PsiQuantum’s optical switch aims at a growing pressure point in AI infrastructure: moving data between GPUs without spending power and time converting optical signals to electrical ones at every network hop. Its 8×8 prototype recorded typical signal loss of 1.2 dB in lab tests, below the 2–3 dB range the company says is workable for data center links. The port count is small, but the design’s low loss without optical amplification gives PsiQuantum a distinct starting point as it develops larger devices.
Less signal loss, fewer conversions
The optical circuit switch, or OCS, routes light directly from one fiber to another. Conventional packet switches convert optical signals into electrical form for processing, then convert them back to light. That adds power use and latency at each hop. An OCS sets a light path and keeps it in place until the network is reconfigured.
That approach is useful in large GPU clusters, where data movement between accelerators can constrain performance. Optical links can carry high bandwidth over longer distances than copper, making them a candidate for connections between GPU racks. But every optical switch reduces the signal reaching the next device, so insertion loss sets a practical limit on how many components a link can tolerate.
PsiQuantum says its silicon photonics design achieves 1.2 dB of typical fiber-to-fiber insertion loss without an amplifier. That corresponds to roughly 24% of the light being lost. The measurement includes coupling light onto and off the chip, but excludes connector losses, the company told Tom’s Hardware Premium. The demo also recorded 40 dB of crosstalk, or about 0.01% of the signal leaking to other ports, and reconfiguration in under 3 milliseconds.
The company says those results exceed its initial targets. Its switch uses circuits etched onto a silicon chip and has no moving parts. That sets it apart from mirror-based MEMS designs, which redirect light using tiny mechanical mirrors.
Eight ports are a starting point
The prototype has eight inputs and eight outputs, with every input able to connect to every output. PsiQuantum describes it as strictly non-blocking: setting up a new connection doesn’t interrupt connections already in place. That flexibility is useful, but the device’s capacity trails several competing products, which offer 32 to 300 ports.
Switch | Ports (in × out) | Insertion loss, and what the figure is | Reconfiguration | How it switches |
|---|---|---|---|---|
PsiQuantum 8×8 demo | 8×8 | 1.2 dB (about 24% of the light lost), typical in the company’s lab tests | Under 3 ms | Silicon photonics chip, no amplification, no moving parts |
Google Palomar | 136×136 | Typically under 2 dB (under 37% lost) across all port pairings; 2 dB worst case | Milliseconds | MEMS mirrors (moving parts) |
Lumentum R300 | 300×300 | Close to 1.5 dB (about 29% lost), typical at start of life, Lumentum’s switching GM told Gazettabyte | Not stated | MEMS mirrors (moving parts) |
Salience Labs OC-32M | 32×32 | Under 2 dB (under 37% lost) port to port, a ceiling, after built-in amplification | Under 300 µs | Silicon photonics chip with amplification |
iPronics ONE-32 | 32 (layout not stated) | Not published | Under 300 µs | Silicon photonics chip with optical amplifiers |
Google’s Palomar switch has 136 ports and typically stays below 2 dB, with 2 dB reported as its worst case across port pairings. Lumentum’s R300 has 300 ports and close to 1.5 dB typical loss at the start of its life, according to the company’s switching general manager, as reported by Gazettabyte. Salience Labs’ 32-port switch claims less than 2 dB after built-in amplification; iPronics also uses optical amplifiers in its 32-port product.
PsiQuantum has a 64×64 version taped out and says a 72×72 device is in progress, with larger designs also under development. As the port count rises, light passes through more components and losses accumulate. The company says losses may increase slightly in larger products while remaining low, but hasn’t shared detailed performance figures for those designs.
The photonics platform behind the prototype
PsiQuantum’s low-loss results draw on work originally aimed at quantum computing, where lost photons can undermine a system’s operation. In a 2025 Nature paper, the company reported silicon nitride waveguide losses as low as 0.5 dB per meter. It also reported losses of 0.5 millidecibels for splitters and 1.2 millidecibels for crossings—small amounts that matter when many components are connected.
The company has also published a best lab result of about 60 millidecibels for chip-to-fiber coupling, equivalent to roughly 1.4% light loss. That measurement was made without a full fiber array attached, so it represents a best case rather than a complete estimate for a larger switch.
The 8×8 device operates in the O-band, the wavelength range used by common data center optical modules. PsiQuantum says its prototype was designed for conventional data center traffic and beats a 3 dB loss target in an Nvidia request for proposals. Coherent has said OCS links can add up to 3 dB of insertion loss while remaining compatible with its transceivers. Lumentum has also said that 3 dB can work with standard FR4, DR4 and DR8 modules used to connect GPU servers and switches over distances of 500 meters to 2 kilometers.
Reconfiguration speed has room to improve
For current OCS deployments, loss and port capacity matter more than switching speed. Google uses its switches to resize TPU pods and route around failed chips; its MEMS devices operate on a millisecond timescale. PsiQuantum’s sub-3-millisecond result fits that use case. Salience and iPronics, by comparison, claim reconfiguration below 300 microseconds, which can help operators change network paths while GPUs continue computing.
PsiQuantum has another route to faster switching through barium titanate, or BTO. Applying voltage to BTO changes how it bends light. The company chose the material over lithium niobate because it produced a stronger effect in its tests, allowing shorter switching elements and lower losses. PsiQuantum already uses BTO in a 64-port switch for its quantum systems, which need switching at gigahertz rates. The material isn’t part of the data center demo or the company’s initial products, which use a thermally switched silicon nitride platform, but PsiQuantum calls it a possible path to faster future devices.
Commercial scale is the test ahead
OCS technology is established inside Google, which has deployed tens of thousands of switches. Outside Google, adoption remains at the proof-of-concept and early-trial stage, according to Cignal AI. The research firm raised its market estimates earlier this year, putting its 2026 estimate at roughly three times its previous level and its 2029 estimate more than 40% higher. Even so, broader use in GPU networks is expected to take time.
PsiQuantum says it has sent samples and testing kits to several computing companies, including GPU vendors and hyperscalers. It expects to announce commercial products in the coming months, alongside updates on larger devices. Salience has already demonstrated a 64-port switch and has outlined 128- and 256-port versions.
There’s a business logic to PsiQuantum entering data center networking: producing switches at scale could help the company build experience in packaging, assembly and supply chains needed for its quantum systems. The company calls AI and quantum computing complementary technologies. Nvidia has a relationship with PsiQuantum through its venture arm, NVentures, and works with the company on GPU-QPU integration and silicon photonics. Nvidia also joined iPronics’ $125 million funding round in September to develop that company’s second-generation optical switch.
PsiQuantum co-founder Pete Shadbolt has praised MEMS switches while arguing for similar performance from a monolithic chip without moving parts. The 8×8 demo offers an early measure of that ambition, but the engineering case will depend on how its losses and reliability hold up as port counts climb. If the company can preserve its low-loss results in larger devices without amplification, it could give GPU operators another way to connect racks while keeping optical signals on the path.