The most surprising entrant in the optical-switching race did not set out to build a switch at all. PsiQuantum has spent nearly a decade building a low-loss silicon photonics platform for a fault-tolerant quantum computer that does not yet exist. On October 28, 2026, the company showed what that platform does when pointed at a nearer customer: the networks inside AI data centers. Its first publicly disclosed optical circuit switch is an 8×8 demonstration chip, and the numbers the company reports for it — 1.2 dB of typical insertion loss, 40 dB of crosstalk, reconfiguration in under 3 milliseconds, all without optical amplification — are the reason the announcement deserves attention. The port count is the reason it deserves restraint. This is a platform finding a nearer market, not a quantum-computing result: nothing here advances PsiQuantum’s quantum computer directly.
What the device is, plainly
An optical circuit switch, or OCS, establishes a direct optical path between two fiber ports. Light enters on one fiber, is steered through the chip, and exits on another fiber, without ever being converted into an electrical signal inside the switch. That is the contrast with the packet switches that carry most data-center traffic, which receive light, convert it to electricity, decide where each packet goes, and convert it back to light. The conversion costs power and latency at every step. A circuit switch does not make per-packet decisions; it holds a connection open for as long as the network needs it, then reconfigures.
PsiQuantum’s demonstration vehicle has eight inputs and eight outputs. The company describes it as strictly non-blocking, meaning any input can reach any output, and it has no moving parts — no mirrors tilting to steer a beam, as in the micro-electromechanical systems (MEMS) switches that dominate optical circuit switching today. The paths are set on a silicon chip, in waveguides etched into the same low-loss platform PsiQuantum developed to carry single photons through a quantum processor without losing them.
Three numbers define the device. Insertion loss, at 1.2 dB typical, fiber-to-fiber, is the light lost passing through the switch — roughly a quarter of it. That sounds large until set against the link budget: standard data-center transceivers are generally designed to tolerate around 3 dB of additional loss from a switch, and reporting on the announcement says PsiQuantum’s figure lands well under a 3 dB target set in a request for proposals from Nvidia. The loss matters because it determines whether a switch can be inserted into an existing link without amplification — extra hardware that adds power, noise, and cost. PsiQuantum reports no amplification was required. Crosstalk, at 40 dB, is leakage between paths: roughly one ten-thousandth of the light in one connection bleeding into another. Reconfiguration time, under 3 milliseconds, is how long the switch takes to move from one topology to another — fast enough to rewire a cluster between training jobs, far too slow to switch packet by packet, which is precisely why circuit switching complements packet fabrics rather than replacing them.
One distinction matters more than the headline numbers. The release headline claims loss “as low as 1 dB” — a best case. The typical figure, the one a customer should expect across ports, is 1.2 dB. It is worth holding the company to the typical number.
Why a quantum platform over-delivers here
The argument PsiQuantum is making, in effect, is that it solved a harder problem first. A fault-tolerant photonic quantum computer is an unforgiving customer for optical loss. Its qubits are single photons, and a photon lost in transit is a computation corrupted. Keeping loss low across thousands of components on a chip, manufactured at wafer scale, is not an optional refinement; it is the precondition for the machine to work at all. An AI data-center switch, by contrast, sends classical light through a handful of components, and can tolerate a few decibels of loss before the link budget breaks.
If that framing holds, a platform built for the harder problem arrives over-specified for the nearer one. PsiQuantum’s chief executive, Victor Peng — the former Xilinx chief executive who joined PsiQuantum after AMD’s acquisition of Xilinx — put it this way in the announcement: “This opportunity exists because we chose to solve the hardest scaling problems at the beginning of our roadmap to utility-scale quantum computing.” Executive vice president Rob Soderbery has pitched the longer vision: rack-to-rack links with GPU-to-GPU speed and latency comparable to connections inside a rack, and eventually light carried all the way to the GPU itself.
