Who Wins If Silicon Spin Wins

Natural silicon contains about 4.7 percent silicon-29, and that single number decides whether a silicon spin quantum computer works at all.

Silicon-29 nuclei have a magnetic moment. Silicon-28 and silicon-30, which make up the rest, have none. An electron trapped in ordinary silicon therefore sits inside a bath of randomly oriented nuclear magnets, and their slow fluctuations scramble its spin within microseconds. Strip the silicon-29 out and measured coherence times – the length of time a qubit holds its quantum state before noise destroys it – stretch by orders of magnitude. Enriched silicon is the first line on the bill of materials, and no amount of clever engineering substitutes for it.

In 2025, the Australian company Diraq and the Belgian research institute imec reported two-qubit gate fidelities above 99 percent on silicon quantum dot devices made on imec’s 300-millimeter industrial wafer line. Gate fidelity is the probability that an operation does what it was asked to do, and 99 percent is roughly the level at which error correction schemes begin to look workable. The devices were pulled from the production run rather than selected after testing. A quantum processor had been built the way ordinary processors are built.

At Quantum Academy we teach silicon spin as a worked case, and not because it leads on qubit count. It doesn’t. We teach it because it inverts the analysis everyone learns first.

Ask who supplies a trapped-ion machine and you get a short list: trap fabricators, narrow-linewidth laser makers, ultra-high-vacuum specialists. Ask about neutral atoms and you get spatial light modulators and high-power laser systems. Ask about superconducting qubits and the answer opens with four or five dilution refrigerator manufacturers. Short lists, specialist firms, most of them small enough that a single order changes their year.

Ask the same question about silicon spin and the honest answer is that most of the supply chain already exists, already has customers, and doesn’t need quantum computing to survive. That shifts the investor’s question from who is the new supplier to which existing supplier gains a line of business, and would that line be large enough for anyone to notice.

This is a technology and market analysis. It is not investment advice, and nothing below is a recommendation on any security.

We’ll go through the machine layer by layer, name who supplies each layer today, say where the chokepoints are, and finish with the five questions we’d apply to the next modality that comes along.

What is actually inside the machine

A silicon spin qubit stores information in the spin of a single electron. Spin is an intrinsic property, and for one electron it takes one of two values – conventionally “up” and “down” – which serve as the 0 and 1 of the qubit, along with the superpositions of both that quantum mechanics allows.

The electron is held in a quantum dot: a trap perhaps 30 to 100 nanometers across, formed by applying carefully chosen voltages to metal gate electrodes patterned on top of a silicon wafer. That description should sound familiar. It is a transistor holding exactly one electron. Bring two dots close enough and their electrons interact through exchange coupling, which is what a two-qubit gate uses.

Three families sit under the silicon spin heading, and they diverge sharply at the fabrication layer.

Gate-defined quantum dots trap electrons using patterned electrodes on either a silicon metal-oxide-semiconductor structure or a silicon-germanium heterostructure. Diraq, Quantum Motion, Quobly, SemiQon, Intel and imec all work here. Fabrication resembles standard transistor manufacturing more closely than anything else in quantum computing.

Donor-based qubits encode information in the electron and nuclear spins of individual phosphorus atoms placed inside the silicon lattice. Silicon Quantum Computing, founded by Michelle Simmons in Sydney, positions atoms one at a time using scanning tunneling microscope lithography, a technique that writes patterns by moving a single atomically sharp probe across a surface rather than by projecting light through a mask. Different physics, and a completely different factory.

Encoded spin qubits use three or more electron spins together to define a single logical unit that is naturally shielded from some kinds of noise. HRL Laboratories in Malibu, owned jointly by Boeing and General Motors, published universal logic with encoded qubits in silicon-germanium in 2023.

Whichever family, a working system needs six things: an enriched silicon substrate, a fabricated processor chip, cryogenic cooling, control electronics near the cold stage, room-temperature control instruments, and classical computing for calibration and decoding. Four of those six layers have interesting supply chains. We’ll take them in order.

The silicon-28 problem

The historical supply of ultra-pure silicon-28 came from metrology, not computing. The Avogadro Project spent years manufacturing near-perfect spheres of isotopically pure silicon so that the Avogadro constant could be measured precisely enough to help redefine the kilogram, which happened in 2019. That effort produced kilogram quantities of silicon enriched well past 99.99 percent, using gas centrifuge cascades adapted from uranium enrichment. The enrichment work was concentrated in Russia.

