Table of Contents
(This profile is one entry in my 2025 series on quantum hardware roadmaps and CRQC risk. For the cross‑vendor overview, filters, and links to every company profile, see Quantum Hardware Companies and Roadmaps Comparison 2025.)
Updated November 2026. In this revision I added Xanadu’s March 2026 listing on Nasdaq and the Toronto Stock Exchange, its August 2026 roadmap to 2031, its position in DARPA’s Quantum Benchmarking Initiative and its 2026 manufacturing partnerships. I also corrected three statements in the 2025 version. Xanadu’s photon-number-resolving detectors operate in cryostats, so its machines are not entirely room-temperature. X8, launched in 2020, was Xanadu’s first photonic processor offered over the cloud; Borealis followed in 2022. Borealis performed Gaussian boson sampling and did not build cluster states for computation.
Introduction
Xanadu is a Toronto company building quantum computers out of light. Christian Weedbrook founded it alone in 2016, after a PhD in quantum information theory at the University of Queensland and postdoctoral positions at MIT and the University of Toronto. He worked on continuous-variable quantum information and co-authored the 2012 Reviews of Modern Physics survey of Gaussian quantum information. Before Xanadu he ran CipherQ, a small continuous-variable quantum key distribution company, as he recounted in a 2026 Decode Quantum interview. Xanadu bases its fault-tolerant design on that line of research. Its machines compute with squeezed light and homodyne detection and encode qubits in continuous variables, whereas PsiQuantum and most other photonic competitors encode them in single photons.
The proposed fault-tolerant architecture has three stages, each on its own photonic chips, connected by optical fiber. Squeezed-light sources on silicon nitride chips feed small Gaussian boson sampling circuits. When photon-number-resolving detectors register the right photon counts, they herald non-Gaussian states. Heralding succeeds only some of the time, so a second set of chips multiplexes the outputs, refines them into Gottesman-Kitaev-Preskill (GKP) qubits and entangles them in pairs. A third set stitches the pairs into a cluster state and computes by measuring it with homodyne detectors whose settings depend on earlier measurement results. Once the GKP states exist, the entangling gates are beamsplitters and therefore deterministic. Only the detectors that herald the input states need cryogenic cooling. Xanadu has demonstrated major parts of this chain in published prototypes. Xanadu has not yet built the complete state factory or the fault-tolerant processor.
Since March 2026 Xanadu has been a public company, listed as XNDU. It has published four hardware papers in Nature, from programmable photonic chips to major components of its fault-tolerant design and their integration. It now targets fault tolerance in 2028–2029 and more than 1,000 logical qubits by 2031. Xanadu has also published a measure of how far its hardware is from those targets. The key loss indicator is the ratio of measured hardware loss to Xanadu’s loss requirement, aggregated across 17 key optical components. It stood at 24.1 in 2026, and Xanadu targets 1.0 by 2030. Xanadu has to bring that number down before it can meet any later target in its roadmap.
Xanadu says its mission is to build quantum computers that are useful and available to people everywhere. It plans a fault-tolerant photonic quantum data center by the end of the decade. Alongside the hardware it develops PennyLane, an open-source quantum programming framework that, by Xanadu’s count, is the second most used among developers after IBM’s Qiskit.
Milestones & Roadmap
2016–2019: Weedbrook founded Xanadu through the University of Toronto’s Creative Destruction Lab. In 2018 the company released two open-source libraries: Strawberry Fields, for designing and simulating continuous-variable photonic circuits, and PennyLane, for differentiable quantum programming and quantum machine learning. It closed its Series A in 2019.
2020–2021: In September 2020 Xanadu put X8, an eight-mode programmable photonic chip, on its own cloud service and described it as the first photonic quantum computer in the cloud. The X8 paper appeared in Nature in March 2021. In 2021 Xanadu raised a Series B of about US$100 million, led by Bessemer.
2022: Xanadu made its quantum computational advantage claim in Nature in June 2022 with Borealis, a 216-mode time-multiplexed Gaussian boson sampler read out by photon-number-resolving detectors. The authors estimated that the best known classical algorithm on a supercomputer would need about 9,000 years for each sample Borealis produced in 36 microseconds. Xanadu offered Borealis on its own cloud and on Amazon Braket. Later in 2022 it raised a US$100 million Series C at a US$1 billion valuation, for total funding of about US$250 million.