The over-specification argument is plausible, but it is also, in part, marketing. It holds for loss, where quantum computing genuinely demands more than data-center switching. It does not automatically hold for everything else a switch must do. A quantum processor’s photonics are optimized for single photons in a controlled environment, at volumes that are still small. A data-center switch must handle high optical power at the O-band wavelengths used by standard pluggable modules, survive in a rack for years, be manufactured in volume at a price a network operator will pay, and be controlled by software that fits into existing network management. None of those requirements is answered by low loss alone.
The incumbent: Google’s mirrors, at scale
The reference point for any optical circuit switch in a data center is Google. Google is the only major hyperscaler currently using optical circuit switching in an AI data center, and it has done so for years. Its Apollo switches, the optical layer of the Jupiter network, steer light with MEMS mirrors: signals arrive through fiber collimator arrays with 136 individual fibers, bounce off two sets of MEMS mirror arrays, and leave on another fiber. Google’s own account of the Jupiter evolution, published in 2022, credits the optical layer with helping deliver a fivefold increase in network speed and capacity, a 30 percent reduction in capital expenditure, and a 41 percent reduction in power consumption while serving live production traffic.
The scale gap is the central fact of this story. Google’s switches have 136 ports each, deployed in the tens of thousands. PsiQuantum’s chip has eight. A GPU cluster needs fabrics that rewire hundreds or thousands of endpoints; an 8×8 device cannot be dropped into that fabric, however good its loss figure. It can only demonstrate that the underlying photonics are capable of the job.
That scaling story, as reported, is specific. According to coverage of the announcement attributed to Tom’s Hardware, a 64×64 version of the switch has already taped out, and a 72×72 design is underway, with PsiQuantum conceding that loss will rise somewhat on the larger parts. That concession is important and honest: loss in a switch grows with port count, because larger switches mean longer paths, more waveguide crossings, and more elements in series. The question is not whether 1.2 dB survives to 64 ports — it almost certainly does not, exactly — but whether the loss at 64 or 72 ports stays inside the 2 to 3 dB window practical deployment requires.
The manufacturing story is less speculative than the scaling story. The chips come out of GlobalFoundries in New York, the same supply chain behind PsiQuantum’s quantum photonic wafers, and the initial products use thermally tuned silicon nitride — a mature, foundry-compatible approach in which heaters adjust the optical path. That matters because the switch does not depend on exotic processes to reach the market; it rides the same 300-millimeter wafer infrastructure the company has qualified for its quantum program.
The faster route, held in reserve
PsiQuantum has a second switching technology, and the announcement is careful to keep it separate from this device. The company has integrated barium titanate, or BTO, at 300-millimeter wafer scale, and uses a 64×64 BTO switch inside its quantum systems. BTO is an electro-optic material: applying a voltage changes its refractive index, steering light far faster than heating a waveguide can. It is the route to faster future products, and, in Soderbery’s telling, to the long-term vision of light carried all the way to the GPU.
But BTO is explicitly not required for the demonstration vehicle or for the initial products. The demo is a thermally tuned silicon nitride device, and that distinction should be kept sharp. Thermal tuning is slower and consumes more power per switch element than electro-optic tuning, but it is simpler and already manufacturable. BTO is the reserve: a faster technology the company already operates in its quantum systems, available when the market demands speeds thermal tuning cannot reach.
The crowded field it enters
PsiQuantum is not the first company to see AI data-center networking as an optical circuit switching opportunity. It is, however, the first to arrive from quantum computing. Two dedicated startups, iPronics and Salience Labs, each raised $125 million to build optical circuit switching from scratch, and both have demonstrated larger devices than PsiQuantum’s — Salience at 32 ports, with both companies claiming reconfiguration under 300 microseconds, an order of magnitude faster than PsiQuantum’s sub-3-millisecond figure. Lumentum, the established optical component vendor, offers MEMS switches up to 300 ports.
What PsiQuantum brings that the dedicated startups do not is a platform that already exists at manufacturing scale. iPronics and Salience had to build their photonics and foundry relationships for this market. PsiQuantum spent a decade building both for a different market, and is now redeploying them. That is a genuine advantage, but it cuts both ways: a platform optimized for quantum computing may carry constraints — wavelength choices, process steps, packaging — that a purpose-built data-center switch would not. Soderbery has positioned Broadcom, Nvidia, Cisco, Marvell, and Arista as partners rather than rivals, on the grounds that circuit switching links their existing packet switches into larger systems rather than displacing them. That is a sensible framing, and also the framing of a company that knows it cannot build the whole fabric itself.
