Updated
Updated · Semafor · Oct 28
PsiQuantum Unveils Optical Switch for AI Data Centers, Cutting Billions of Signal Conversions
Updated
Updated · Semafor · Oct 28

PsiQuantum Unveils Optical Switch for AI Data Centers, Cutting Billions of Signal Conversions

3 articles · Updated · Semafor · Oct 28

Summary

  • PsiQuantum is showing customers a prototype optical circuit switch that routes GPU traffic across AI data centers without converting light signals into electrical ones.
  • That targets a key bottleneck in AI training: copper links and repeated optical-electrical-optical conversions slow communication between large GPU clusters and add heat, power use and bulk.
  • Robert Soderbery said the device could make separate server racks behave like one system, letting GPUs in different racks exchange data with the speed and latency of a single rack.
  • Google is the only major hyperscaler using purely optical circuit switching in AI data centers today, while rivals including Lightmatter, Nvidia, Lumentum and Coherent are pushing broader photonics adoption.

Insights

If GPUs are no longer the main limit in AI clusters, could optical circuit switches become the next must-have data-center technology?
PsiQuantum’s prototype avoids optical-electrical conversions—but can it scale from 8×8 demo to hyperscale AI deployment fast enough to matter?