{"slug": "psiquantum-8x8-optical-switch-hits-1-2-db-loss-2026", "title": "PsiQuantum 8×8 Optical Switch Hits 1.2 dB Loss [2026]", "summary": "PsiQuantum demonstrated an 8×8 non-blocking optical circuit switch on Oct. 28, 2026, reporting a typical insertion loss of 1.2 dB, a best measured value near 1 dB, 40 dB of crosstalk and reconfiguration in under 3 milliseconds, with no optical amplifier in the signal path. The Palo Alto company's figures come from its own lab testing relayed by The Quantum Insider, Converge Digest and Xenospectrum; no independent lab has verified them, and PsiQuantum has not disclosed pricing, volume, customer names, wavelength sweep, polarization or temperature-range data. The result matters because data center optics typically tolerate up to 3 dB of loss, and the network linking GPU clusters is emerging as a bottleneck as hyperscalers expand spending.", "body_md": "PsiQuantum put a new kind of product in front of the AI infrastructure crowd on Oct. 28, 2026. The Palo Alto company, best known for its fault-tolerant quantum computer ambitions, demonstrated an 8×8 optical circuit switch built on its low-loss silicon photonics platform. Its own testing showed a typical insertion loss of 1.2 dB, with the best measured value near 1 dB and no optical amplifier in the signal path, according to reports from The Quantum Insider and Converge Digest.\n\nThat 1.2 dB loss figure matters because data center optics have little light to spare. Tom’s Hardware framed the result around a rule of thumb that data center optics tolerate up to 3 dB of loss, which is roughly half the signal. At 1.2 dB, the PsiQuantum optical switch loses about a quarter of the light in lab tests. The company paired that with 40 dB of crosstalk performance and reconfiguration in under 3 milliseconds.\n\nThe timing is pointed. Hyperscalers keep pouring cash into GPU clusters, as our breakdown of [Big Tech’s $720 billion data center blitz](https://shattered.io/big-tech-data-center-spending-720-billion-cash-burn-wall-street-2026/) shows, and the network that ties those GPUs together is turning into the next bottleneck. A switch that steers light without converting it to electricity and back speaks directly to that problem.\n\nThis analysis covers what PsiQuantum showed, what 1.2 dB means in practice, who it competes with, what remains unproven, and what to watch next. One caution up front. Every performance figure here comes from PsiQuantum’s own lab testing as relayed by the outlets named below. No independent lab has verified any of it.\n\nDon't miss new tech stories on Google\n\nAdd Shattered once in the Google app and our stories show up in your news suggestions.\n\n[Add now](https://www.google.com/preferences/source?hl=en&q=https%3A%2F%2Fshattered.io%2F)\n\n## PsiQuantum’s 8×8 Optical Switch by the Numbers\n\nThe demonstration vehicle is an 8×8 optical circuit switch, meaning eight input ports can connect to any of eight output ports. Xenospectrum describes the design as non-blocking, with no moving parts, no mirrors and no optical amplifiers. Converge Digest says the company reports a typical insertion loss of 1.2 dB and a lowest measured value of 1 dB. The Quantum Insider adds 40 dB crosstalk and path reconfiguration in under 3 milliseconds.\n\nThe table below collects the figures that more than one outlet reported, plus the details that only one outlet carried. Where a detail came from a single source, the table says so. Where the company has not disclosed something, the table says that too.\n\n| Spec | Reported figure | Source and notes | \n|---|---|---|\n| Configuration | 8×8, non-blocking | Xenospectrum | \n| Typical insertion loss | 1.2 dB | Company lab testing, relayed by several outlets | \n| Lowest measured loss | 1 dB | Converge Digest, Xenospectrum | \n| Crosstalk | 40 dB | Measurement basis undisclosed (Xenospectrum) | \n| Reconfiguration time | Under 3 ms | The Quantum Insider, Converge Digest | \n| Optical amplification | None | All outlets reviewed | \n| Operating band | O-band | LAVX (news.lavx.hu), single source | \n| Unveiled | Oct. 28, 2026 | Xenospectrum | \n| Customer status | Samples to selected prospects | The Quantum Insider | \n| Pricing, volume, customer names | Not disclosed | Converge Digest | \n\nNote what the table lacks. There is no wavelength sweep, no polarization data, no temperature range and no sample count. Xenospectrum flagged exactly those gaps. A single typical number from a single lab tells you the physics works. It does not tell you how the part behaves across a production lot at 45 degrees Celsius inside a rack.