Who Really Wins If Photonics Wins

In February 2025, PsiQuantum published A manufacturable platform for photonic quantum computing in Nature, describing single-photon sources, superconducting detectors, nanosecond optical switches, and low-loss waveguides fabricated together on 300-millimeter silicon wafers at GlobalFoundries’ Fab 8 in Malta, New York. Those wafers came off a production line that also makes chips for phones and cars, through more than 20 photolithography levels, under the process control and yield tracking a foundry applies to any paying customer.

That is the photonic wager, stated as an engineering result. Superconducting and trapped-ion machines scale by solving problems with almost no industrial precedent, one dilution refrigerator or one vacuum chamber at a time. Photonic machines are meant to scale the way the semiconductor industry scales everything else, by printing them.

If the wager pays off, much of the gain goes to companies outside the quantum industry. Firms that have spent decades building photonic and telecom infrastructure would supply a large share of the machine, and they will keep building that infrastructure whether or not fault-tolerant quantum computing ever arrives. Working out which of them actually gains, and by how much, takes a layer-by-layer look at what a photonic quantum computer is made of and at which layers are bought rather than invented.

The map below is the one we use with learners in our modality and supply chain material. It works upward from the components, then asks where value concentrates. We’ve written it for someone assessing the photonic bet from outside the lab, and it covers technology and market structure rather than investment advice.

The Four Things Every Photonic Machine Needs

Photonic quantum computers differ from one another more than superconducting machines do, so the vocabulary has to be fixed before the supply chain makes any sense.

Some designs encode information in properties of individual photons, such as polarization, arrival time, or which waveguide the photon travels down. A waveguide is a channel etched into a chip that confines light the way a copper trace confines current. PsiQuantum, Quandela, and ORCA Computing all work this way. Other designs encode information in the amplitude and phase of a light field rather than in countable particles. Xanadu built its machines on that approach, using squeezed light, meaning light prepared so that fluctuations in one measurable property fall below the ordinary vacuum level, paid for with larger fluctuations in the paired property.

The computational models differ too. PsiQuantum uses fusion-based computation. Small entangled clusters of photons are generated and then joined together by measurements, and the answer comes out of the measurement record rather than out of a sequence of gates. Xanadu combines measurement-based computing with GKP qubits, named for Gottesman, Kitaev, and Preskill, an encoding that spreads one logical qubit across a continuous light field so that small errors can be spotted and corrected. QuiX Quantum and Quandela build gate-based photonic processors, closer to the circuit model most people learn first.

Underneath the architectural differences, the parts list is nearly identical. Four component families do the work.

Photon sources produce the quantum light, either single photons on demand, heralded photon pairs, or squeezed states.

Photonic integrated circuits (PICs) route and interfere that light. A PIC is a chip full of waveguides, splitters, and phase shifters, and the interference patterns it produces are the quantum operations.

Optical switches send photons down one path or another in response to a measurement result. Feed-forward logic and active error correction both depend on them.

Single-photon detectors register each arriving photon. Almost every serious machine uses a superconducting nanowire single-photon detector (SNSPD).

Two supporting layers surround these four. Optical fiber carries photons between chips, modules, and racks. Control electronics generate the timing, drive the switches, read the detectors, and close the feed-forward loop in nanoseconds. And the detectors have to be cooled, so a photonic machine isn’t a room-temperature machine end to end, whatever the brochure says.

What the Foundry Actually Supplies

Silicon photonics is a real industry with real volume, and it exists because data centers need optical interconnect. Three waveguide platforms carry most of the quantum work. Silicon-on-insulator is compact and fully CMOS-compatible, and it’s what PsiQuantum uses at GlobalFoundries. Silicon nitride loses far less light per centimeter and works across a wider band, and it’s the platform at Xanadu, QuiX Quantum, and LioniX International. Indium phosphide and thin-film lithium niobate fill the gaps silicon can’t cover, supplying lasers and the fastest modulators.

A foundry that already prints low-loss waveguides, splitters, and phase shifters on 300-millimeter wafers needs very little persuading to print them for a quantum customer. The design rules overlap, the metrology overlaps, and the yield methodology overlaps. This is the genuine inheritance, and it’s also the layer with the least technical risk anywhere in the machine.

The rest of the parts list is a different story. None of the other three component families has a high-volume classical twin, and every one of them has to be added to the foundry flow as something the fab has never made before.

The Three Components With No Classical Twin

Photon sources. PsiQuantum generates heralded photons by four-wave mixing in silicon waveguides at 1550 nm. Detecting one photon of a pair confirms that its partner exists and marks the instant it becomes available. The process is spontaneous, so the timing isn’t under the operator’s control, and the machine has to run many sources in parallel and switch the successful ones into the circuit. The compensating advantage is large: the source prints on the same wafer as everything else.

