PsiQuantum shows AI data-center switch built from its quantum-computing technology

PsiQuantum has put technology developed for quantum computers into an AI networking prototype. On October 28, 2026, it disclosed an 8×8 optical circuit-switching demonstration and said it is sharing samples with selected prospective customers. The device targets connections between AI processors, routing light without first turning it into an electrical signal.

The switch addresses the work of moving data, rather than doing the AI calculation itself. Semafor describes a networking bottleneck as large groups of GPUs—the processors doing those calculations—exchange their output. Training increasingly large AI models requires coordination across many processors, making the links between them an important constraint even when the chips can calculate quickly.

Optical cables carry information as light. In the routing process described by Semafor, that information is converted into an electrical signal to direct it to its destination, then converted back into light for transmission. Those repeated conversions consume energy and produce heat. PsiQuantum says its optical circuit switch can direct the signal without that electrical detour.

The company's ambition goes beyond replacing one connection. Executive vice president Robert Soderbery told Semafor that separate server racks could effectively behave as one, with GPU-to-GPU speed and latency—the delay in transferring data—comparable to connections inside a rack. PsiQuantum also says the device could enable rack-to-rack connections producing almost no heat and consuming negligible energy. Those are prospective benefits claimed by the company.

The demonstration uses PsiQuantum's low-loss silicon photonics platform, which handles light on silicon chips. The company has spent nearly a decade developing that platform for utility-scale quantum computing. Its explanation for entering AI networking is that the optical performance requirements it pursued for fault-tolerant quantum computers exceed many of those needed for optical switching in AI data centers.

CEO Victor Peng framed the switch as a way to put that work to use now while continuing toward useful, fault-tolerant quantum computers. Not every part of the quantum platform is needed for this application: PsiQuantum says it has integrated barium titanate, or BTO, at the 300-millimeter wafer scale and uses a 64×64 BTO switch in its quantum systems. BTO is not required for this demonstration or the initial commercial products.

Purely optical switching already has a deployment precedent. Semafor identifies Google as the only major hyperscaler currently using it in an AI data center. Google's system directs light using tiny mirrors known as MEMS. That provides context for the application PsiQuantum is pursuing, rather than a head-to-head performance comparison: the new announcement concerns PsiQuantum's own 8×8 demonstration.

PsiQuantum's next stated step is commercial development. It says it is advancing the platform toward products it plans to announce in the coming months, while selected prospective customers receive demonstration samples. The distinction is important: the disclosed measurements come from company laboratory testing, while the rack-level speed and energy advantages describe what executives believe the technology could enable. Customer sampling is the current milestone, not a completed commercial deployment.