\n\n## What 1.2 dB of Insertion Loss Means in Plain Terms\n\nDecibels hide how much light actually disappears. The conversion is simple: the fraction of power that survives equals 10 raised to the power of minus the loss divided by 10. Run that formula and a 3 dB loss leaves about 50 percent of the light. A 1.2 dB loss leaves about 76 percent. LAVX put the PsiQuantum figure at roughly 24 percent of light lost, which matches the math.\n\n| Insertion loss | Light that survives | Light lost | Context | \n|---|---|---|---|\n| 1.0 dB | 79.4% | 20.6% | PsiQuantum’s best measured value | \n| 1.2 dB | 75.9% | 24.1% | PsiQuantum’s typical value | \n| 1.5 dB | 70.8% | 29.2% | Lumentum R300 typical at start of life | \n| 2.0 dB | 63.1% | 36.9% | Illustrative midpoint | \n| 3.0 dB | 50.1% | 49.9% | Rule-of-thumb ceiling cited by Tom’s Hardware | \n\nThe percentages in the table are my own calculations from the standard decibel formula, not figures any outlet printed. The context column draws on the reporting named in each row.\n\nWhy does a few tenths of a decibel carry so much weight? Because loss in the switch eats into the link budget. Every dB a switch consumes is a dB the transceivers, fiber and connectors cannot spend. Lower switch loss means optics designers can use cheaper lasers, relax receiver sensitivity, or put more switching stages in a path. Skipping the amplifier is the bigger deal. An amplifier adds power draw, cost, noise and another part that can fail.\n\nThe 40 dB crosstalk number translates to a leakage level of one ten-thousandth of the signal power. That sounds tiny, and for an 8-port part it is a solid result. The catch, as Xenospectrum pointed out, is that PsiQuantum has not said how it measured it.\n\n## Why AI Clusters Want Optical Circuit Switches\n\nA conventional data center switch receives light, converts it into electrical signals, reads packet headers, routes the packet, then converts it back into light. That chain works, and it burns power at every hop. An optical circuit switch skips all of it. It sets up a physical light path between two ports and leaves it alone until the controller reconfigures it. The light never touches electronics inside the switch.\n\nThat tradeoff suits AI training traffic well. Large training jobs tend to hold stable communication patterns for long stretches, so a fabric that can rewire itself between jobs, or when a node fails, gets real value from circuit switching. Lumentum estimates that adopting its optical circuit switch can cut overall AI data center network power by 65 percent or more in 100,000-GPU deployments compared with InfiniBand or Ethernet designs. That is a vendor estimate, so treat it as a ceiling rather than a promise.\n\nThe same logic is pushing optics closer to the processor itself. Nvidia’s own work on putting optics on the package is a different answer to the same pressure, and our coverage of the [Nvidia optics-on-interposer prototype](https://shattered.io/nvidia-optics-on-interposer-ooi-cpo-prototype-organic-substrate-2026/) explains why the industry wants fewer electrical hops. Circuit switching and co-packaged optics are complementary. One cuts conversions at the edge of the chip, the other cuts them in the middle of the fabric.\n\nGoogle proved the concept at scale years ago. According to the 2022 paper summarized in coverage of its Apollo system, Google built its own MEMS-based optical circuit switches and used them to replace the top-tier spine layer of electrical switches across its data centers, saving cost and power. That put a stake in the ground. Optical circuit switching works in production. The open question is who supplies it to everyone who is not Google.\n\n## How the Switch Is Built\n\nReports differ on a few construction details, and the differences are worth sorting out. PsiQuantum has not published a full device description, so the picture assembled from outlets is partial.\n\n### Silicon photonics without moving parts\n\nXenospectrum says the demonstrator has no moving parts, mirrors or amplifiers. LAVX adds that the current demo uses thermally switched silicon nitride waveguides on a silicon substrate. That is consistent with a reconfiguration time measured in milliseconds, since heating a waveguide to shift its optical path is fast but not instant. It also explains the contrast with MEMS designs, which steer light with tiny mirrors that physically tilt.\n\nNo mirrors means nothing to wear out mechanically, which is a reliability argument PsiQuantum can make against MEMS rivals. It also means the device can ride the same wafer-scale manufacturing flow as other silicon photonics parts. Whether that argument holds up in a long-term reliability test is something nobody has published.