For a research group building a nine-qubit device, kilogram quantities are generous. For an industry, they’re a rounding error. And the geopolitical concentration is unworkable for a technology that governments treat as strategic.

Two commercial suppliers have moved into the gap. ASP Isotopes, listed on Nasdaq, began commercial silicon-28 production at a facility in Pretoria using an aerodynamic separation process applied directly to silane gas, which removes a chemical conversion step that centrifuge routes require. Silex Systems in Australia has publicly described a laser-based silicon-28 program, which gives the Australian companies a domestic source for the one material they cannot do without.

The material is only half the problem. Quantum chips don’t use bulk enriched silicon. They use a thin epitaxial layer, typically well under a micrometer thick, grown by chemical vapor deposition onto a conventional wafer. That layer has to be uniform across the whole wafer. If the residual silicon-29 concentration varies from one side of a 300-millimeter wafer to the other, qubits on one side will behave differently from qubits on the other, and every device will need its own calibration. Uniform enriched epitaxy at 300 millimeters is a young capability with very few practitioners.

The isotope layer is the weakest part of this supply chain today, and it isn’t close. Two visible commercial suppliers, output measured in tens of kilograms, and a downstream epitaxy step that almost nobody has industrialized. The constraint is not geology. Silicon is the second most abundant element in the crust and separation physics is well understood. The constraint is that nobody has yet had a reason to build capacity, so the capacity doesn’t exist.

That is a very different risk profile from helium-3, which is genuinely scarce and cannot be manufactured to order. Enrichment capacity responds to contracts. It just responds slowly, over the several years a separation plant takes to design, permit and commission.

What transfers from a chip fab, and what doesn’t

The claim that silicon spin qubits can be made in existing chip factories is true and routinely overstated at the same time.

The true half: CMOS – complementary metal-oxide-semiconductor, the process family behind essentially every logic chip made since the 1980s – already produces exactly the structures a gate-defined spin qubit needs. Metal gates over a thin oxide over a silicon channel, patterned by lithography, deposition and etch. Andrew Dzurak’s group at UNSW built the first two-qubit logic gate in silicon in 2015 using a device that a process engineer would recognize immediately. Dzurak went on to found Diraq. When Quantum Motion, a spinout from University College London and Oxford, delivered a full-stack quantum computer built on standard CMOS fabrication to the UK’s National Quantum Computing Centre in 2025, the processors came off a commercial 300-millimeter line.

The overstated half: you cannot point an existing logic line at a quantum design and get working qubits. Four things have to change.

• The channel must be enriched silicon-28, which means custom epitaxy rather than standard wafer supply.
• The silicon-oxide interface must be far cleaner than logic requires, since defects there generate the charge noise that dephases qubits. Getting there took imec most of a decade of process work.
• Gate dimensions and the spacing between dots set the coupling strength between qubits, so overlay accuracy matters more than it does for a transistor that only has to switch.
• The thermal budget through the process flow has to preserve isotopic purity and avoid introducing defects.

Each of those is an engineering program, not a setting. Together they explain why quantum-capable process development sits with a handful of organizations rather than every fab on earth.

Who those organizations are:

imec operates the most advanced openly accessible 300-millimeter spin qubit process in the world, and its role is closer to a process development house than a merchant foundry. The Diraq fidelity results came off imec’s flow.

Intel has the largest manufacturing footprint of anyone working on spin qubits and the smallest demonstrated device count among the serious players. Its 12-qubit Tunnel Falls chip, made on a 300-millimeter production line, was distributed to university research groups rather than sold. Intel treats quantum as one program among many, with the process discipline that implies and the urgency that implies too.

STMicroelectronics and Quobly have announced a partnership to adapt ST’s 28-nanometer fully depleted silicon-on-insulator (FD-SOI) process at Crolles for quantum processors. FD-SOI puts a thin insulating layer under the transistor channel and adds a back gate underneath, which gives designers a second control knob. Quobly, founded out of CEA-Leti in Grenoble by Maud Vinet and Tristan Meunier, is betting that the same platform can host qubits and their control circuits on one die.

GlobalFoundries has been named by Diraq as a manufacturing partner, which fits Diraq’s stated strategy of renting existing capacity rather than building any.

The donor camp sits outside all of this. Silicon Quantum Computing places phosphorus atoms with a scanning tunneling microscope and manufactures its own chips end to end. No photolithography, no foundry, no wafer supply agreement beyond the substrate. It is the most vertically integrated model in the field and the one least able to borrow anyone else’s capacity.

