The Light Is On: Wiwynn, Ayar Labs and GUC Bring real “endgame” Optical I/O to a Running System

13 Oct 2026 · @Anders Storm

At Computex, optical I/O (or the more common name today - co-packaged optics CPO) next to the XPU was an architecture. At OCP Global Summit 2026 it is a running system. Three things to unpack: the missing piece in Goldman’s roadmap, a standard everyone wrote off, and a laser socket nobody wants to talk about.

In June I wrote that Goldman Sachs sees the destination for AI optics, but could not put a value on it, and that Computex 2026 may have shown us the path. The path just got confirmed.

Today on 13 October, Wiwynn, Ayar Labs and GUC announced a live demonstration at OCP Global Summit 2026: a Wiwynn compute system with a GUC ASIC package, Ayar Labs TeraPHY optical engines and SuperNova light sources, streaming real-time health telemetry. It reads like another trade-show release. It is not. I always seen Ayar labs as the best company remove the copper wall. I followed them for years and been fortune enough to get to know and talk to their CEOs over the years. I think they have a fanatics product and it is great to see the progress they are making.

Hence it is extra great to see that the bar has moved. The question is no longer whether optical I/O works. It is whether it can be operated. This release is the first public answer, and it raises three questions worth asking out loud.

The three questions this post answers:

1. What is actually new? The same partners showed a rack at Computex. Why the OCP demo is a different order of event, and the missing piece it fills in Goldman Sachs’ optical roadmap.
2. Is CW-WDM MSA dead? The common view is that OCI MSA replaced it. Why that is a category error, and what the public record says about the light source inside this demo.
3. Who is behind it, and whose lasers are used? Which hyperscaler an ODM like Wiwynn is really building for, and whether Sivers is still the laser inside SuperNova, or has quietly lost the socket?

The free preview covers the first question. The standard-layer argument, the hyperscaler detective work, the laser investigation and the timeline are below the fold.

From architecture to operation: what is new

OCP is the third public step in nine months, and each step answers a harder question. Each row below is sourced to the companies’ own announcements:

Event: OFC, March 2026

Ayar Labs and Wiwynn partnership; HVDC-enabled rack-level architecture, designed to scale to 1,024 accelerators with more than 100 Tbps of optical connectivity each

Ayar Labs and Wiwynn joint release, 11 Mar 2026

Event: Computex, May 2026

Full optical scale-up rack: 32 compute trays, 64 AI ASICs, more than 800 TB/s optical bandwidth, SuperNova light sources

Wiwynn’s Computex demonstration, as analysed in my June post

Is it deployable on paper.

Event: OCP, October 2026

Live compute system: GUC multi-chip package, TeraPHY engines and SuperNova, link up, telemetry streaming

Wiwynn, Ayar Labs and GUC joint release, 13 Oct 2026

Does it run, and can you manage it?

In the OCP release, the three companies describe this as CPO moving from architectural validation to system-level functional proof. Three details in the release back that up:

• The link is live inside the system. According to the release, telemetry is polled continuously from the optical engines inside the multi-chip packages and displayed in real time, proving the optical link is up.
• Drift is tracked, and the eye holds. The release states that as the tuning subsystem tracks thermal and wavelength drift (TX heater activity), eye height, OMA and extinction ratio hold steady alongside optical engine and ELSFP temperatures. My read: thermal drift is the classic objection to ring-based optical I/O, and this is the first time it has been shown managed inside a production-format chassis.
• It is serviceable. The release lists a detachable fiber attach unit and in-chassis fiber routing, presented as proof of serviceability and manufacturability at the compute-system level.

Ayar CEO Mark Wade put it plainly in the release: the bar is no longer whether it works, but whether it can be deployed and managed at scale. My interpretation: telemetry, drift tracking and serviceability are exactly what a hyperscaler’s operations team asks for before signing off on a new interconnect. Nobody builds that for a demo.

Closing Goldman’s open box

Goldman Sachs’ April 2026 report, Optical Networking: The Next Mega Trend in AI Infrastructure, models optical networking growing from roughly $15 billion to $154 billion by 2028. In Goldman’s numbers, about $106 billion of that is scale-up and about $91 billion comes from CPO. Almost all of the CPO modelling in the report is switch-centric.

The third box on Goldman’s roadmap, CPO with the XPU, carries a single annotation in the report: adoption schedule still unclear. In my Computex piece I argued that this was the most important sentence in the report. Goldman knew where the industry was going, but not when.

My view: OCP does not give us the date, but it removes the biggest reason the date was unknowable. Before this release, there was no public evidence that optical I/O at the XPU could run stably inside a production-format system. The OCP telemetry data is that evidence.

It also firms up the light-source arithmetic. In June, my reading of Wiwynn’s Computex system diagrams was up to 16 SuperNova assemblies per compute tray, which across 32 trays implies as many as 512 per rack. Goldman’s report assumes roughly 30 to 60 external light sources per rack for switch-centric CPO. If those figures are comparable, compute-side optical I/O carries 8 to 15 times the light-source density of switch CPO. Because it replaces copper rather than pluggables, my conclusion is that it is additive to Goldman’s TAM, not a share shift within it.

A working system does not prove the multiplier. It does prove the architecture the multiplier depends on is real.