IN Brief:
- Nineteen companies have moved an earlier co-packaged-optics architecture project into an official OCP workstream.
- A roughly 300-page architecture paper targets interoperable optical AI infrastructure scaling from 72 to more than 1,024 nodes.
- The coalition expects its first formal specifications to be submitted during the fourth quarter of 2026.
Lightmatter and 18 other companies have moved an industry co-packaged-optics architecture programme into the Open Compute Project, creating a formal workstream for specifications intended to support interoperable optical AI infrastructure.
The Open Silicon Photonics for AI Systems initiative has published a roughly 300-page foundational architecture document covering Modular Hardware System and Open Rack v3 environments. The proposed framework is intended to support clusters ranging from 72 to more than 1,024 nodes while allowing hardware from multiple suppliers to operate within the same system architecture.
The coalition now consists of 19 companies, comprising nine new participants and 10 founding members. Named participants include Celestica, Dell Technologies, Flex, Foxconn Interconnect Technology, Global Unichip Corporation, Hyve Solutions, Keysight, Lightmatter, Qualcomm, and Quanta Cloud Technology.
The work is broader than defining a particular optical engine or switch. Co-packaged optics places optical conversion much closer to switching or compute silicon, forcing mechanical, electrical, thermal, optical, manufacturing, and service requirements to be considered together rather than as separate chassis-level decisions.
That architectural pressure is increasing as electrical interface rates rise. The initiative cites SerDes rates approaching 448G, where the practical reach of copper connections is reduced and the distance between a high-speed processor or switch and its optical interface becomes increasingly difficult to manage electrically.
Moving the optics into or close to the package shortens that electrical path, but the trade is not free. Optical engines, fibres, switching silicon, power delivery, cooling, and packaging have to coexist within a smaller physical envelope, while production and field-service procedures become more complicated than replacing a conventional pluggable transceiver.
Co-packaged optics changes the packaging, thermal, fibre-attachment, manufacturing, and test requirements around high-speed AI interconnects. The new OCP programme moves beyond that engineering direction and attempts to define system boundaries that multiple vendors can design against.
The architecture is intended to remain technology-agnostic. Lightmatter says it can accommodate silicon photonics, VCSEL-based approaches, and micro-LED technologies, allowing the specification work to concentrate on the wider system interfaces rather than selecting one optical implementation.
That distinction is useful because optical technologies can differ substantially in their light source, coupling method, modulation, packaging, thermal behaviour, and production test. A shared rack architecture does not eliminate those differences, but it can define where each supplier’s implementation must connect mechanically, electrically, optically, and through management interfaces.
Common boundaries reduce the amount of system engineering that has to be repeated whenever an optical component changes. Hyperscale operators and equipment manufacturers can still qualify individual devices, but they do not have to treat every supplier combination as a completely different rack architecture.
The coalition is also targeting larger scale-up domains. As accelerator clusters grow, processors have to exchange increasing volumes of data while keeping communication latency and power consumption under control. Adding more accelerators produces diminishing returns if the interconnect fabric cannot keep those devices supplied with data at sufficient speed.
Optical links address reach and bandwidth density, but their manufacturing requirements differ from copper interconnects. Fibre placement, optical alignment, photonic die attachment, laser strategy, package yield, and test have to be controlled at volume, while high-power switching and compute silicon create additional thermal constraints nearby.
Contract manufacturers and test suppliers therefore sit alongside photonics and semiconductor companies in the initiative. The finished architecture has to be manufacturable through real assembly lines and test flows rather than existing only as a logical connection diagram.
Serviceability creates another complication. Pluggable optical modules can normally be replaced independently of the switch ASIC, whereas co-packaging ties more of the optical subsystem to expensive compute or switching hardware. Fault isolation, replaceable units, fibre handling, and repair policy consequently have to be considered at architecture stage.
The OCP workstream gives those decisions a formal standards route. Lightmatter says the first specification submissions are expected during the fourth quarter of 2026, following the publication of the architecture paper.
A specification will not resolve every outstanding issue around cost, yield, laser placement, repairability, or thermal design. It can define enough common assumptions for equipment suppliers, photonics companies, semiconductor developers, test houses, and manufacturers to build compatible hardware without each creating an entirely proprietary system.
The move from a coalition paper to an OCP workstream is therefore the substantive development. Co-packaged optics is already technically viable in several forms; the harder industrial task is establishing interfaces and manufacturing practices that allow those optical technologies to be deployed across a multi-vendor AI infrastructure market.