Gaussian boson sampling is a specialized task with no established practical computational advantage. In 2024 Changhun Oh, Minzhao Liu, Yuri Alexeev, Bill Fefferman and Liang Jiang published a tensor-network algorithm in Nature Physics that exploits the high photon loss in these experiments. On the benchmarks the experimental teams used to claim advantage, their classical sampler matched the ideal distribution better than the largest experiment published at the time. In Xanadu’s own Aurora paper, the authors cite that work and acknowledge that photon loss leaves photonic sampling experiments vulnerable to classical simulation. Borealis remains an engineering milestone in programmable photonics; its advantage claim is now contested.
Jan 2025: Xanadu published Aurora in Nature: 35 photonic chips in four server racks, linked by 13 km of fiber, with 84 squeezers and 36 photon-number-resolving detectors supplying 12 physical qubit modes per clock cycle. Over two hours at a 1 MHz clock rate, with only squeezed-light inputs, the machine generated and measured a cluster state of 86.4 billion modes entangled across separate chips, made up of 12 spatial modes over 7.2 billion time steps. Those modes were produced one after another; at any moment the machine was working with 12. In a separate experiment, Aurora ran a distance-2 repetition code on low-quality GKP states, with a decoder making decisions in real time. The racks operate at room temperature, but the 36 transition-edge-sensor detectors are cooled to 12 millikelvin in two dilution refrigerators.
The authors call Aurora a “(sub-performant) scale model” and state that it was too noisy to show error suppression. It demonstrated that Xanadu’s modules work together over fiber. Xanadu has yet to show fault tolerance. After Aurora, it moved its engineering focus to optical loss.
Mar 2025: Blayney Walshe and colleagues on Xanadu’s architecture team showed in Physical Review Letters that GKP qubits combined with static linear optics can implement any quantum error-correcting code, including the qLDPC codes that need long-range connections between qubits. In the preprint’s simulations, hyperbolic surface codes reached a threshold comparable to the 2D surface code at about ten times its encoding rate. Under the paper’s noise model, such codes need fewer modes per logical qubit in the outer code. Encoding rate is only part of the hardware cost, which also includes preparing GKP states and the optical modes that make up the code. Xanadu’s hardware has yet to reach the error rates the simulations assume.
Apr 2025: Xanadu entered Stage A of DARPA’s Quantum Benchmarking Initiative, in which the agency is examining whether any approach can reach utility-scale quantum computing by 2033. It also began a four-year research partnership with the U.S. Air Force Research Laboratory.
June 2025: Xanadu reported the first GKP states generated on an integrated chip, in Nature. The source was a multilayer silicon nitride chip fabricated on 300 mm wafers, coupled by fiber to photon-number-resolving detectors that Xanadu built in-house and that, the company reports, exceed 99% efficiency. The states showed at least four resolvable peaks in each quadrature and a 3×3 grid of negative regions in the Wigner function, the structure that GKP error correction requires. They were not of fault-tolerant quality, and the authors attribute the gap mainly to optical loss. The same month, Xanadu opened a photonic packaging facility in Toronto.
Mid-2025: Xanadu began work with Applied Materials on the first 300 mm fabrication process for transition-edge sensors, the superconducting detectors Aurora used for photon-number resolution. In August 2025 it partnered with Japan’s DISCO Corp. on dicing and polishing processes for ultra-low-loss photonic integrated circuits.
Nov–Dec 2025: On 3 November Xanadu announced its plan to go public through a merger with Crane Harbor Acquisition Corp., a blank-check company, at a pre-money equity value of US$3.0 billion. In the same announcement it projected up to 100,000 physical qubits and up to 1,000 logical qubits in its fault-tolerant computer by 2029. Three days later it advanced to Stage B of DARPA’s initiative, with up to US$15 million in funding. It also expanded its research partnership with Singapore’s A*STAR. In December the Canadian government selected Xanadu for its Quantum Champions Program, with up to C$23 million in support.
Mar 2026: The merger closed on 26 March. With its shares trading on Nasdaq and the Toronto Stock Exchange from the next day, Xanadu became the first publicly listed company focused solely on photonic quantum computing. It raised about US$302 million in gross proceeds from the transaction, including a US$275 million private placement at US$10 a share. Earlier that month Xanadu had opened negotiations for up to C$390 million from the governments of Canada and Ontario for Project OPTIMISM, a domestic manufacturing program. It also announced a collaboration with TELUS on quantum data center infrastructure.
May–Aug 2026: Xanadu set up a US$300 million synthetic at-the-market equity facility with Yorkville Advisors, under which it may sell shares over three years but is not obliged to. It drew US$67.2 million from the facility in the second quarter. It reported an average edge-coupling loss of 0.085 dB per facet between fiber and chip. It also partnered with EV Group on wafer bonding and lithography. Xanadu also expanded its U.S. operations in Albany, New York.