\n\n### Barium titanate comes later\n\nThe Quantum Insider and Converge Digest both mention barium titanate, or BTO, a material PsiQuantum integrates at 300mm wafer scale. A 64×64 BTO switch already operates inside the company’s quantum systems, per both outlets. BTO switches by changing refractive index with an electric field, which can be much faster than heating a waveguide. LAVX reports PsiQuantum picked BTO over lithium niobate for faster future products.\n\nThe two reports line up on one point. BTO is not required for the initial AI optical circuit switch products, and the 8×8 demonstrator does not necessarily use it. So the quantum-grade 64×64 switch is a roadmap signal, not the thing being sampled today.\n\n## From Quantum Chips to Networking: PsiQuantum’s Pivot\n\nPsiQuantum was founded in 2016 and builds photonic quantum computers. Photonic quantum designs punish optical loss brutally, because every lost photon is lost information. That requirement forced the company to drive waveguide, coupler and switch losses far below what ordinary data center parts need. Converge Digest says the platform took nearly a decade to develop and rests on a foundry relationship with GlobalFoundries that dates to 2019.\n\nNow the company is pointing that same low-loss toolbox at a market that pays this year rather than in the 2030s. Fault-tolerant quantum computers remain a long-horizon bet. AI networking is a spending boom happening right now. Selling optical switches gives PsiQuantum revenue, manufacturing volume and customer feedback while the quantum program continues. The company says it will keep working on fault-tolerant quantum machines alongside the new product line, according to The Quantum Insider.\n\nRuntime Wire also reported a second motive. Building switches in volume helps harden the manufacturing supply chain before quantum machines get deployed. That reads as sensible industrial strategy. Every wafer that ships as an AI switch teaches the fab something about yield on the exact process PsiQuantum needs for quantum hardware.\n\nThe Quantum Insider named Victor Peng as CEO and Rob Soderbery as executive vice president. LAVX says Soderbery joined in July to push the silicon photonics effort. The hire timing fits a product push that landed about three months later.\n\n## The Competitive Field: Google, Lumentum, Salience and iPronics\n\nPsiQuantum is not entering an empty market. LAVX lists several rivals with their port counts, and Lumentum’s own materials fill in loss data for its R300. The comparison below mixes sources, so read the notes column before drawing conclusions.\n\n| Vendor | Device | Ports | Loss and speed data | Technology | \n|---|---|---|---|---|\n| PsiQuantum | 8×8 demonstrator | 8 | 1.2 dB typical, under 3 ms | Silicon photonics, no moving parts | \n| Lumentum | R300 | 300 | 1.5 dB or less typical at start of life, up to 3 dB cited | MEMS mirrors | \n|  | Palomar (Apollo) | 136 (LAVX) | Not disclosed in reports reviewed | In-house MEMS, internal use | \n| Salience Labs | OC-32M | 32 (LAVX) | Not disclosed in reports reviewed | Not specified | \n| iPronics | Not specified | Not specified | Switching under 300 microseconds (LAVX) | Programmable photonics | \n\nThe table shows why the headline number needs context. Lumentum’s R300 handles 300 ports and still posts a typical loss around 1.5 dB at the start of its life, per Gazettabyte and Lumentum’s own blog. PsiQuantum’s 1.2 dB comes from a part with 8 ports. Comparing the two directly would be like comparing a sprinter’s 100-meter time with a marathoner’s pace. The honest reading is that PsiQuantum beats the R300 on a per-switch loss figure at tiny scale and says nothing yet about whether it holds at hundreds of ports.\n\nOn speed, PsiQuantum is not the leader either. LAVX reports iPronics claims switching in under 300 microseconds. PsiQuantum’s under-3-millisecond ceiling sits an order of magnitude above that. For training workloads that reconfigure rarely, the gap may not matter. For fabrics that want to rewire between short bursts, it could.\n\nLAVX also says Nvidia holds a stake in PsiQuantum through NVentures and took part in iPronics’ $125 million round in September. If accurate, Nvidia is hedging across several optical switching approaches, which is exactly what you would expect from the company that sells most of the GPUs these switches would connect.