PsiQuantum has demonstrated an 8x8 optical circuit switch in the lab that loses just 1.2 dB of light in typical tests, well under the 3 dB loss that data center optics can tolerate, according to Tom's Hardware's reporting on the company's announcement. The catch is scale: the prototype has eight inputs and outputs, against rivals shipping switches with 32 to 300 ports, though PsiQuantum says larger devices are already in development.
The device is an optical circuit switch, which passes light straight from fiber to fiber without converting it to electrical signals at each hop, unlike the packet switches that currently handle most GPU-to-GPU traffic. That conversion costs power and adds latency, and skipping it matters more as AI clusters grow and inter-GPU bandwidth becomes the bottleneck, Tom's Hardware reports. PsiQuantum's switch is built on a silicon chip etched with no moving parts, in contrast to Google's and Lumentum's designs, which use tiny mechanical mirrors, or MEMS. Tom's Hardware reports the 8x8 is "strictly non-blocking," meaning a new connection never disturbs ones already established, and every input can reach every output.
In lab testing, the switch measured a typical 1.2 dB insertion loss, meaning about a quarter of the light is lost crossing the chip, along with 40 dB of crosstalk, or roughly 0.01% of signal leaking to other ports, and under 3 ms of reconfiguration time, figures that PsiQuantum says exceed its own initial targets. That loss figure is measured fiber to fiber, including getting light on and off the chip, but not connector losses, PsiQuantum told Tom's Hardware Premium. By comparison, Google's Palomar switch, with 136 ports, typically stays under 2 dB; Lumentum's 300-port R300 sits near 1.5 dB at start of life, according to the company's switching GM speaking to Gazettabyte; and Salience Labs' 32-port OC-32M also stays under 2 dB, but relies on built-in amplification that PsiQuantum's design does not need.
The low-loss numbers matter because PsiQuantum says an eight-port device designed for conventional data center traffic already comes in well under the 3 dB loss ceiling that Nvidia specified in a request for proposals, per Tom's Hardware. The company told Semafor its switch needed insertion loss under 2 to 3 dB to be workable at all. PsiQuantum has a 64x64 version already taped out and a 72x72 underway, and the company's 2025 Nature paper reported silicon nitride waveguides losing as little as 0.5 dB per meter, with splitters losing 0.5 millidecibels and crossings losing 1.2 millidecibels, the low per-component losses that would need to hold as more ports, and more light-crossing components, are added.
Speed is a weaker point. PsiQuantum's switch reconfigures in under 3 ms, in the same class as Google's millisecond-scale MEMS switches, but Salience Labs and iPronics both claim under 300 microseconds for their amplified chip-based designs, which iPronics says it needs to rewire networks mid-workload without stalling compute. PsiQuantum does have a faster path available: barium titanate, a material already used in a 64-port switch for its quantum computing systems that switches light at gigahertz rates, roughly a million times faster than the AI demo, though the company says that material is "not required for the demonstration vehicle or initial products," which instead use a thermally switched silicon nitride platform.
Adoption outside Google remains limited. Google's switches already run in the "tens of thousands" inside its own data centers resizing TPU pods and routing around failed chips, and research firm Cignal AI, cited by Tom's Hardware, expects most deployments to stay there through the end of the decade, with the rest of the market still at "proof-of-concept and early trial." Cignal AI has nonetheless raised its market estimates roughly threefold for 2026 and by more than 40% for 2029. PsiQuantum says it has already shipped samples and testing kits to GPU vendors and hyperscalers, and Nvidia, which holds a stake in PsiQuantum through its NVentures arm, also backed competitor iPronics' $125 million funding round in September.
For architects weighing copper against optical fabric roadmaps, the number that matters is port count, not loss: PsiQuantum's physics checks out at eight ports, Salience has already shown 64 ports at this year's European Conference on Optical Communication with 128- and 256-port products on its roadmap, and nothing here is independently verified past the lab bench, so don't reprice an interconnect strategy until someone ships a 64-port monolithic device at these loss numbers.