AIToday
AI Business & IndustrySemafor TechPublished: Oct 29, 2026, 04:00 JST

PsiQuantum unveils optical device to link GPUs across racks

Send on LINE
PsiQuantum unveils optical device to link GPUs across racks

3 Key Points

  1. What happened

    PsiQuantum announced a new device that routes data between GPUs without converting photons into electrical signals, which executive vice president Robert Soderbery said would let separate server racks behave as one.

  2. Why it matters

    Eliminating optical-to-electrical conversions could cut the heat, energy and bulk generated by optical switches in large AI data centers, addressing a networking bottleneck during training of large models.

WHO IT HITSData center network architects and AI infrastructure teams at hyperscalers and large model trainers, who currently rely on optical switches that generate heat and consume energy during GPU-to-GPU data transfer, may see an alternative for rack-to-rack connectivity.

Not sure about something? Ask the AI

Questions and answers are published on this page.

Summaries like this, every morning.

Context & Analysis

AI data centers have scaled by networking hundreds of thousands of GPUs, but the copper links between them have become the weak point, especially during training runs that demand heavy coordination. Optical cables are increasingly used, yet they still require converting light to electrical signals and back again billions of times per second, generating heat and bulk through the optical switches involved.

PsiQuantum's device sidesteps those conversions entirely, which Robert Soderbery said would let data flowing between GPUs in separate racks match the speed and latency of GPUs inside one rack. That matters because server racks are approaching limits on compute density, size and energy efficiency, so stretching performance across racks could offer a way around physical constraints.

The company is not alone in pursuing photonics for data center networking. Google already operates a purely optical circuit switching system using MEMS mirrors, while Lightmatter has pushed for standardization and Nvidia has invested billions in optics companies like Lumentum and Coherent. PsiQuantum's entry suggests the race to move data center networking from copper to light is drawing in quantum computing players as well.

FAQ
How does PsiQuantum's device differ from existing optical networking?
Existing optical systems convert light into electrical signals to route data, then back to light. PsiQuantum says its device routes photons without any electrical conversion.
Who else is working on optical networking for AI data centers?
Google uses a purely optical circuit switching system with tiny mirrors called MEMS. Lightmatter helped launch an initiative to standardize optics, and Nvidia has invested billions in optics companies like Lumentum and Coherent.
What problem does this solve for AI training?
Training large AI models requires coordination among many GPUs, but copper connections bottleneck data transfer. PsiQuantum's device could connect racks at the same speed and latency as GPUs within a single rack.
Send on LINE

AI news that matters for your work, delivered every morning.

Pick your industry and the AI tools you use, and get news related to your work every day.

Free · 30 seconds with Google · unsubscribe anytimeWhat is AIToday? →

Ask AI

Ask AI anything about this article. The AI reads this article, earlier AIToday articles, and Wikipedia, and cites its sources. Q&As are published on this page for other readers too.

Questions and answers are published on this page.

Related Articles

Next articleBroadcom up 4.2% on $63 billion Anthropic financing deal