On 28 August the federal government committed C$195 million through its Strategic Response Fund, the full federal share of Project OPTIMISM, toward a C$893 million project with 275 jobs. The project includes “Inception,” a 158,000-square-foot former Campbell Soup plant in Toronto that Xanadu is fitting out for photonic chip packaging, testing and rack assembly. Ontario’s share was still under negotiation when the federal agreement was signed.
Sep–Oct 2026: Xanadu announced a lithography collaboration with ASML; released Backline with AMD, an open-source layer connecting PennyLane to CPUs, GPUs and FPGAs; partnered with Bluefors on a modular cryogenic unit for single-photon detectors operating near 2 kelvin; and signed a multi-year agreement with GlobalFoundries to industrialize its superconducting nanowire single-photon detectors (SNSPDs) and ultra-low-loss silicon nitride photonics on GF’s 300 mm line in Malta, New York. In the GlobalFoundries release Xanadu says it has adopted SNSPDs as its photon-number-resolving platform in place of the transition-edge sensors used in Aurora. It expects the change to make its systems faster and cut chip overheads.
On 19 October Xanadu received provisional approval for up to US$10 million from DARPA for a one-year extension of its Stage B work. Twelve days earlier, DARPA had moved four other companies into Stage C.
2026–2031 (Planned): Xanadu’s August 2026 roadmap lists a “qubit factory” built in 2026–2027; a fault-tolerance milestone in 2028–2029; a quantum data center in 2029–2030; up to 200 logical qubits in 2029, up to 500 in 2030 and more than 1,000 in 2031; and logical error rates improving from roughly 10−3 to 10−8 in 2028 to 10−16 by 2030. Xanadu gives no physical qubit counts in the roadmap and does not say whether the error rates are per cycle, per operation or per computation. It has also moved its published schedule. In the November 2025 merger announcement it projected up to 1,000 logical qubits by 2029; in the August 2026 roadmap it puts more than 1,000 in 2031. The two announcements may not describe identical machines, but Xanadu has not explained the shift. Weedbrook has described about one million physical qubits as Xanadu’s yardstick for useful quantum computing.
In its roadmap slides, Xanadu defines its key loss indicator as hardware loss divided by loss requirement, aggregated over 17 key optical components, and targets a fall from 24.1 in 2026 to 1.0 in 2030. Xanadu has not published how the requirement is set, whether it marks the threshold of the error-correcting code or a budget with margin for a target logical error rate, or how the 17 components are weighted. Without those definitions, outsiders cannot tell how far a complete machine is from threshold today, or how Xanadu gets from a ratio of 1.0 in 2030 to the logical error rate of 10−16 it gives for the same year. Until Xanadu publishes them, I read 2030 as the year it expects its components to meet their loss budgets, and the logical error rates as projections from its models. DARPA’s Quantum Benchmarking Initiative is the nearest thing to an independent check on these plans. Xanadu is currently in a one-year extension of Stage B.
Focus on Fault Tolerance
For fault tolerance, Xanadu uses bosonic encoding, specifically the Gottesman-Kitaev-Preskill code. A GKP state encodes one qubit in a single optical mode as a grid of peaks in phase space, a comb of squeezed states. Small displacements of the mode move those peaks. The displacements can be measured and corrected with Gaussian operations (beamsplitters, phase shifts and homodyne measurement) together with fresh GKP states. Optical loss does more than displace the state, but with suitable compensation its effect can be treated as displacement noise of this kind.
Xanadu’s error correction has two layers, with GKP encoding as the inner one. Each optical mode encodes a qubit with some built-in protection, which makes it a logical GKP qubit in the strict sense. Xanadu then uses those modes as the physical inputs to an outer code, now a qLDPC code, to build the logical qubits on which algorithms would run.
Why GKP: with this encoding, Xanadu moves the hard part of photonic quantum computing from gates to state preparation. Once GKP states exist, Clifford gates and entanglement need only beamsplitters, phase shifters and homodyne detectors, all deterministic and all at room temperature. Optical fiber can connect any two modes, so the outer code does not have to be geometrically local. In the 2025 Physical Review Letters paper, Xanadu’s team simulated qLDPC codes with a threshold comparable to the surface code and roughly ten times the encoding rate. In exchange, Xanadu has to produce GKP states of high quality. In the Aurora analysis, fault tolerance requires an effective GKP squeezing of 9.75 dB, a property of the error-correction scheme and decoder. The Aurora authors also assume 15 dB of squeezing at the state factory’s light sources. Loss between source and measurement lowers the effective squeezing the hardware delivers.