\n\n## Why an 8-Port Result Does Not Settle Anything\n\nSkeptics have solid ground here. Xenospectrum listed the missing evidence plainly: no peer-reviewed validation, no third-party verification, no disclosed wavelength, polarization, temperature or sample count, no measurement of data center power draw, and no AI workload results.\n\n### What a lab number can and cannot prove\n\nA typical value of 1.2 dB says the median measurement came out there. It does not say the worst port pair, the worst temperature or the worst unit in a batch. Production switches get judged on worst-case specs, because one bad path can stall a whole training job. Buyers will want a distribution of losses across all 64 path combinations of an 8×8 part, then across many chips.\n\nCrosstalk deserves the same scrutiny. A 40 dB number with an undisclosed measurement basis could mean the best path pair or the average. It could also mean a single-wavelength test that will not carry over to a full band.\n\nThen comes the packaging question. Chip-to-fiber coupling, the step where light enters and leaves the silicon, often decides whether a lab result survives contact with a real product. LAVX cites a best-case lab coupling loss near 1.4 percent. Production assemblies rarely match best-case lab numbers, and the reports do not say whether the quoted 1.2 dB includes the coupling penalty.\n\nNone of this makes the demo a stunt. It makes it an early data point. PsiQuantum is handing samples to selected prospective customers, and those customers will run the tests that matter.\n\n## The Scaling Question: 64×64 and 72×72\n\nThe real test is port count. LAVX reports that a 64×64 version has taped out and a 72×72 version is in development. Runtime Wire says the company names 64 ports and beyond as a goal, and notes that scalability past 8 ports is not yet demonstrated. Tom’s Hardware’s headline likewise points to plans for larger devices.\n\nLoss usually climbs as port counts grow. A bigger switch needs more waveguide length, more crossings and more splitting stages, and each one nibbles at the budget. LAVX says PsiQuantum declined to share future product details but indicated that loss will stay low, though slightly higher at larger sizes. That is a reasonable thing to say and an unfalsifiable one until real numbers appear.\n\nEven a 72×72 part would sit well below Lumentum’s 300-port R300 and Google’s reported 136-port Palomar. Large AI fabrics typically want high radix, because more ports per switch means fewer layers and a flatter network. Lumentum itself pitches the R300’s high radix as the thing that enables flat architectures for 100,000-plus GPU deployments. A 64 or 72-port part can still find homes, such as rack-level or pod-level reconfiguration, but PsiQuantum will have to show either larger radix or a cost advantage that pays for stacking several smaller switches.\n\n## Power, Cost and the Data Center Bill\n\nPower is the argument that sells optical switching to finance teams. AI campuses now plan around electricity first and silicon second, a theme we covered in [the AI power problem, from 800V DC to 20-year nuclear deals](https://shattered.io/ai-power-problem-delta-800v-dc-cloverleaf-microsoft-20-year-2026/). Anything that trims the network’s share of that bill gets a hearing.\n\nHere the evidence for PsiQuantum is thin. Xenospectrum noted that the company has not measured overall data center power consumption with its switch in the loop. The 1.2 dB loss figure implies an amplifier-free path, which saves the power an amplifier would draw. It says nothing about the control electronics, the thermal tuning power or the laser budget needed to compensate for loss elsewhere.\n\nThermal switching has its own power cost. Heating waveguides to hold a path draws continuous power per element, unlike a MEMS mirror that can hold position with little energy. LAVX says the current demo is thermally switched, so a fair accounting needs a watts-per-port figure that nobody has published. This is the number I would ask for first, ahead of any further loss data.\n\nPricing is also a blank. Converge Digest confirms pricing, production volume and customer names are undisclosed, and Xenospectrum says volume shipments have not begun. Without a price, there is no way to compute cost per port against a Lumentum or Google alternative.\n\n## Funding, Government Money and the GlobalFoundries Link\n\nPsiQuantum has capital to burn on this detour. Converge Digest reports a Series E in September 2025 that raised $1 billion at a reported $7 billion valuation, led by BlackRock affiliates, Temasek and Baillie Gifford, with NVentures, Macquarie Capital, Ribbit Capital and the Qatar Investment Authority among the other backers.