Quantum dots emit on demand instead. A quantum dot is a nanometer-scale island of one semiconductor embedded in another, typically indium arsenide in gallium arsenide, and a laser pulse makes it release a single photon with very high purity. Quandela, Sparrow Quantum, and Aegiq all build on this. The catch sits upstream. Quantum dots are grown by molecular beam epitaxy (MBE), a process that deposits material one atomic layer at a time in ultra-high vacuum, and the tools cost several million dollars each. Turning epitaxial wafers into usable emitters then takes electron-beam lithography and cavity fabrication. Nothing in that sequence resembles high-volume CMOS.

Xanadu’s squeezed-light sources use optical parametric amplification driven by a classical pump laser. The underlying nonlinear optics is well established. The difficulty is loss, since every fraction of a decibel degrades the squeezing that the whole architecture is built on.

Fast, low-loss switches. A switch in a measurement-based photonic computer has to change state in nanoseconds, add almost no loss, and be manufacturable by the million. PsiQuantum’s answer is barium titanate, a ferroelectric with an unusually strong electro-optic response, grown as thin films on silicon and bonded to the photonic wafer in a separate step. Barium titanate isn’t a standard foundry material, and PsiQuantum had to develop the growth and bonding processes itself. Xanadu and QuiX Quantum mostly use thermo-optic switches instead. A thermo-optic switch changes the refractive index of a waveguide by heating it, and it’s far easier to make and roughly a thousand times slower. Their architectures are designed around that difference.

Detectors. An SNSPD is a superconducting strip a few nanometers thick and about 100 nanometers wide, cooled below its critical temperature and biased just under its critical current. A single photon breaks superconductivity locally and produces a voltage pulse, and the strip recovers in nanoseconds. Commercial devices deliver detection efficiency above 90 percent at 1550 nm, with timing jitter well under 100 picoseconds and effectively no dark counts. No other detector technology comes close on that combination.

PsiQuantum builds its detectors into the wafer, depositing niobium nitride films inside the foundry flow. That removes the fiber coupling losses discrete detectors introduce, and it is the single most impressive manufacturing claim in Omega, the chipset the Nature paper describes.

Everyone else buys. The commercial SNSPD supply is short. ID Quantique in Geneva, now owned by IonQ, is the most established vendor. Single Quantum in Delft and Pixel Photonics in Münster serve much of the European market, with Pixel Photonics specializing in detectors integrated directly onto photonic chips. Photon Spot in California and PhoTec in Shanghai cover the United States and China. That is close to the entire list.

The cooling itself is unremarkable. SNSPDs run at 1–4 Kelvin, not the 10–15 millikelvin a superconducting qubit needs, so the job goes to standard closed-cycle cryocoolers from Sumitomo, Oxford Instruments, or Cryomech, now part of Bluefors. PsiQuantum has gone further and designed liquid-helium-cooled cabinets holding hundreds of chips, closer in spirit to data center liquid cooling than to a dilution refrigerator.

Put the three together and the semiconductor inheritance looks narrower than the headline suggests. The foundry contributes the layer with the deepest existing process base. The three layers that will decide whether a photonic machine reaches fault tolerance are all additions, and two of them, quantum-dot growth and superconducting film deposition, come from outside the silicon photonics tradition entirely.

Five Companies, Five Supply Chains

PsiQuantum is the most vertically committed. Chips come from one foundry, barium titanate growth is in-house, detectors are on the wafer, and cooling is a bespoke rack design. The company has skipped near-term demonstrations altogether and is building directly for fault tolerance, with utility-scale sites under development in Brisbane and Chicago. If it succeeds, remarkably few outside suppliers get rich alongside it.

Xanadu in Toronto took the opposite structural path. Aurora, its prototype, linked 35 photonic chips across four server racks with 13 kilometers of optical fiber to run a networked system of 12 physical qubits. The photonic processing runs at room temperature and the detectors don’t, and the room-temperature claim usually gets repeated without that second half. Xanadu builds on silicon nitride, has invested in its own packaging capability, and maintains PennyLane, an open-source software framework used well beyond its own hardware.

Quandela in France is the clearest counterexample to the monolithic thesis. Its machines are built around deterministic quantum dot sources, which means its supply chain runs through MBE growth and nanofabrication rather than through a CMOS foundry. Its current commercial system offers 12 physical qubits, and its Perceval SDK is used across the research community.