Xanadu has demonstrated several pieces of an error-corrected qubit separately. The June 2025 chip produced GKP states with the right structure but below fault-tolerant quality. Aurora showed real-time decoding and feedforward: its full signal chain, from an optical pulse arriving at the homodyne detectors to an updated measurement setting, took about 976 nanoseconds, of which the decoder took 64. The loop fitted inside one 1 µs clock cycle, as Aurora’s feedforward scheme required. For large qLDPC codes, decoder performance is measured differently: decoding has to keep up with incoming data at an acceptable latency, though a pipelined decoder may take longer than one cycle per decision. In Aurora’s repetition-code experiment, the non-Gaussian inputs were squeezed cat states, low-quality approximations of GKP states, and they made up about 3% of the modes.
In its 2025 annual results, Xanadu describes Aurora as demonstrating “12 logical GKP qubits with real-time error correction.” At the inner layer the wording has a basis: each of Aurora’s 12 modes per clock cycle can encode a GKP qubit. In the Nature paper, the authors report something narrower: a system too noisy to show error suppression, low-quality GKP states and a distance-2 repetition-code error-detection experiment. In its annual report to the SEC, Xanadu comes closer to the paper’s description: it says it has “only demonstrated twelve logical qubits with photon loss not meeting the required levels.” No outer-code logical qubit has run below threshold on Aurora. Below-threshold operation means that the logical error rate falls as the code distance grows, the test for scalable error suppression. Xanadu has not yet published a result that meets it.
The Aurora authors estimate that the best known configurations of Xanadu’s architecture require loss budgets of about 1% and multiplexing depths of about ten. On Aurora they measured about 56% loss on the heralding paths and slightly over 95% on the paths carrying the heralded states. They estimate that a 20- to 30-fold improvement in each component’s insertion loss, measured in decibels, would allow fault-tolerant operation with the architecture’s requirements as they stand. The authors expect better protocols to relax the budgets by a few percentage points at most. Xanadu reported a 60% reduction in optical loss during 2025 and a 20-fold improvement over three years in its 2025 annual results. It gives no metric for either figure and has not published how they relate to Aurora’s component budgets or to the 2026 loss indicator. Outsiders therefore cannot compare the three sets of numbers directly.
A GKP-based logical qubit whose error rate falls as the outer code grows would be Xanadu’s first below-threshold result; neither Xanadu nor PsiQuantum has published one. Universal computation also requires magic states. The Aurora authors note that more work is needed to confirm that producing them does not worsen the architecture’s loss tolerance. The qLDPC codes Xanadu now plans to use require fast decoding at sizes far beyond the distance-2 repetition code decoded on Aurora. Xanadu’s 2028–2029 fault-tolerance milestone requires all three.
CRQC Implications
A cryptographically relevant quantum computer (CRQC) is a machine able to break widely deployed public-key cryptography, such as RSA or elliptic-curve keys, with Shor’s algorithm in a useful time. In the most widely discussed estimate for RSA-2048, Craig Gidney’s May 2025 paper, the attack uses fewer than a million noisy physical qubits and about 6.5 billion Toffoli gates. The algorithm itself needs about 1,400 logical qubits, plus space for computation and magic-state factories. Its expected runtime is just under a week, made up of repeated attempts of about 12 hours each. Gidney assumes a 0.1% physical error rate, a 1 µs surface-code cycle, a 10 µs reaction time and nearest-neighbor connections. I covered the paper in my analysis of the Gidney estimate.
Estimates based on qLDPC codes, which require long-range connections, are lower. Iceberg Quantum put RSA-2048 at about 100,000 physical qubits for a month of expected runtime with its Pinnacle Architecture. In its October 2026 Pinnacle 1.2 technical release, a company publication, Iceberg puts it at 72,000 qubits and an expected 3.7 days at the same 0.1% error rate and 1 µs cycle, and at 99,000 qubits and 5.0 days for fixed-connectivity superconducting systems, a variant Iceberg calls Frozen Pinnacle. In Google Quantum AI’s 2026 estimate for the elliptic curve used by Bitcoin, an attack needs 1,200 to 1,450 logical qubits and 70 to 90 million Toffoli gates, against Gidney’s 6.5 billion for RSA-2048. Some deployed elliptic-curve systems may therefore become practical targets first, although the order depends on the curve, the circuit and the hardware.