The parallel worth developing

In 1987 Morris Chang founded TSMC on a proposition the industry found strange at the time: that designing chips and manufacturing them were separate businesses, and that a company doing only the second could serve everyone doing only the first. The split created two distinct pools of value. Fabless designers captured the margin on architecture and market access. The foundry captured the margin on yield, capital efficiency and process leadership. Neither could exist without the other, and the returns to each moved on different cycles.

Gate-defined silicon spin is set up to repeat that split. Diraq, Quantum Motion and Quobly are, functionally, fabless quantum companies. imec, ST and GlobalFoundries are the foundry side. If the modality scales, an investor gets two different exposures with two different risk profiles, and only one of them requires picking the winning quantum computer company.

Silicon Quantum Computing is the integrated device manufacturer in this analogy, in the mold of Intel before the split – controlling everything, capturing everything, and carrying the full capital burden alone.

The cryogenics branch point

Silicon spin qubits need to be cold. How cold is the question, and the field is genuinely split on the answer.

Most gate-defined devices today run between 10 and 100 millikelvin. At those temperatures the only option is a dilution refrigerator, a machine that reaches millikelvin by circulating a mixture of helium-3 and helium-4 and exploiting the heat absorbed when helium-3 crosses between two phases. Four or five companies build them at research grade – Bluefors, Oxford Instruments, Leiden Cryogenics and Maybell Quantum among them – and helium-3 supply is genuinely constrained, since it is produced mainly as a decay product of tritium held in national stockpiles.

So far, that’s the superconducting supply chain with a different chip in it.

The divergence comes from a result Andrew Dzurak’s group published in 2020, showing a silicon quantum processor unit cell operating above 1 kelvin. Others have since pushed higher-temperature operation further. The engineering consequence is large, because the cooling power available in a dilution refrigerator scales roughly with the square of temperature. Moving the operating point from tens of millikelvin to 1 kelvin buys several orders of magnitude of thermal headroom.

At 1 kelvin you no longer need a dilution refrigerator. A cryocooler – a Gifford-McMahon or pulse tube machine of the kind already produced in volume for MRI scanners, semiconductor tools and defense systems – will do. That market has more suppliers, established industrial capacity, and companies like Sumitomo Heavy Industries and Cryomech, now part of Bluefors, already competing in it.

Two caveats keep this from being a clean win. The higher-temperature demonstrations belong mostly to the metal-oxide-semiconductor quantum dot camp; donor qubits and many silicon-germanium systems still want millikelvin. And a demonstration of a unit cell above 1 kelvin is not the same as a full processor with thousands of qubits running there.

The read: silicon spin currently shares superconducting’s cryogenic dependency and has a credible route out of it that no other gate-based modality can currently claim. Anyone modeling this supply chain should carry both branches, because which one prevails changes the cost structure of a machine by an order of magnitude and changes which suppliers benefit entirely.

The wiring wall

Every quantum computing approach eventually meets the same problem. Each qubit needs control lines and readout lines running from room-temperature electronics down into the cold stage, every one of those lines conducts heat inward, and the cooling budget runs out long before the qubit count reaches anything useful. Superconducting systems are already fighting this at a few hundred qubits.

Silicon spin has an answer the others don’t: put the control electronics on the same chip as the qubits.

Because spin qubits are made in CMOS-compatible processes, the transistors that control them can in principle be fabricated alongside them. That would replace thousands of coaxial cables with on-chip signal generation and multiplexing, which is the difference between an architecture that stops at hundreds of qubits and one that reaches millions.

Cryo-CMOS is the name for this: conventional transistor circuits designed to operate at cryogenic temperatures. Getting there is harder than it sounds. Threshold voltages shift when you cool silicon to 4 kelvin, carrier mobility changes, noise behaves differently, and the power a circuit dissipates has to fit inside a cooling budget measured in milliwatts. Standard design libraries are characterized at room temperature and simply don’t describe the devices anymore.

Three efforts are worth tracking. Quobly and CEA-Leti have presented cryogenic readout circuits in FD-SOI that multiplex several qubits through a single amplifier, which is the direct attack on the cable count. SemiQon, a spinout from Finland’s VTT operating out of the Micronova facility in Espoo, designs transistors for cryogenic operation from first principles rather than adapting room-temperature designs, and sells components into the industry instead of building complete machines. Intel’s Horse Ridge control chip, fabricated in a 22-nanometer FinFET process and demonstrated at around 4 kelvin, was the first serious signal that a large manufacturer intended to solve this in silicon.

