After OCP 2026, the question for co-packaged optics is deployment.

The most useful sentence from October's OCP Global Summit in San Jose came from Mark Wade, CEO of Ayar Labs: "the bar is no longer whether it works; it's whether it can be deployed and managed at scale." He was describing a live demonstration by Wiwynn, Ayar Labs and GUC of a compute system that integrates TeraPHY optical engines and a SuperNova light source into an ORv3-ready platform. Wade's framing should be read as a vendor's, but the demo's content supports it. The focus was on operations: live telemetry from the optical engines, tuning that tracks thermal and wavelength drift, and detachable fiber attachment with in-chassis routing for serviceability.

Those are the concerns of someone who has to run this equipment. They are also where the announcement stops short. The release says key eye metrics hold steady, and gives no values. It offers no bandwidth, power or reliability figures. That gap runs through much of this season's coverage.

Broadcom has the most concrete hardware. It describes TH6-Davisson as its third-generation co-packaged Ethernet switch, with 102.4 terabits per second of optically enabled capacity at 200 gigabits per channel, double the line rate of the previous TH5-Bailly generation. Writing from OFC in April, Geoff Tate reported in Semiconductor Engineering that Broadcom and Nvidia had deployed co-packaged optics into initial production for scale-out switches and planned to use it for scale-up.

The structural change may matter more than any single product. In August, the Open Silicon Photonics for AI Systems initiative launched as an official Open Compute Project workstream, led by Lightmatter with support from Celestica, Dell, Keysight and Qualcomm, following a March announcement. Its opening white paper runs to 294 pages and sets out a shared blueprint for clusters from 72 to more than 1,000 nodes, designed to fit existing OCP rack platforms and to keep passive fiber in place as electrical rates rise. Buyers tend to resist single-vendor interconnects, and an open specification is how a technology like this becomes safe to design around.

The packaging debate is not settled either. On November 3, Tower Semiconductor and Innolight said Innolight's first near-package optics device had entered early production on Tower's silicon photonics process. Near-package designs keep the optics just outside the main package. Semiconductor Engineering lays out the trade: moving optics closer lowers power but widens the blast radius when a part fails, while near-package reduces that risk at some cost in power and performance. Pluggables are not leaving soon either. In October, Semtech and Tower agreed to expand silicon-germanium capacity across three regions, on a roadmap from today's 800G and 1.6T links toward 3.2T.

Reliability is the claim that would move the market fastest, and the public evidence is thin. Semiconductor Engineering reports Meta saying at OFC that the co-packaged optics it evaluated proved more reliable than pluggable transceivers, but that is secondhand and covers one evaluation. Until operators publish field failure rates and power per bit from production clusters, the case rests on vendor demonstrations. What changed this autumn is that those demonstrations now answer the operators' questions, and the specifications are being written in the open.