Xanadu’s targets are of the same numerical order. More than 1,000 logical qubits in 2031, at the error rates Xanadu projects, would approach the algorithmic logical-qubit count of Gidney’s construction. The up to 100,000 physical qubits it projected at its merger announcement match the scale of the latest qLDPC-based estimates. The units differ, though. A physical qubit in a Pinnacle estimate is a circuit-level qubit, while Xanadu’s are optical GKP modes. Converting one into the other requires accounting for GKP preparation, cluster-state construction, loss, decoding and magic-state supply.
Photonic machines get long-range connectivity from optical fiber. On superconducting hardware those connections have to be engineered: IBM designs long-range couplers into its chips and, in October 2026, linked two processors inside one dilution refrigerator with 80-centimeter couplers. Qubit connectivity is the CRQC capability where photonics has its clearest architectural advantage. Until Xanadu publishes a resource estimate for its own architecture, including its state factories and the rate of its logical operations, it has no RSA-breaking capability even on paper.
On Xanadu’s side, every figure in that comparison is a roadmap target. Xanadu has not published a below-threshold logical qubit, nor an RSA or elliptic-curve resource estimate for its own architecture, so its runtime for Shor’s algorithm is unknown. The Aurora authors include an illustrative sizing: a machine with 100 logical qubits at a 100-to-1 error-correction overhead and a multiplexing depth of 12 would need tens of millions of Gaussian boson sampling cells and, even with a hundredfold improvement in component density, tens of thousands of server racks. Xanadu bases its 2026 roadmap on qLDPC codes, SNSPDs and better components but has not published a revised budget. A CRQC would also need real-time decoding of large qLDPC codes and fault-tolerant operation sustained for many hours per attempt, the continuous operation capability that no one has yet shown on any platform for a computation of that length.
DARPA’s evaluation is the independent milestone to watch. On 7 October 2026 DARPA moved Atom Computing, Diraq, IBM and IonQ into Stage C, the final phase of the Quantum Benchmarking Initiative. They joined Microsoft and PsiQuantum, which entered through DARPA’s precursor program. The agency expanded PsiQuantum’s work with a US$125 million agreement in July 2026. Xanadu received a one-year extension of its Stage B work instead. DARPA says QBI “is not a competition between companies; it is not intended to pick winners and losers,” and it expects more companies to advance to Stage C in the near future.
My assessment is that Xanadu is a credible contender for a CRQC-scale machine but not a front-runner. Google has run a surface-code memory below threshold on superconducting qubits. Xanadu has not yet published a below-threshold result of any kind. If Xanadu’s components reached their loss budgets, its fiber connectivity would suit the qLDPC codes behind the newest RSA-2048 estimates. The first evidence of progress would be a below-threshold logical qubit on its own hardware. Until that result appears, I won’t change my estimate of when a CRQC arrives on the strength of Xanadu’s roadmap. Migration planners have nothing to adjust either: regulators and standards bodies have already set PQC migration deadlines that apply whichever modality gets there first.
Modality & Strengths/Trade-offs
Xanadu’s modality is photonic quantum computing in its continuous-variable, measurement-based form, with GKP qubits as the physical layer. I cover the modality in general in my photonic quantum computing explainer. The strengths and trade-offs below are specific to Xanadu’s version of it.
Strengths:
Mostly Room-Temperature Operation: In Xanadu’s architecture the sources, chips, switches, fiber delays and homodyne detectors work at room temperature, as do the passive beamsplitters that serve as entangling gates. The exception is photon-number-resolving detection, which heralds the non-Gaussian input states. Aurora’s 36 transition-edge sensors were cooled to 12 millikelvin in two dilution refrigerators. Xanadu now plans to use superconducting nanowire detectors operating near 2 kelvin, and Bluefors is developing a modular cryogenic unit for them. Cooling to about 2 kelvin is far simpler than holding a superconducting processor at millikelvin temperatures. In Xanadu’s design, only one stage of state preparation needs that cooling. A Xanadu data center will still contain cryostats.
High Qubit Connectivity: Optical modes can be routed by fiber to any other part of the machine, including modules in other racks. In Xanadu’s architecture the choice of error-correcting code is largely a question of how the fibers carrying entangled pairs are routed, so Xanadu has considerable freedom to use codes with non-local checks. Long fibers add loss and need phase stabilization. Xanadu has not published that cost for a large machine. Other platforms run qLDPC codes with their own routing and interconnect costs; IBM, for example, designs long-range couplers into its superconducting chips.