ORCA Computing in London uses time-bin encoding, where the information sits in which time slot a photon arrives in, combined with quantum memory to buffer photons between operations. Its systems are rack-mounted and built from telecom-standard parts, and it has delivered to UK defense and research customers. ORCA buys much of its photonic front end from European specialists rather than making it.

QuiX Quantum in Enschede builds on LioniX International’s silicon nitride platform and sells processors to research institutions including the German Aerospace Center. It positions itself as a component supplier rather than a full-stack system vendor, which is the photonic analogue of a chip company selling into other people’s machines.

Read those five together and a pattern shows up that matters for anyone estimating spillover. The more of the machine a company prints on one wafer, the more of the supply chain it internalizes, and the fewer outside firms benefit if it wins. The more a company buys, the wider the group of suppliers that gains. PsiQuantum winning and Quandela winning are not the same event for anybody else in the industry.

The Telecom Inheritance

Most photonic architectures operate at 1550 nm, the wavelength the world’s fiber networks already use. Photonic modules can therefore be connected with ordinary telecom fiber, with no frequency conversion step in between. Superconducting machines have no equivalent option, since their qubits live at microwave frequencies and converting microwave to optical signals efficiently remains unsolved.

This is the strongest structural advantage the modality has. Scaling out means adding racks and fiber rather than enlarging a single cryostat, and Xanadu’s 13-kilometer demonstration was designed to prove exactly that point. It also makes photonic systems the easiest quantum hardware to imagine inside an existing data center.

It doesn’t follow that fiber suppliers are winners. Corning, Prysmian, and Furukawa Electric ship hundreds of millions of kilometers of fiber a year for the internet. Quantum computing will not register in those volumes this decade or probably next. The inheritance lowers the cost of scaling photonic machines without creating meaningful new demand for the people who make fiber.

And the advantage is architectural rather than proven. Connecting two modules with fiber is straightforward. Generating entanglement between them fast enough to support fault-tolerant computation is the open engineering question, and rate is the number to watch, not distance.

Where the Value Actually Concentrates

Some of you will disagree with the ranking below, and the disagreement usually comes down to how much weight you give the AI interconnect market. Here is how we’d order it.

The system builders carry the most exposure and the most risk. PsiQuantum, Xanadu, Quandela, ORCA, QuiX, and Photonic Inc., the Canadian company working with silicon T-centers, defects in silicon that store a spin qubit and emit a matching telecom photon, are the firms whose value is directly tied to the outcome. They’re almost all private, and listing status changes fast enough that it should be checked against current filings rather than against any article.

Specialist component suppliers are where concentration is highest. SNSPD makers, quantum dot source companies, and the ultra-low-loss silicon nitride foundries have small revenue bases and no large classical business to fall back on. A photonic win would move them by multiples. Almost all of them are private, and a striking share are European: Single Quantum, Pixel Photonics, Sparrow Quantum, Aegiq, LioniX International, Ligentec, ID Quantique. This is the tier most winner maps skip, and it’s the one where the outcome would be felt hardest.

Silicon photonics foundries are involved but barely exposed. GlobalFoundries manufactures for PsiQuantum, and Tower Semiconductor and IMEC serve adjacent quantum work. All three are scaling silicon photonics to serve AI data center demand, and that demand dwarfs anything the quantum industry will order this decade. Buying a large foundry for photonic quantum exposure gets you AI optical interconnect with a lottery ticket stapled to it, and the ticket is priced at approximately zero for a reason.

Broad photonics and cryogenics suppliers gain directionally and unmeasurably. Coherent, Thorlabs, Gooch & Housego, Sumitomo, Oxford Instruments, and Bluefors will all sell more parts into a growing photonic quantum market. None of them will notice it in a quarterly result for a long time.

There is a fifth beneficiary that doesn’t appear on supply chain maps at all, which is people. Photonic packaging engineers, PIC designers, cryogenic technicians, and MBE growth specialists are scarce today and would be scarcer still if two or three utility-scale photonic sites came online at once. Hiring pressure is one of the earliest observable signals that a modality is moving from research into production.

Sovereignty Follows the Photonics Industry

Because photonic quantum computers are printed in semiconductor fabs, a country’s position in this modality is largely a function of the photonics industry it already has. That makes the sovereignty picture different from superconducting, where the constraint is a handful of cryogenics vendors, and different again from trapped ions, where it’s precision lasers and vacuum systems.

The Netherlands has an unusually strong hand: QuiX Quantum, LioniX International, Single Quantum, the PhotonDelta public-private program, and ASML supplying the lithography underneath all of it. Belgium contributes IMEC as a shared pilot line for photonic integrated circuits. France has Quandela and the quantum dot research base at CNRS. Switzerland has Ligentec and ID Quantique. Germany has Pixel Photonics and a national program tying photonics to quantum technology. Taken together, Europe is more competitive in photonic quantum computing than in any other modality.

