The next AI bottleneck is shifting from compute performance toward the physical systems required to move, package, connect, and deploy photons at scale.
TL;DR
The next bottleneck in AI is no longer compute. It is the physical infrastructure required to move, package, integrate, and manufacture photonics at scale.
Investors remain focused on device performance, but history suggests the larger opportunity may sit closer to packaging, manufacturing, portability, qualification, and ecosystem orchestration.
The winners may not be the companies with the best demos. They may be the companies that make deployment possible.
=========================================
AI’s next bottleneck is no longer compute. It’s in the physical systems that move, package, and sense photons.
The same bottleneck is about to break two multi-billion-dollar markets at the same time: AI data-center optics and LiDAR-class sensing.
Everyone is still staring at GPUs, memory, and “AI chips.”
That’s already a lagging indicator.
The real constraint is at the layer below, in the physical infrastructure that moves and senses photons.
In my recent free Substack,
“Why Photonics Scaling Is an Orchestration Challenge — Not a Capacity Crunch”
(https://open.substack.com/pub/pratimanagement/p/why-photonics-scaling-is-an-orchestration?utm_campaign=post-expanded-share&utm_medium=web),
I laid out why the bottleneck is shifting from compute → memory → networking → optics, and why this is fundamentally an orchestration problem, not a simple capacity issue.
This piece picks up where that left off.
Because once you accept that framing, the next question is unavoidable:
Where does the system actually break?
In the data center, this shows up as co-packaged and near-package optics hitting power and density limits, and optical fabrics struggling to scale to tens of thousands of links per cluster.
At the edge, it shows up as LiDAR and sensing stacks trying to move from demo fleets to millions of units — with autonomy programs discovering that “cool sensor” is not the same as “manufacturable platform.”
On paper, these look like separate markets.
In practice, they share the same failure mode:
The same III–V materials
The same packaging and test bottlenecks
The same underlying question almost nobody is pricing correctly:
Can this be manufactured — reliably, repeatedly, and at scale?
Most commentary still treats this as a capacity story: add tools, add wafers, add fabs.
That’s the wrong lens.
This is not a supply problem.
It’s a synchronization problem.
And in synchronization problems, the money rarely ends up where the slide decks say it will.
