Jun. 23 at 2:19 PM
The AI transition from training to inference is where things start to get interesting, but it’s not the obvious stuff like GPUs or even optics anymore.
Every step from 400G to 800G to 1.6T isn’t just about bandwidth going up. It quietly makes synchronization way harder. More GPUs talking to each other, more optical links, more switching points… and suddenly timing errors start to matter a lot more than people think.
Inference clusters especially are getting hit with way more timing content demand than training ever saw, sometimes 2–4x depending on system design.
Everyone’s watching the obvious names like
$AAOI,
$COHR,
$LITE… optical modules, lasers, transceivers, all that visible stuff.
But the interesting part is the layer underneath. Every time these systems scale, someone is getting paid just for keeping everything in sync.
Feels like the hidden AI tax isn’t optics at all… it’s timing.
$SITM