SemiQon’s business model deserves a note for anyone building a market map. Selling cryogenic components to every quantum computer builder is a different bet from building a quantum computer, and cryogenic electronics have customers in space systems and detector instrumentation regardless of what happens to quantum computing. That combination – modality-wide exposure plus a non-quantum revenue floor – is rare in this industry and usually undervalued in modality-level analysis.

Announced, demonstrated, and the distance between them

A physical qubit is one piece of hardware holding one quantum state. A logical qubit is an error-corrected unit assembled from many physical qubits, protected well enough that its error rate falls below the error rate of its parts. The ratio between them for a useful machine is expected to be in the hundreds or thousands to one, depending on the error rate and the code.

Every qubit count in this article is physical. Published error correction work on spin platforms remains at the level of small codes on a handful of qubits, and no below-threshold logical qubit has been demonstrated on silicon spin. Diraq’s target of a million qubits on a chip, Quobly’s target of a hundred-qubit device, and every roadmap date any of these companies has published are announcements, not results. They may well be met. They have not been met yet.

The most useful independent filter available is DARPA’s Quantum Benchmarking Initiative, which exists specifically to test whether a company’s own timeline matches what outside reviewers can verify. Diraq, Quantum Motion and Silicon Quantum Computing have all advanced to Stage B, which tells you these programs cleared a bar set by someone with no stake in the outcome. It does not tell you they’ll deliver.

Who wins if silicon spin wins

The winners who are already winning

If quantum processors can be made on 300-millimeter CMOS lines, then the capital already sunk into semiconductor manufacturing becomes dual-use, and the equipment suppliers gain a new process module rather than a new industry. ASML, Applied Materials, Lam Research, KLA and Tokyo Electron would all see incremental demand for quantum-specific process development.

Quantum chip volumes are invisible next to logic and memory production, so nobody’s revenue line moves. What moves is optionality: a customer relationship, a process recipe, and a position in a business that might be large in fifteen years.

The foundries and process houses that develop quantum-capable flows – imec, GlobalFoundries, STMicroelectronics, and any of the large merchant foundries that choose to enter – capture direct revenue and, more valuably, the learning curve. This is the layer where the fabless parallel pays off for investors, since it offers exposure to the modality without requiring a view on which quantum company wins.

Classical computing companies benefit in an unexpected way. Tuning a quantum dot array is brutally hard: each dot needs its gate voltages set precisely, and the correct settings interact nonlinearly across the whole array. Groups working on gate-defined qubits have turned to machine learning to automate tuning, and the classical compute demand for calibration grows with qubit count rather than staying flat. Nvidia’s partnerships across several silicon spin companies reflect the same recognition.

The winners who don’t exist yet

The isotope layer is where a genuinely new market would form. Enriched silicon-28 is a small business today – low tens of millions of dollars at most, spread across metrology, research and early quantum demand. A silicon spin industry at manufacturing scale would grow that by one to two orders of magnitude, and the suppliers would be building capacity into a demand curve nobody can yet forecast. ASP Isotopes is the visible pure play. Silex Systems offers a second technical route. National laboratories with separation capability constitute a third.

Cryogenic electronics is the other place where new companies fit. SemiQon’s component model has no incumbent to displace, since nobody was making transistors characterized for 1 kelvin operation in volume.

And the cryogenics answer depends entirely on which temperature branch prevails. Stay at millikelvin and the same four or five dilution refrigerator makers benefit, along with the helium-3 constraint that comes with them. Move to 1 kelvin and demand shifts to cryocooler manufacturers with real industrial capacity, which broadens the supply chain rather than concentrating it. Investors modeling the cryogenics layer should hold both cases open. Anyone claiming to know which one wins is guessing.

Where sovereignty enters

The geography of this modality is unusual, and it matters for policy.

Australia has the deepest national position: two internationally competitive companies in Silicon Quantum Computing and Diraq, strong university programs at UNSW, Sydney and Melbourne, and a domestic isotope separation capability. That combination means the full chain from raw material to finished chip can stay under national control, which is close to a textbook definition of technological sovereignty. The limitation is market size, and both companies are expanding toward the United States for customers and capital.