The United States has the foundry relationship that matters most, along with the largest capital pool. China has a substantial domestic photonics industry, its own SNSPD manufacturing, and the Jiuzhang experiments at USTC, which detected up to 255 photons in Gaussian boson sampling. That task is a sampling problem chosen because classical computers find it hard, not because anyone needs the output. Australia has bought in rather than built up, hosting and co-funding a PsiQuantum site instead of developing capability from the component level. Canada is doing both, hosting that kind of investment while backing domestic capability in Xanadu, which runs its own packaging facility.

The chokepoint is not the fab. Silicon photonics capacity exists in the United States, Europe, and Asia, and more is being built for reasons that have nothing to do with quantum. The narrow points are quantum-dot growth capability, SNSPD supply, and photonic packaging, all of which sit with a small number of firms in a small number of countries. A national quantum strategy that funds fab access while ignoring those three has funded the easy part.

The same fabs that would print quantum photonic wafers are being filled by AI interconnect orders from customers who buy in far greater volume. Near term this helps, since it pays for process development that quantum programs then inherit. Longer term, a capacity crunch would not be resolved in the quantum customer’s favor.

What Would Break the Thesis

Loss has not been beaten. Every photon absorbed or scattered before detection is a failed operation, and losses compound across millions of components. The Nature paper reports component-level fidelities for state preparation and for fusion, and those are conditional on the photon being detected in the first place. The end-to-end loss budget for a full-scale machine has not been demonstrated by anyone, and it is the number that decides the modality.

Photonic two-qubit gates are probabilistic by nature. In linear optics, the interaction that entangles two photons succeeds only some of the time. This is physics rather than fabrication, so no amount of manufacturing skill removes it. Fusion-based computation and GKP encoding are both designed to work around it with redundancy, and the overhead they require is a modeling result, not a measured one. If the real overhead is higher than projected, the machine needs more chips, more wafers, and more fab capacity than any roadmap currently assumes.

Another modality could get to fault tolerance first. Neutral atoms in particular have scaled quickly, and superconducting and trapped-ion programs are running error correction experiments today. Much of the photonic supply chain is modality-specific. SNSPD lines, barium titanate growth capacity, and quantum dot fabs do not transfer easily to another winner.

No photonic machine has yet done anything useful that a classical computer can’t. The Jiuzhang sampling results and their successors are demonstrations of hardness, not of utility. PsiQuantum’s decision to skip intermediate demonstrations is defensible on engineering grounds and it leaves the company with no system-level proof point, only component results.

How to Read a Photonic Roadmap

We give learners a short set of questions for exactly this situation, and they work on a vendor pitch as well as on a press release.

Ask which layer the claim is about. A foundry milestone, a source milestone, a switch milestone, and a detector milestone are four different things, and the first is by far the easiest.

Ask whether the fidelity numbers include loss. Component fidelities conditional on detection are legitimate and incomplete. The end-to-end number is what determines resource overhead.

Ask whether the qubits are physical or logical, and whether the system was demonstrated or announced. A 12-qubit photonic prototype demonstrating rack-to-rack networking is a genuine result, and it is not a compute milestone. Both statements are true at once.

Ask what fraction of the machine is bought. A company that buys sources and detectors has shorter development risk and longer supplier risk. A company that prints everything has the reverse. Neither is better, and they fail differently.

Ask how many suppliers exist for the hardest part. For most photonic architectures, that answer is somewhere between one and five, and it usually isn’t the foundry.

Ask about entanglement rate, not fiber length. Modularity claims live or die on how fast two modules can be entangled, so a demonstration over 13 kilometers tells you the link works and nothing about whether it’s fast enough.

The Short Version

The photonic wager is a manufacturing wager, so judging it is mostly a supply chain skill. You need to know which layer a claim refers to, whether a number is physical or logical, whether a system was demonstrated or announced, and how many firms could supply the part that would be hardest to replace. None of that needs a physics degree, and all of it can be learned.

Our modality and supply chain material at Quantum Academy applies this same layer-by-layer method to photonic, superconducting, trapped-ion, and neutral-atom systems, so that the comparison is structural rather than anecdotal. If the people side is what interests you, the roles this industry is hiring for are mapped at QuantumCareers.com. And for a deeper technical treatment of the photonic stack, including the architectures we’ve only sketched here, the extended analysis at PostQuantum.com goes several layers further down.