PsiQuantum's first product has eight inputs and eight outputs. It is an optical circuit switch, not a quantum computer, and it is aimed at the network bottleneck inside AI data centers.
The Palo Alto company disclosed the device on October 28. Its lab results: 1.2 dB of typical insertion loss, 40 dB of crosstalk, reconfiguration under 3 milliseconds, no optical amplification, and no moving parts.
Port count is the comparison that frames the investment case. Google's Palomar switch runs 136 inputs and 136 outputs in production inside its Jupiter network. PsiQuantum's demonstration runs eight.
Google reported a 5× gain in network speed and capacity across its Jupiter evolution, which uses optical circuit switches and software-defined networking, plus a 30% cut in capital expenditure and a 41% cut in power consumption. Those figures come from live traffic, not a laboratory.
PsiQuantum is sharing samples with selected prospective customers. Commercial products are promised "in the coming months." There is no price, no named customer, and no published result at a larger port count.
What the demonstration does and does not prove
Light surviving the 8×8 switch
About 76% of the signal passes through; the best reported figure was 1 dB. PsiQuantum, October 2026
The device is a strictly non-blocking 8×8 switch, meaning any input can reach any output. It is built on the silicon-photonics platform PsiQuantum has developed since 2016 for fault-tolerant quantum computing, using processes from GlobalFoundries.
Crosstalk of 40 dB is a 10,000-to-1 separation between neighbouring channels. Reconfiguration under 3 milliseconds is slow by packet-switching standards. It suits scheduled, predictable flows, which is what large training jobs generate.
The omissions are specific. PsiQuantum has not published loss at a larger radix, named a customer, set a price, or confirmed which foundry will build the commercial part. Evaluation samples are not a product line.
There is a second switch inside the company's quantum systems: a 64×64 device built on barium titanate. PsiQuantum says that material is not required for the demonstration or the first commercial products. The initial AI part stays on a simpler silicon-photonics process.
The loss budget is the constraint that matters
Data-center optics usually tolerate about 3 dB of loss across a link. Half the light can disappear before the connection fails. A switch that spends 1.2 dB leaves the rest of the allowance for fiber, connectors, splices and patch panels.
That arithmetic is the reason low-loss switching has value. It is also the reason the number has to hold as port count rises. Loss generally grows with radix, and a 64-port device cannot inherit an 8-port result without new evidence.
A 1.2 dB figure describes one device. A deployed cluster adds patch panels, fiber runs, splices and connectors, all of which spend the same 3 dB budget. The switch's share has to stay small after everything else is counted.
Optical circuit switching does not replace packet switching. It sets a light path between ports and leaves it there until the traffic pattern changes. Bursty, unpredictable traffic still needs electrical packet forwarding. The realistic design is hybrid: circuits for all-reduce and all-to-all exchanges, packets for the rest.
Google's production experience supports that split. Its Jupiter network keeps a packet layer and uses optical circuits where the traffic is predictable enough to schedule. The 30% capex and 41% power savings came from the combination, not from optics alone.
Google already proved the concept at 136 ports
| Parameter | PsiQuantum 8×8 (demo) | Google Palomar (production) |
|---|---|---|
| Ports | 8 × 8 | 136 × 136 |
| Insertion loss | 1.2 dB typical, 1 dB best | Under 2 dB typical, 2 dB worst case |
| Reconfiguration | Under 3 ms | Millisecond-class |
| Switching | Silicon photonics, no moving parts | MEMS mirrors |
| Status | Samples with select customers | Live in Jupiter |
Company disclosures and the ACM SIGCOMM 2022 Jupiter paper. Figures are vendor-reported and not independently benchmarked.
PsiQuantum's loss is lower at eight ports than Palomar's at 136. That is a real result. It is also the easiest version of the problem, because loss, crosstalk and yield all get harder as the switch matrix grows.
The gap between 8 and 136 is where the investment question sits. PsiQuantum has to show that its platform keeps the advantage at cluster radix. Nobody outside the company has measured the larger device because there is no larger device to measure.
The manufacturing argument behind silicon photonics
MEMS switches steer light with tiny mirrors. They work, and Google runs them at 136 ports. The trade-off is mechanical: mirrors need alignment, packaging and control loops, and every additional port adds beam-steering complexity.
Silicon photonics moves the switching function into a chip. There are no moving parts, and the device can be built on CMOS-style processes. PsiQuantum's pitch is a switch that inherits the cost curve of a semiconductor fab rather than the cost curve of precision optics.