Europe is fragmented and collectively formidable. Quobly, CEA-Leti, CNRS and STMicroelectronics anchor France. Quantum Motion, the NQCC and UCL anchor the UK. SemiQon and VTT anchor Finland. imec supplies the process development backbone for all of them. The connective tissue between national labs, foundries and startups is the genuine advantage here, and it isn’t something a country can buy quickly.

The United States presents an inversion of the normal pattern. Corporate research is strong – Intel, HRL Laboratories, Sandia, and university programs at Princeton and Wisconsin – and venture-backed silicon spin startups are notably scarce. The most commercially advanced companies in this modality are headquartered in Sydney, London and Grenoble.

The policy point underneath all of this is uncomfortable for late entrants. Silicon spin is the modality most tightly bound to existing semiconductor manufacturing, which means the countries with advanced foundries, lithography access and process engineering talent start with a structural head start. Countries without a domestic fab can still participate through isotope supply, cryogenic electronics or design, but they cannot build the whole chain. That’s a different sovereignty calculation from superconducting or trapped ions, where the specialist supply chains are small enough that a determined newcomer can buy into them.

What this does to threat timelines

For security and risk teams, silicon spin raises one specific question, and it isn’t a new source of urgency.

The urgency for post-quantum migration already exists and comes from published standards and procurement deadlines, not from any hardware announcement. NIST finalized ML-KEM, ML-DSA and SLH-DSA in 2024, and migration timelines are being set by regulators and customers now.

What silicon spin changes is the shape of the scaling curve, not its starting point. Other modalities face supply chain constraints that act as natural brakes: a handful of dilution refrigerator makers, a few specialist trap fabricators, capacity-limited laser suppliers. Those constraints don’t stop progress, they pace it. Silicon spin, if CMOS-compatible fabrication holds up in production, faces far fewer of them. The manufacturing base exists, the process engineering talent pool is enormous compared to cryogenics or precision optics, and the capital required to add quantum capacity is trivial next to what the semiconductor industry already spends on equipment annually.

Silicon spin is behind on demonstrated qubit counts and will likely stay behind for some years. The point is narrower: if it crosses from research result to production process, the subsequent scaling could be faster than for any competing approach, because it would be scaling on someone else’s machine.

For anyone building a cryptographically relevant quantum computer – a CRQC, meaning a machine capable of breaking currently deployed public-key cryptography – that is a variable worth tracking in a timeline model. It doesn’t change what a migration program should do this year. It changes the confidence interval around when the program has to be finished. Organizations running structured migrations under a framework like the one at pqcframework.org should treat modality scaling behavior as an input to their timeline assumptions, reviewed annually rather than set once.

Five questions for the next modality

The value of a worked example is the method it leaves behind. Here is what we ask about any modality’s supply chain, in the order we ask it.

1. What is on the bill of materials that has no substitute? For silicon spin the answer is isotopically enriched silicon-28. Identify the input that cannot be engineered around, and you’ve found where the chokepoint sits before you look at any company.

1. Which layers are borrowed and which are bespoke? Borrowed layers scale with someone else’s capital and rarely constrain anything. Bespoke layers set the pace. Silicon spin borrows fabrication and is bespoke in enrichment and cryogenic control, which is why those two layers deserve most of the attention.

1. How many suppliers exist at each layer, and could a new one enter? Two suppliers with a five-year build time is a different risk from two suppliers where a third could enter in eighteen months. Enriched silicon is the second kind. Helium-3 is the first.

1. Which technical branch points would reroute the supply chain entirely? Millikelvin versus 1 kelvin operation changes which cryogenics companies benefit, by how much, and whether the constrained helium-3 chain matters at all. Find these branches early and model both sides.

1. What separates the announcements from the results? Roadmap dates, physical versus logical qubit counts, selected devices versus random samples from a run. The Diraq and imec result carried weight precisely because the chips weren’t cherry-picked, and that distinction is the kind that gets lost in a press release.

Run those five questions against neutral atoms, photonics or topological qubits and you get a comparable map, built from public information, without needing a physics degree to assemble it.

Where to take this next

Reading a supply chain well is a learnable skill, and it’s the one that separates useful quantum market analysis from a list of company names. Our programs at Quantum Academy work through modality comparison, supply chain mapping and the announced-versus-demonstrated discipline with the same structure we used here, applied across every major hardware approach rather than one.

For deeper technical treatment of this modality, including the fabrication detail we compressed, see the silicon spin coverage on PostQuantum.com.