A step-by-step account of why digital signal processing can't scale with the AI boom — and what we built instead.
GPUs crunch every matrix multiply electrically — electrons racing through silicon at up to 1,000 W per chip.
A 100,000-GPU AI cluster is a dense network — every node must reach every other at hundreds of gigabits per second.
Fiber optics span the datacenter — but electrons must convert to photons to enter the fiber, then back to electrons at the far end.
Chromatic dispersion stretches the optical pulse as it travels — different wavelengths arrive at different times. The signal is a mess.
Every transceiver contains a DSP that reconstructs the signal from the optical chaos. That digital processing burns 20 W — per link.
Instead of converting the distorted light into electrons just to fix it, what if the correction happened directly in the optical domain — before the photons ever become electricity?
It replaces the transceiver's DSP outright — taking the same distorted optical signal in, and putting a clean optical signal out, without ever converting to electricity. All at under 1 W.