Editing Openai/69109937-2170-8005-9304-da49840bb5b9 (section)

==== What they are doing together ====
* Celestial AI co-developed a custom Optical Interposer platform with POET.
* POET supplies Starlight light engines, which Celestial uses as precision optical power sources for its Orion AI accelerator / Photonic Fabric. YouTube<ref>{{cite web|title=YouTube|url=https://www.poet-technologies.com/news/2022-feb-10|publisher=poet-technologies.com|access-date=2025-11-10}}</ref>

The architecture (simplified):

: 

Why coupling efficiency directly hits system-level AI metrics

Celestial’s value proposition isn’t “we sell transceivers”, it’s:
* Higher effective bandwidth between AI chips than electrical links can provide,
* At lower energy per bit.

If POET’s light engines have poor coupling:
* You need more lasers / more laser power to hit the same on-chip optical power → higher TDP for the AI box.
* Fan-out to multiple chips/tiles consumes more optical power on splitters and interfaces; any extra insertion loss at Starlight eats into what’s left for the actual compute fabric.
* System designers may be forced to shorten reach, reduce radix, or increase cooling — undermining the “better than copper” story.

That’s why Celestial’s CEO talks about “precision optical power sources in a highly integrated form factor”: the “precision” part is both power stability and predictable, low insertion loss from source into fabric. semiconductor-today.com<ref>{{cite web|title=semiconductor-today.com|url=https://www.semiconductor-today.com/news_items/2022/feb/poet-110222.shtml|publisher=semiconductor-today.com|access-date=2025-11-10}}</ref>

So in this relationship, coupling efficiency is a hidden but crucial factor in the AI system’s total flop/€ and flop/W. It’s not just a module spec; it’s part of whether optical fabrics actually beat Nvidia’s internal electrical fabrics on real workloads.