\n\nPublic money has followed. The Quantum Insider reports a $100 million award from the U.S. Department of Commerce in September 2026 for research, domestic manufacturing and component performance. It also reports a $150 million conditional loan from what it calls the U.S. Department of War in October 2026. Converge Digest adds an expanded $125 million DARPA quantum benchmarking agreement in 2026. These reports describe the awards differently, so treat the exact program names and amounts as the outlets’ claims until the agencies publish their own notices.\n\nManufacturing runs through GlobalFoundries, per Converge Digest, which gives PsiQuantum access to an established high-volume foundry rather than a boutique line. That matters for the pitch to hyperscalers, who care about supply continuity as much as peak specs. A wafer-scale process also gives PsiQuantum a plausible route to cost curves that hand-assembled MEMS arrays struggle to match.\n\n## Historical Context: From Google’s Apollo to Merchant Optical Switching\n\nOptical circuit switching is not a new idea. It spent decades as a telecom curiosity. Google’s 2022 paper changed the narrative by showing that an operator could build its own MEMS switches and run them across a fleet. The paper revealed that Google had swapped electrical spine switches for optical circuit switches and booked cost and power savings.\n\nThe merchant market followed. Lumentum announced its optical circuit switch for next-generation AI data center scalability in March 2025, and by now the R300 anchors the high-radix end. The 2026 wave brings challengers on different physics: programmable photonics from iPronics, a 32-port design from Salience Labs, and now silicon photonics from a quantum computing company.\n\nPsiQuantum’s entry fits a pattern worth watching. Technology developed for a hard science problem finds a commercial home in a nearer market while the original goal keeps maturing. Photonic quantum computing demands loss performance that conventional telecom parts never needed. Now the AI boom has made loss a premium specification, and PsiQuantum happens to own a platform tuned for it.\n\n## Market Impact: Who Gains and Who Loses\n\nIn the near term, the market impact is small. An 8-port sample moves no revenue. The impact is in expectations. The announcement tells hyperscalers and GPU makers that a second silicon-photonics source for optical circuit switching is coming, and that source has deep funding, a major foundry partner and a loss figure that undercuts the MEMS incumbent’s typical number.\n\nLumentum has the most to defend. Its R300 owns the high-radix position today, and its argument rests on 300-port scale that nobody else can match. PsiQuantum cannot touch that yet. If the company delivers a 64 or 72-port part at lower cost, though, it can undercut the R300 in pod-level roles and pressure margins from below.\n\nGPU vendors benefit from any technology that makes bigger clusters practical. The latest rack-scale platforms are bandwidth-hungry, and we reported earlier on [Nvidia trimming Vera Rubin HBM4 bandwidth by 13 percent](https://shattered.io/nvidia-vera-rubin-bandwidth-cut-hbm4-19-2-tbs-nvlink-2026/), a reminder that every layer of the memory and network stack is under strain. Cheaper, lower-loss fabric switching eases one piece of that squeeze. Our [hardware coverage](https://shattered.io/hardware/) tracks how those pieces fit together.\n\nTransceiver and laser suppliers sit on the other side of the trade. Lower switch loss shrinks the optical power each link needs, which can trim laser content per port. That cuts both ways. A cheaper link budget lowers the price of the fabric, and it also trims the revenue pool for parts sold into it.\n\n## Four Predictions for Optical Switching in AI Networks\n\nThese are my own analytical calls based on the reporting above. They are not company statements or confirmed plans.\n\n- **A product announcement lands in the first half of 2027.** The Quantum Insider and Converge Digest both say commercial products are expected within the next few months. I expect a named 64-port part, probably with a disclosed worst-case loss spec.\n- **The first worst-case loss spec will exceed 1.2 dB.** Typical-value demos almost always lead to looser datasheet limits. Expect the 64-port datasheet to quote a ceiling well above the 8-port typical.\n- **At least one hyperscaler will publicly confirm a pilot by the end of 2027.** Samples are already out to selected prospects. A pilot with a big name would validate the platform faster than any spec sheet.\n- **Watts per port becomes the headline metric.