Modularity and Networking: Aurora connected four racks of independently packaged modules over fiber and ran one cluster state across them through phase- and polarization-stabilized fiber delay lines. Scaling therefore means adding modules and fibers, much as classical data centers add servers. Here Xanadu can draw on decades of telecom investment in fiber, connectors and optical components. Companies working in other modalities are building modular links as well, such as IBM’s couplers between processors inside one dilution refrigerator. Photonic qubits can travel between racks over ordinary fiber, which superconducting qubits cannot do without microwave-to-optical conversion or cryogenic links.
Semiconductor Manufacturing: Aurora was built from chips made on existing commercial lines: silicon nitride sources from a 200 mm process, switches in thin-film lithium niobate, and silicon homodyne chips from a 300 mm platform. The June 2025 GKP source came from 300 mm silicon nitride wafers. Under the GlobalFoundries agreement, Xanadu’s detectors and silicon nitride photonics will move onto a high-volume 300 mm line. Xanadu’s direct access to foundry capacity, which few quantum hardware companies have, is an advantage on engineering scale and manufacturability.
Clock Rate and Parallelism: Photonic machines process pulses in flight. Borealis ran at 6 MHz and Aurora at 1 MHz, with all 12 modes measured on every clock cycle and the decoder’s decision applied on the next one. With time multiplexing, one physical channel handles many modes in sequence. Xanadu used it in Aurora to generate a cluster state of 86.4 billion modes over two hours while the machine worked with 12 modes at a time.
Limited Idle Decoherence: Propagating photons avoid the storage constraints of stationary qubits such as electron spins or superconducting circuits. Xanadu designs its machines to create fresh qubits continuously instead of storing them for long. Photons still have to wait in fiber delay lines, where some are lost and the others pick up phase noise and polarization drift. Aurora therefore actively stabilizes the phase and polarization of its delays. Aurora’s delay modules lost about 0.28 dB, excluding connectors; Xanadu’s prototypes are below 0.1 dB.
Trade-offs:
Optical Loss and Detection Efficiency: Loss is the dominant error in Xanadu’s architecture, and it compounds multiplicatively along every optical path. Aurora’s measured losses exceeded its architecture’s budgets by wide margins. Detection is part of the same budget. Aurora’s transition-edge sensors ranged from 69.3% to 97% efficiency, against the more than 99% that the authors say the heralding paths need. Xanadu reported in-house detectors above 99% in its 2025 GKP experiment; producing thousands of them with the required speed and photon-number resolution is the remaining task. The clearest published gap is in the switches. The deepest path in the fault-tolerant design passes through as many as 15 Mach-Zehnder switches. The Aurora authors calculate that each switch can lose no more than about 7 millidecibels. Xanadu reported roughly 30 millidecibels from its optimized designs at the time.
Probabilistic State Preparation: Xanadu’s entangling gates are deterministic, but its inputs are not. Each Gaussian boson sampling cell produces a useful non-Gaussian state only when its detectors register the right photon pattern. Xanadu therefore multiplexes many cells through trees of switches and selects the best outputs. Every extra layer of switching raises the success probability and adds loss, and Xanadu designs its state factory around that trade-off. PsiQuantum faces a version of the same problem from the other side: its fusion-based architecture uses single photons and entangling measurements that succeed only with some probability, so PsiQuantum multiplexes too. Both companies are limited by loss, and neither has shown a below-threshold logical qubit.
Resource Overhead for Encoding: The GKP layer lowers the overhead of the outer code but raises the cost of each physical qubit. The integrated GKP states Xanadu published in 2025 are not yet of the quality its fault-tolerance model requires. For the state factory, Xanadu separately assumes sources with 15 dB of squeezing. A large machine needs continuous production of fresh GKP states, a magic-state supply for non-Clifford gates, and thousands of lasers, squeezers and interferometers held at stable phase.
Comparative Qubit Precision: Other modalities are further along on error correction. Google has run a surface-code memory below threshold on superconducting qubits. The Aurora authors themselves cite neutral-atom and trapped-ion experiments that produced convincing evidence of sub-threshold operation with logical gates. Xanadu has not yet published either kind of result. Xanadu is betting that once its components cross threshold, photonic machines will scale more easily than competing machines. The risk is that photonics takes longer to cross threshold than the other modalities.
In summary, Xanadu has three advantages in its version of photonics: deterministic room-temperature logic once GKP states exist, fiber connectivity well suited to qLDPC codes, and access to semiconductor foundries. Against those, it has a loss gap that the Aurora authors put at 20 to 30 times in decibel terms, probabilistic state preparation whose multiplexing adds loss of its own, and a cryogenic detector stage that is often left out when its machines are described as room-temperature. If Xanadu meets its loss budgets, its architecture could scale further and faster than monolithic platforms. If it misses them in the 2028–2030 window, companies building superconducting, neutral-atom and trapped-ion machines will extend their lead in logical qubits.
