PsiQuantum Optical Circuit Switch: What the AI Data Center Demo Shows and What Operators Would Monitor

What PsiQuantum announced

PsiQuantum's press release of 28 October 2026 calls the device its first publicly disclosed optical circuit switch (OCS): an 8x8 radix demonstration vehicle. It runs on the low-loss silicon photonics platform the company has spent nearly a decade building for its quantum computers. PsiQuantum's OCS product page adds that the switch is strictly non-blocking, so every input can reach every output, and that it has no moving parts.

All of the performance figures come from PsiQuantum. The press release says tests in PsiQuantum's labs showed typical results of 1.2 dB insertion loss, 40 dB crosstalk and less than 3 ms reconfiguration time, without amplification. The company is sharing samples with select prospective customers and says commercial products will be announced in the coming months. The Quantum Insider carried the release the same day.

The release also says PsiQuantum uses a 64x64 switch built with barium titanate (BTO) inside its quantum systems, and that BTO is not needed for the demonstration unit or the first products. Quantum Computing Report, in its 28 October 2026 analysis, notes that winning volume in AI clusters would mean scaling well past 8x8 ports without losing the loss budget during automated fiber packaging.

What an optical circuit switch does

In a typical AI cluster, light from a transceiver reaches an electrical packet switch. The switch turns it into an electrical signal, reads where each packet should go and sends it out as light again. PsiQuantum's OCS page describes the other approach. An optical circuit switch passes light from one port to another with no conversion at all.

The switch never looks at packets. It holds a fixed path, a circuit, between two ports until it is told to change that path. Changing it is a reconfiguration, which is why PsiQuantum publishes a reconfiguration time. Every switch in the path also loses a little light. PsiQuantum's page notes that this loss adds up across a network, so insertion loss is the number the company leads with.

Why AI cluster networks look at OCS

Google is the public example. Its TPU v4 paper, submitted to arXiv on 4 April 2023 for the ISCA 2023 conference, describes a supercomputer of 4,096 chips whose optical circuit switches reconfigure the interconnect topology, in service since 2020. The paper says the switches and their optical components are under 5% of system cost.

Google's Cloud blog post on the paper, dated 6 April 2023, says those switches use MEMS mirror arrays. It also says circuit switching makes it easy to route around failed components, so a training job can use thousands of processors for weeks.

Google built its own switches with mirrors. PsiQuantum's page describes a silicon chip with no mirrors and no moving parts. Whether that approach holds up in a production GPU fabric, at a useful port count, is still open.

What the demo shows and what it does not

The demo shows that PsiQuantum can make an 8x8 switch on its photonics platform and measure typical loss, crosstalk and switching time that it calls better than its initial targets.

The release names no customer and no network equipment partner. The figures are typical lab values, and neither the release nor the coverage we read reports an independent test. There is no data sheet, no price, no port count for a product, and no report of the switch carrying traffic in a working AI data center. Until those arrive, operators can follow it without planning around it.

What operations teams would monitor

If OCS does reach a GPU fabric, from PsiQuantum or anyone else, the monitoring changes in a few ways. The switch reads no packets, so it keeps no packet counters. Trouble on a circuit shows up at the two ends, as errors, packet loss or a port going down on the switches and network cards that the circuit connects.

Start with link loss. Each circuit has a loss budget, and every switch, connector and splice uses part of it. A dirty connector or a worn port can push a link over budget, so teams would track loss per circuit and the receive levels on the transceivers at each end.

Every reconfiguration also needs a record: the time, the ports involved and what asked for the change. When a training job slows down a minute after a reconfiguration, that record is the first lead.

Transceiver temperature and error counters show which end is failing, and the asset record has to say which fiber runs to which switch port, or no one can trace a fault.

What the operator sees in Sensaka

Sensaka for AI Data Centers monitors and does not control network hardware. It puts GPU utilization, network throughput, packet loss, errors and port state, storage throughput and node health on one timeline, collected from the switches and interfaces in the cluster. Which indicators are available depends on the network equipment and the telemetry it exposes. Those end point counters are where a failing optical circuit would appear.

AI workload operations ties each training job to the GPUs and nodes it runs on and shows network throughput on the same timeline as GPU activity. NVLink errors and GPUs slower than their peers are flagged. If an OCS controller logs its reconfigurations, the team can line those times up with the job view and see whether a slowdown followed a path change.

The AI infrastructure CMDB keeps one inventory for servers, GPUs, networks and storage, with rack and rack unit relationships, which is where port and fiber records belong. When a port keeps erroring, the alert becomes a ticket, and through CMDB and ITSM integration it syncs with ServiceNow or Jira.

Takeaway

PsiQuantum has shown a small, low-loss switch and published its own lab numbers. Google's TPU v4 shows that optical circuit switching can work at cluster scale. Before any OCS goes in, operators should know their loss budgets, log every reconfiguration and be able to tell when network waits are cutting GPU utilization.

Frequently asked questions