 picojoules-per-bit efficiency figure as a target. That wording matters. A target is not an independently measured wall-plug result for a complete server or cluster.

For an infrastructure reader, the useful development is a denser optical path offered for evaluation. Whether that path improves a workload depends on where communication constrains the system and how the link is integrated.

What an evaluation needs to establish

A comparable test would identify the direction of traffic, usable payload, error treatment and complete powered configuration. Laser, control and conversion energy should be accounted for explicitly before comparing energy per delivered bit.

The reviewed release does not supply an independent whole-system benchmark. It also does not establish that every interested buyer can obtain a production deployment on a particular schedule. Lead-customer sampling is the reported commercial stage.

Our assessment is that the announcement gives engineering teams a concrete component to examine, with a clear bandwidth-density question. The next useful evidence would show sustained link behaviour and a complete power boundary under a stated configuration. Until then, the per-fibre throughput and the efficiency target should remain separate entries in any comparison.