Track Record
Xanadu’s record is strongest in published science and software and thinner in demonstrated performance and revenue.
Scientific Publications: Nature has published four of Xanadu’s hardware results: X8 (2021), Borealis (2022), Aurora (2025) and the on-chip GKP source (2025). Xanadu’s researchers have added the 2025 Physical Review Letters paper on arbitrary error-correcting codes and a steady flow of algorithm work, including a May 2026 result on reducing the cost of quantum read-only memory. The authors are careful about the limits of their results: the Aurora team calls its machine a sub-performant scale model, and the GKP paper’s authors say their states are not yet fault-tolerant. Xanadu’s press materials are less careful: they describe Aurora as demonstrating “12 logical GKP qubits.” Borealis, the most widely reported result, is also the one most weakened since publication, by the 2024 classical simulation work.
Industry Firsts: Xanadu launched X8 in 2020 as what it called the first photonic quantum computer in the cloud. Borealis followed in 2022, on Xanadu’s own service and on Amazon Braket. Xanadu describes Aurora as the first modular, networked photonic quantum computer. In its June 2025 paper, Xanadu reported the first GKP states generated on an integrated chip. In March 2026 it became the first publicly listed company focused solely on photonic quantum computing.
Software: PennyLane is Xanadu’s most widely adopted product. At its August 2026 roadmap presentation Xanadu reported about 1.1 million package installs a month across the PennyLane platform family, 51 simulators and hardware devices reachable through one API, and use by 30.8% of developers in an industry survey, second to IBM’s Qiskit. These are company figures. PennyLane is hardware-agnostic and has no separate revenue line; Xanadu plans to earn from it through enterprise training, applied algorithm development, cloud services and eventual hardware sales. Strawberry Fields, its original photonic library, is now used mainly in research.
Funding and Public Listing: Xanadu raised about US$250 million in venture funding through its 2022 Series C, at a US$1 billion valuation, and about US$302 million gross from the March 2026 merger. At 30 June 2026 it held US$312.8 million in cash and cash equivalents. In its 31 August roadmap release it presented about US$686 million of “total available funding”: that cash, US$232.8 million of undrawn capacity under the equity facility, and about US$140.2 million of federal support for Project OPTIMISM (the C$195 million commitment). Only the cash is in hand. Under the facility Xanadu can sell shares when it chooses. At lower prices, each dollar it raises that way dilutes existing shareholders more. The federal support is tied to the project and, The Logic reports, repayable.
Government Programs and Partnerships: In the United States, Xanadu holds DARPA Stage B funding of up to US$15 million, the provisionally approved extension of up to US$10 million, and its four-year Air Force Research Laboratory partnership. In Canada it has the C$195 million federal commitment, the Quantum Champions Program and a 2025 selection by Canada’s defence innovation program for NORAD modernization, under which it applies quantum computing to battery research. Its manufacturing partners now include GlobalFoundries, ASML, Applied Materials, DISCO, EV Group, Tower Semiconductor, Corning and Bluefors. Application partners include Volkswagen, Toyota, Mitsubishi Chemical, Rolls-Royce, Lockheed Martin, AMD, TELUS, the University of Alberta, Canada’s National Research Council and Oak Ridge National Laboratory.
Revenue: Xanadu’s sales are small and come mostly from government agencies and algorithm co-development contracts. It reported revenue of US$4.6 million for 2025, up from US$1.6 million in 2024, with a net loss of US$70.7 million. In the second quarter of 2026, revenue was US$1.5 million and the net loss US$42.1 million, including a US$10.6 million non-cash charge for remeasuring earn-out liabilities; the adjusted EBITDA loss, a non-GAAP measure, was US$21.3 million. Third-quarter results are due on 11 November 2026.
In aggregate, Xanadu has moved in ten years from a single founder with a theoretical architecture to a public company with 264 employees at the end of 2025, a published modular prototype and a manufacturing chain that includes several of the world’s largest semiconductor companies. Its researchers have demonstrated what their papers claimed. To meet its later roadmap targets, Xanadu needs components good enough for error correction to work, which it has not yet shown. In its public statements, Xanadu sometimes puts that progress further along than its researchers do in their papers.
