PsiQuantum Sends 8×8 Optical Switch Samples to AI Network Buyers

PsiQuantum is shipping samples of an 8×8 optical circuit switch to prospective customers, reporting 1.2 dB typical insertion loss, 40 dB crosstalk performance, and reconfiguration in under three milliseconds in its own laboratory tests. The demonstrator operates without optical amplification, and commercial products are expected to be announced in the coming months.

The company did not identify the customers receiving samples, and no commercial deployment has been announced. That gap separates this week's news from an operator contract: what exists today is a lab-validated demonstrator in the hands of evaluators, not a shipping product.

Where would the switch sit?

The target application is connectivity between GPUs and other computing resources inside large AI clusters. Moving data between accelerators can constrain the performance of those clusters, and vendors across the industry are looking for ways to ease that bottleneck.

Optical circuit switching establishes paths that carry light directly between ports. It avoids the optical-to-electrical conversions used in conventional electronic switching. PsiQuantum says this approach could reduce network power consumption and latency, helping accelerators spend more time computing rather than waiting on the network.

The demonstration device has no moving parts or mirrors. The company describes it as strictly non-blocking, meaning a connection can be set up between any available input and output without rearranging existing connections.

Two technical caveats frame the headline numbers. Low insertion loss helps preserve the optical power available to receivers as signals traverse the network, so 1.2 dB matters for chainability. And the reported reconfiguration time refers to changing the optical path, not to the time traffic takes to traverse an established connection. PsiQuantum says it is developing larger devices, although it has not published specifications for them.

How does a quantum company end up in AI networking?

The switch extends a silicon photonics platform developed over nearly a decade for PsiQuantum's planned fault-tolerant quantum computers into conventional computing infrastructure. The company is effectively selling a second life for components built to solve quantum scaling problems.

CEO Victor Peng framed the move as a consequence of early engineering choices. "This opportunity exists because we chose to solve the hardest scaling problems at the beginning of our roadmap to utility-scale quantum computing," Peng said in the announcement.

The photonics lineage runs deeper than the AI demonstrator. PsiQuantum separately uses a 64×64 barium titanate (BTO) switch in its quantum systems. It explained that BTO is not required for the AI networking demonstrator or its initial planned products, leaving that material as a possible route to faster future devices.

How does it compare with rivals?

PsiQuantum is not alone in chasing optical switching for AI networks. Salience Labs recently raised $125 million and unveiled a 64-port switch, following an earlier 32-port design. PsiQuantum's 8×8 demonstrator is smaller, with larger devices still in development.

Port count alone, however, does not establish relative performance. Real-world comparisons also depend on optical loss and reconfiguration speed, and PsiQuantum has published figures for both of those parameters for its demonstrator. Until independent testing exists, the competing claims sit in the realm of vendor marketing rather than measured field results.

The commercial backdrop is partly policy-driven. PsiQuantum's manufacturing work is supported by a $100 million award from the U.S. Department of Commerce under the CHIPS and Science Act, covering optical switches, photon detectors, and advanced packaging. That funding ties the company's component roadmap to a U.S. industrial policy push aimed at domestic semiconductor and photonics manufacturing.

For now, the story rests on lab numbers and unnamed evaluators. The coming months will show whether the 8×8 samples convert into announced products, and whether the larger devices in development can close the port-count gap with competitors already fielding 64-port designs.