** Once loss figures converge near 1 to 1.5 dB across vendors, buyers will compare power draw, switching speed and price per port. Thermal switching will face the toughest questions there.\n\nThe risk to all four calls is simple. Scaling from 8 to 64 ports could expose losses the small demo hid, and the reports give no evidence either way.\n\n## What Buyers and Engineers Should Watch Next\n\nIf you run network architecture for a GPU cluster, the PsiQuantum optical switch is a watch-and-request-samples item, not a purchase-order item. Here is the checklist I would use when the company shares a datasheet.\n\n- Worst-case insertion loss across all port pairs, wavelengths and a stated temperature range.\n- Polarization-dependent loss, which the reports do not mention at all.\n- Whether the quoted loss includes fiber coupling and packaging.\n- Watts per port for the thermal tuning and control electronics.\n- Reconfiguration time under load, not just a best-case figure.\n- Lot-to-lot yield and long-term reliability data from GlobalFoundries production.\n- Radix roadmap with dates, plus pricing against a MEMS-based alternative.\n\nSupply chain context matters as well. Memory pricing is already squeezing AI hardware budgets, as we saw in [the Q4 DRAM price rise and the $450 DDR5 kit](https://shattered.io/dram-prices-q4-2026-trendforce-ddr5-450-micron-54b-2026/), so any network savings will get compared against a rising bill of materials. A switch that saves power but costs more per port may lose the budget fight.\n\nThe takeaway: PsiQuantum posted a strong small-scale result and wants the AI networking market to take it seriously. The next datasheet, not the current demo, will decide whether it does.\n\n## FAQ: PsiQuantum’s Optical Switch\n\n**What did PsiQuantum announce?** On Oct. 28, 2026, the company demonstrated an 8×8 optical circuit switch built on its silicon photonics platform. It reports a typical insertion loss of 1.2 dB, crosstalk of 40 dB and reconfiguration in under 3 milliseconds, with no optical amplification.\n\n**Is 1.2 dB of loss good?** It is competitive. A 1.2 dB loss leaves about 76 percent of the light, compared with roughly 71 percent at Lumentum’s 1.5 dB typical figure for the R300. The catch is port count: PsiQuantum’s switch has 8 ports and the R300 has 300.\n\n**Can I buy the switch today?** No. PsiQuantum is sharing demonstration samples with selected prospective customers. Converge Digest says pricing, production volume and customer names are undisclosed, and Xenospectrum says volume shipments have not started.\n\n**How is it different from Google’s optical switches?** Google’s Apollo system uses in-house MEMS mirrors and serves Google’s own data centers. PsiQuantum’s demonstrator has no moving parts, per Xenospectrum, and the company plans to sell to outside customers.\n\n**Does the switch use barium titanate?** Not necessarily. The company uses BTO in 64×64 switches inside its quantum systems, but reports say BTO is not required for the initial AI products and the 8×8 demonstrator may not use it. LAVX says the current demo is thermally switched silicon nitride.\n\n**Who makes the chips?** Converge Digest says PsiQuantum works with GlobalFoundries, a partnership that began in 2019.\n\n**Have the results been independently verified?** No. Xenospectrum stressed that the figures come from in-house testing with no peer review, no disclosed test conditions and no third-party evaluation in AI workloads.\n\n**Will larger versions exist?** The company plans them. LAVX reports a 64×64 version has taped out and a 72×72 version is in development, and says PsiQuantum expects loss to stay low but rise slightly with size. Larger-radix performance has not been disclosed.", "url": "https://wpnews.pro/news/psiquantum-8x8-optical-switch-hits-1-2-db-loss-2026", "canonical_source": "https://shattered.io/psiquantum-8x8-optical-switch-ai-data-centers-1-2-db-loss-2026/", "published_at": "2026-11-02 01:29:22+00:00", "updated_at": "2026-11-02 01:49:41.233469+00:00", "lang": "en", "topics": ["ai-infrastructure", "ai-chips"], "entities": ["PsiQuantum", "The Quantum Insider", "Converge Digest", "Xenospectrum", "Tom's Hardware", "LAVX"], "also_reported_by": [], "alternates": {"html": "https://wpnews.pro/news/psiquantum-8x8-optical-switch-hits-1-2-db-loss-2026", "markdown": "https://wpnews.pro/news/psiquantum-8x8-optical-switch-hits-1-2-db-loss-2026.md", "text": "https://wpnews.pro/news/psiquantum-8x8-optical-switch-hits-1-2-db-loss-2026.txt", "jsonld": "https://wpnews.pro/news/psiquantum-8x8-optical-switch-hits-1-2-db-loss-2026.jsonld"}}