Japanese consortium advances optical-chiplet manufacturing for AI infrastructure

A NEDO-selected project brings device fabrication, testing and packaging into one pilot-production effort, with NTT acting as adviser.

A Japanese consortium has been selected for a NEDO-supported project to develop manufacturing technology for optical chiplets used in next-generation AI computing. The [20 October announcement](https://group.ntt/en/newsrelease/2026/10/20/261020a.html) names NTT Innovative Devices, Japan Semiconductor, Furukawa Electric and Shinko Electric Industries as participants, with NTT advising.

The project spans photonic devices, silicon photonics, testing and packaging. It will establish and evaluate a pilot production line, targeting more than a tenfold increase in manufacturing productivity compared with research-stage processes. That is a development objective, rather than an announced production result.

Moving data is part of the AI workload

An AI system spends energy moving information as well as calculating with it. Processing chips exchange data with memory, other chips and connected systems. As computing grows, the links between those components become an increasingly important part of the overall design.

Optical interconnects use light for transmission. Bringing photonic and electronic components together can provide another way to handle high-capacity links, but a promising device still has to be made, connected, tested and packaged in a repeatable way.

That is the industrial problem this project addresses. A laboratory process can demonstrate that a component works while still requiring too much time, manual alignment or variation for commercial manufacturing. Production engineering seeks to make the result consistent across many devices.

The manufacturing chain has several specialists

The participating companies divide the work across optical-chiplet manufacturing and test methods, silicon-photonics production, compound-semiconductor optical devices and packaging processes. The announcement also includes inspection during manufacturing.

This distribution reflects the nature of a chiplet system. Its final behaviour depends on several components and interfaces, not simply the performance of one material. A useful manufacturing process must preserve those relationships as parts move through the production chain.

Testing is particularly important at those interfaces. Finding a problem before expensive assembly can change the economics of production. Packaging must then maintain the necessary electrical, optical and mechanical connections while fitting the larger computing system.

A pilot line turns a design into a production question

A pilot line provides a place to evaluate the combined process before broad commercial manufacturing. It allows the consortium to examine throughput, repeatability, inspection and the interaction between stages.

The productivity target gives the project a measurable industrial direction. The relevant comparison is with today's research-stage process, not with an already mature high-volume product. That helps explain both the ambition and the current stage.

For AI-infrastructure developers, the significance is a supply-chain investment around a possible next generation of interconnects. The announcement does not introduce a chip that operators can immediately buy and install. It advances the manufacturing path that would have to exist before such components could be supplied reliably at scale.

The project therefore adds another dimension to AI hardware development: making the connections between processors more practical to manufacture, alongside the continuing work to make the processors themselves faster.

Sources

- [Accelerating Development of Mass-Production Technologies for Photonics-Electronics Convergence Devices](https://group.ntt/en/newsrelease/2026/10/20/261020a.html) — NTT