On October 28, 2026, PsiQuantum unveiled a demonstrator of an optical circuit switch with eight inputs and eight outputs. The device switches connections while keeping signals as light, without converting them to electrical signals. It applies silicon photonics technology that the company has spent roughly a decade developing for quantum computers, now extended to AI networking. According to PsiQuantum, in-house testing of the demonstrator achieved a typical insertion loss of 1.2 dB. The company has already provided samples to some prospective customers and plans to announce a commercial product within the next few months. However, these measurements alone cannot show how much AI network performance as a whole would improve. The test conditions and validation in real operating environments will matter.
What Does a Typical Insertion Loss of 1.2 dB Mean?
PsiQuantum's figures come from in-house testing of the 8×8 optical circuit switch demonstrator, not from simulation-based estimates. Insertion loss measures how much optical power is lost as light passes through a device; the smaller the value, the more light remains at the output. Crosstalk, by contrast, refers to signal leaking onto unintended optical paths, and is used to evaluate interference between paths. Besides insertion loss and crosstalk, the announcement also gave the time needed to switch connection paths.
| Disclosed item | Value | What was measured / caveats |
|---|---|---|
| Typical insertion loss | 1.2 dB | Typical value from in-house testing of the 8×8 demonstrator |
| Lower-end measured value | 1 dB | A lower value reported for the same demonstrator; not a typical or guaranteed value |
| Crosstalk | 40 dB | Typical value from in-house testing of the 8×8 demonstrator; sign convention and measurement basis not disclosed |
| Path reconfiguration time | Under 3 ms | Typical value from in-house testing of the 8×8 demonstrator; the time needed to switch connection paths |
The 1 dB and 1.2 dB insertion loss figures are not a before-and-after comparison of improvements. The 40 dB crosstalk figure is used exactly as PsiQuantum reported it; because no reference signal or measurement method was given, it cannot be reinterpreted as a different metric by adding a sign. The press release and product page also do not state conditions such as the number of devices tested, the number of measurements, the wavelength of light used, polarization, or temperature. The distribution and variation of measured values across all connection paths, as well as confidence intervals, are also unknown. In addition, these materials do not point to a peer-reviewed paper or preprint reporting the 8×8 demonstrator's results, or to independent third-party replication or measurement. It is therefore not possible at this point to judge whether the reported typical values would hold in other usage environments or in larger devices.
Switching Network Paths While Keeping Signals as Light
An optical circuit switch routes incoming optical signals to the desired output port without converting them to electrical signals. The demonstrator has eight input ports and eight output ports and can switch the connection paths of optical signals. PsiQuantum describes this configuration as "non-blocking," meaning each input can be connected to any free output without being obstructed by other connections. It does not mean that all eight inputs can be connected to the same output at once.
According to PsiQuantum, the demonstrator uses no moving parts, mirrors, or optical amplifiers. It also needs no step to convert optical signals to electrical ones, though that does not eliminate loss as light passes through the device. The typical insertion loss of 1.2 dB evaluates exactly this kind of optical power loss. The under-3 ms figure, meanwhile, is the time required to reconfigure connection paths; it does not indicate data propagation delay or GPU response time. Nor is it a measurement of AI model training time. Switch reconfiguration speed, overall network communication performance, and the computing performance of an AI system need to be considered separately.
Extending Quantum Computing Technology to AI Networks
What stands out in this announcement is that PsiQuantum aims to use manufacturing technology developed for quantum computers in optical components for AI networks as well. The company is seeking to extend the silicon photonics manufacturing base it built while developing quantum computers to another application. However, using technology developed for quantum computers is not the same as having demonstrated the performance of an AI networking switch. In particular, where manufacturing technology and device scale differ, measurement results from one device cannot simply be applied to another.
PsiQuantum says it has developed technology to integrate barium titanate (BTO) at 300 mm wafer scale, and that its quantum systems use 64×64 BTO switches. By contrast, the 8×8 demonstrator unveiled now and the initial commercial products do not need to use BTO. In other words, the demonstrator and the 64×64 switches for quantum systems are not the same configuration. The 1.2 dB typical insertion loss and under-3 ms reconfiguration time obtained with the 8×8 demonstrator cannot be treated as the performance of the 64×64 switches.
On the manufacturing side, PsiQuantum stresses that it has a volume manufacturing base using GlobalFoundries, and it is also developing larger optical circuit switches. However, this announcement concerns the performance of the 8×8 demonstrator and sample provision; volume shipments of a commercial product have not begun. Specific specifications, such as port counts, for the commercial product to be announced have also not been disclosed. Providing samples to prospective customers and offering a commercial product with finalized specifications, pricing, and supply terms are separate stages.
Switching paths without converting optical signals to electrical ones could reduce the power consumed by communications. However, these materials contain no measurements of overall data center power consumption, GPU utilization, communication bandwidth, or AI model training performance. What was presented are in-house test results for the optical circuit switch alone, not an experiment that directly verified effects in an AI system or a correlation analysis based on real operating data. How much a component's low loss would contribute to power savings or performance gains across an AI network has therefore not yet been demonstrated.
TSN also points out that the announcement did not disclose pricing, customer names, or shipping dates, and cites balancing larger port counts with low loss and price competitiveness as a challenge for commercialization. Going forward, it will be necessary to see detailed measurement conditions for larger switches and independent third-party performance evaluations, along with how much benefit is gained under real AI communication workloads. Once conditions such as product specifications, pricing, and supply timing are clarified, it will be possible to evaluate more concretely how much practical value PsiQuantum's low-loss optical circuit switch would bring to AI data center networks.