Challenges
Reducing Optical Loss and Detector Inefficiency: Xanadu has to reduce optical loss before it can meet anything else on its roadmap. It has to get every optical component within its loss budget and keep it there in mass production. Xanadu reports progress on several: edge couplers at 0.085 dB per facet, fiber delays below 0.1 dB in prototypes and in-house detectors above 99% efficiency. Each figure has to be judged against the budget allocated to that component and the number of such components along a path. At 0.085 dB, a single facet already loses about 1.9% of the light. Weedbrook told analysts on the company’s first-quarter 2026 call that each chip run takes two to five months, which limits how fast Xanadu can bring the loss indicator down.
Crossing the Fault-Tolerance Threshold: Once components meet their budgets, Xanadu still has to show a logical qubit whose error rate falls as the code grows, the result I would look for first in any CRQC assessment. After that it has to add magic-state production without degrading loss tolerance and decode qLDPC codes in real time at its clock rate. It also has to keep the machine phase-stable across kilometers of fiber and thousands of interferometers for hours at a time. Xanadu has not yet run error correction at scale. Some error mechanisms may appear only in larger systems.
Scaling from Prototype to Data Center: Xanadu’s plan for scale is industrial: the Inception facility in Toronto, GlobalFoundries production for detectors and silicon nitride chips, ASML and EV Group process work, and Bluefors cryogenic modules. Each of these is at an early stage. Fit-out of Inception begins in early 2027, The Logic reports. The 2026–2027 “qubit factory” is the first test of whether Xanadu can make GKP sources at volume. Xanadu has published neither its specifications nor a system budget to replace the illustrative one in the Aurora paper. Control electronics, laser distribution, phase stabilization and classical decoding all grow with the machine too. Engineers building large systems often find bottlenecks in those subsystems that never showed up in prototypes.
Competition and Technological Uncertainties: PsiQuantum, pursuing a discrete-variable photonic architecture, is better funded, already in DARPA’s Stage C and building utility-scale sites in Brisbane and Chicago on GlobalFoundries silicon. Quandela, ORCA Computing and QuiX Quantum compete in European photonic procurement. Among other modalities, IBM targets a 200-logical-qubit fault-tolerant system in 2029. Neutral-atom companies are reporting rapid progress in logical qubits. Trapped-ion companies report some of the highest two-qubit gate fidelities. If any of them demonstrates large-scale error correction well before photonic machines cross threshold, customers and investors will care less about Xanadu’s argument that photonics scales more easily.
Market and Financing Pressures: Since its listing, Xanadu has faced public-market scrutiny. Its Nasdaq shares reached US$42.44 shortly after listing and closed at US$2.62 on 4 November 2026, below the US$10 price paid by private-placement investors. Over the same period Xanadu registered about 294 million existing shares for resale in April. The six-month lockup on pre-merger shareholders expired on 22 September. Operating losses are widening as Xanadu spends more on engineering. The C$893 million Project OPTIMISM depends partly on Ontario funding that had not been agreed when the federal commitment was signed. Drawing on the equity facility at lower share prices means more dilution for existing shareholders. Management says Xanadu has enough capital to fund operations through its near-term milestones. Xanadu has not published what a data center at the scale of its 2029–2030 targets will cost.
Talent and Operational Complexity: At the end of 2025 Xanadu reported 264 employees, 152 of them physicists and engineers. It is now hiring in Toronto and Albany, New York, and has been appointing the executives a public company needs. The move from research lab to manufacturer is the harder transition. Running a packaging and assembly facility, qualifying suppliers and managing yield across several foundries are different skills from publishing in Nature. Xanadu also has to compete with large technology companies for the same photonics and error-correction specialists.
In summary, Xanadu’s challenges have become narrower and more concrete since 2025. Xanadu showed modular integration across four racks with Aurora. Yield, calibration and reliability at larger scale have yet to be tested. The other open problems are component loss, below-threshold error correction, and the capital and manufacturing scale to build whatever state factory Xanadu’s final design requires. Xanadu has published a measurable indicator for the first, which I will track against the 24.1 baseline. Outsiders could check the remainder of the roadmap if Xanadu published five things:
- the definition and history of the loss indicator, including how the requirement is set and how the 17 components are weighted
- logical error rates with their units and the code sizes behind them
- the quality and production rate of the qubit factory’s GKP states
- a revised state-factory and cryogenic budget for the 2029–2031 machines
- a resource estimate for an application, such as factoring RSA-2048, on Xanadu’s own architecture
Third-quarter results on 11 November, the qubit factory and any move into DARPA’s Stage C are the next public checkpoints for the 2028–2029 fault-tolerance milestone.