There is a policy leg as well. In September 2026, PsiQuantum signed definitive documentation with the US Department of Commerce to accelerate research, domestic manufacturability, and performance of its critical components, including high-performance optical switches. The agreement guarantees neither manufacturing capacity nor customers, but it ties the switch program to a domestic manufacturing agenda that makes GlobalFoundries’ New York production more than a supply-chain detail.
What is not yet shown
The list of what the announcement does not contain is as important as what it does. No customers are named. No power figures are published. No loss distribution across ports, no reliability data, no pricing, and no data-center deployment are reported. The company says samples are going to select prospective customers and that commercial products will be announced in the coming months. Until a customer puts the device in a rack and reports results, the 1.2 dB figure is a company lab result, not an independently verified benchmark.
The reconfiguration time deserves its own plain-language note. Under 3 milliseconds suits stable, long-lived reconfigurations between racks — rewiring a cluster for a new training job, routing around a failed component, resizing a pod. It does not suit per-packet switching, and it is slower than the sub-300-microsecond figures claimed by iPronics and Salience. Whether that matters depends on the traffic pattern. As the IEEE Communications Society’s technology blog noted on the day of the announcement, optical circuit switching is a complement to conventional switching fabrics, not a wholesale replacement for Ethernet or InfiniBand, and its benefits depend on network architecture, traffic patterns, connection utilization, and reconfiguration overhead.
There is also a question the announcement does not address: what happens to the quantum program. PsiQuantum presents the switch as putting its platform to work today while continuing toward fault-tolerant quantum computers, not as a pivot away from them. That should be taken at face value for now, but it bears watching: a company that finds a nearer, larger market for its core technology will face pressure to allocate engineering talent and foundry capacity to that market. The test will be whether quantum milestones continue to arrive on schedule while the switch business scales.
What to watch
Three things will determine whether this announcement is remembered as a curiosity or as the opening move in a new product line. The first is the commercial product announcement promised in the coming months: its port count, loss figure, and price will tell us how much of the demonstration survives contact with a product specification. The second is the first named customer or deployment. A switch in a hyperscaler’s lab is a different fact from a switch in a hyperscaler’s network, and the difference is the whole story. The third is the 64×64 device: whether its loss stays inside the 2 to 3 dB window, and whether its reconfiguration time improves on the thermal tuning of the demo.
If those three things arrive, PsiQuantum will have done something unusual: turned a quantum-computing supply chain into an AI infrastructure business. If they do not, the 8×8 demo will stand as a reminder of a familiar pattern in photonics — a beautiful device, measured in a lab, waiting for a product to carry it. The physics, at 1.2 dB, is not the problem. Everything after the physics is.
Sources and further reading
PsiQuantum, “PsiQuantum Unveils Optical Circuit Switch Technology Built on Advanced, Low-Loss Silicon Photonics Platform,” October 28, 2026 — the primary source for the 8×8 demonstration, the 1.2 dB typical insertion loss, 40 dB crosstalk, sub-3-millisecond reconfiguration, the sampling status, the BTO distinction, and the September 2026 Department of Commerce agreement. Tom’s Hardware coverage of the announcement, October 30, 2026, as reported in secondary coverage — the source for the Nvidia request-for-proposals 3 dB target, the 64×64 tape-out, the 72×72 design in progress, the GlobalFoundries manufacturing, and the thermally tuned silicon nitride construction. Google, “Jupiter: Evolving a Datacenter Network” (ACM SIGCOMM 2022) and the IEEE Communications Society Technology Blog’s October 28, 2026 analysis — for the Palomar MEMS precedent, the 136-port design, and the framing of optical circuit switching as a complement to packet fabrics. Semafor’s reporting on the announcement, as summarized in same-day coverage — for Google as the only major hyperscaler currently using optical circuit switching in an AI data center and for the rack-to-rack latency framing.