That pitch has limits. Photonic chips still need fiber attach, packaging and thermal control, and those steps dominate cost at low volume. Yield also falls as the switch matrix grows, because one defective element can take out a path. PsiQuantum has not published yield data at any port count.
The published specification is also narrow. It covers insertion loss, crosstalk and reconfiguration time. Bandwidth, polarization dependence and operating temperature across a data-center range are not in the release. Those parameters decide whether a component can sit inside a real link budget.
PsiQuantum and GlobalFoundries have worked together on photonic chip manufacturing since 2019, and PsiQuantum has qualified 300 mm barium-titanate integration. Neither fact proves that a commercial switch can be manufactured at data-center volumes or prices. They show that the process exists and that a tier-one foundry is already involved.
The comparison with iPronics is instructive. iPronics also uses silicon photonics, but it sells a complete rack-ready system with telemetry and APIs. PsiQuantum is selling a component and a platform story. Systems integration is where margins are made, and it is the part PsiQuantum has not shown.
The market is crowded before PsiQuantum arrives
iPronics raised a $125 million Series B in September, co-led by Maverick Silicon and Light Street Capital, with NVIDIA participating. Its iPronics ONE is already sold as a rack-ready optical circuit switch with control software and APIs. Total funding reached $177 million.
Salience Labs is developing an optical switching chip. Lightmatter is pitching optical scale-up interconnects. Resolight has built a nanosecond switching architecture and is working through standards bodies. Oriole Networks is integrating photonic networking with AMD accelerators.
As we wrote in September, iPronics raised $125 million and brought a rack-ready optical switch to market. PsiQuantum's entry lands in the same category with a different maturity profile: better component physics, less system integration.
Established suppliers fill the rest of the field. Lumentum, Calient, Coherent, Polatis and Telescent sell MEMS, liquid-crystal, piezoelectric or robotic cross-connect systems. Google is the only major hyperscaler running pure optical circuit switching in an AI data center today, according to Semafor.
That matters for pricing. A component demo from a quantum company does not reset the market. It creates an option on a better switching fabric, and the option is worth something only if the port count scales.
Why a quantum company is selling switches
PsiQuantum's platform was built for a harder customer than a data center. Fault-tolerant photonic quantum computing demands low loss, tight phase control and wafer-scale manufacturing. Switching in an AI cluster is a looser set of requirements, which makes the platform reusable.
This opportunity exists because we chose to solve the hardest scaling problems at the beginning of our roadmap to utility-scale quantum computing.— Victor Peng, CEO, PsiQuantum
The balance sheet explains the timing. PsiQuantum raised a $1 billion Series E in September 2025 at a reported $7 billion valuation, with BlackRock, Temasek and Baillie Gifford funds leading. In September 2026 it signed definitive documentation for a $100 million award from the U.S. Department of Commerce to support domestic photonics manufacturing, including optical switches.
Public support goes further. PsiQuantum holds a conditional commitment for a $150 million loan from the U.S. Department of War's Office of Strategic Capital, and an expanded $125 million DARPA agreement announced in 2026. GlobalFoundries has manufactured its photonic chips since 2019.
The switch gives that roadmap a nearer-term output. Management frames it as a by-product of quantum work. Investors will treat it as evidence that the platform can produce sellable parts before the quantum computer exists.
What has to be true by 2027
Four milestones separate a credible second product line from a well-timed announcement.
- A switch at cluster-relevant radix, with loss and crosstalk measured by an independent customer rather than the company's lab.
- At least one named hyperscaler or neocloud running it inside a production fabric, even at modest scale.
- System-level evidence: watts or capex saved per deployed rack, measured against the 3 dB link budget rather than the component figure.
- A confirmed foundry and a volume price that survives comparison with MEMS incumbents that already ship at 136 ports.
The first two are engineering and sales problems. The last two decide whether optical switching becomes a business line or stays a demonstration.
For PsiQuantum, the strategic value may not depend on winning the OCS market outright. A working commercial switch validates the manufacturing platform, opens a customer channel into hyperscalers, and gives the quantum roadmap more years of runway. That is a defensible reason to sell a component that has nothing to do with qubits.
PsiQuantum has the best reported loss at this port count, and the comparison that matters is Google's 136-port production switch.
The commercial gap is radix and integration, not component physics. Those take customer trials and a foundry commitment.
Price this as an option on silicon-photonics scaling, funded by a $7 billion quantum balance sheet and government awards. It is not a networking revenue line today.
Eight ports is a start. The next number that matters is the one PsiQuantum has not published.