Marin Ivezic
Procurement asks three questions of every quantum hardware vendor: what does it cost, what does it do, and when does it ship. A fourth question decides more of the build schedule than the other three combined, and it is the one that rarely makes it onto the form. If this part does not arrive, how many weeks of engineering does the replacement cost?
Answer that question layer by layer and the supply chain stops being a map of countries and starts being something a program manager can act on. Some components in a quantum build can be swapped in a fortnight. Some cost months of re-integration. A few cannot be swapped at all, at any price, on any timeline that matters to a program starting today. Those are three different problems, and they call for three different procurement strategies. Most build plans treat them identically.
The substitution ladder
Grade every component by replacement cost rather than by unit price. Three rungs give enough resolution to make decisions with.
| Layer | What you can buy instead | Replacement cost |
|---|---|---|
| Calibration and characterisation software | Two or three independent vendors | Weeks |
| Error-correction decoder | Dedicated hardware, or GPU-hosted open source | Weeks |
| Classical servers, switches, fabric | Commodity market | Weeks |
| Quantum processor (superconducting) | Three vendors across three countries | Weeks to months, by re-integration scope |
| Control electronics | Three Western vendors | Weeks to months to re-integrate; delivery set by field-programmable gate array (FPGA) allocation |
| Dilution refrigerator | Three Western manufacturers | Months, set by configuration and queue depth |
| Superconducting flex cabling | Coaxial, at a channel-count penalty | No like-for-like |
| Helium-3 | Nothing | No substitute |
| Control-plane FPGAs | Shared by every control vendor | No substitute |
Rung one: replace in weeks
Calibration software is the tooling that tunes pulse parameters, tracks qubit drift, and re-characterises the device after a thermal cycle. Q-CTRL in Australia, QuantrolOx in the United Kingdom, and Quantum Machines with its bundled toolchain all sell into this layer. The interfaces are software, the outputs are parameter sets, and moving between them costs integration time rather than hardware.
The error-correction decoder sits one layer down. It reads the stream of syndrome measurements coming off the device, works out which errors occurred, and returns a correction fast enough that the machine does not outrun it. Riverlane sells dedicated FPGA-based decoders for this. The alternative is decoding on GPUs using open-source matching libraries such as PyMatching, running on hardware you can buy from any server vendor. Two independent paths, built on different silicon, sourced from different countries.
This is the layer buyers worry about most and should worry about least.
Rung two: replace in months, not weeks
The quantum processing unit is the chip carrying the qubits themselves. For superconducting builds it’s the least concentrated layer in the stack, which surprises people. QuantWare in the Netherlands, Rigetti in the United States, and IQM in Finland all sell processors to external integrators. Rigetti’s Novera ships as a packaged module, specified by Rigetti at nine physical qubits. Substituting one vendor for another means redoing the wiring map, the calibration routines, and the characterisation campaign. That’s weeks to months of engineering depending on how much of the integration has to be redone, not a rack swap, and it’s survivable.
Control electronics generate the microwave pulses and read out the results. Qblox in the Netherlands, Quantum Machines in Israel, and Zurich Instruments in Switzerland are the three Western options. All three drive superconducting qubits and none of them does it the same way: different synchronisation architectures, different pulse-sequencing languages, different toolchains. A team that has written six months of sequences against one platform is porting those sequences, not repointing a driver, and that port runs weeks to months. Delivery is a separate wait, often several months, because the boards depend on FPGA allocation.
The dilution refrigerator is the cryostat that holds a superconducting processor at a few tens of millikelvin. Bluefors in Finland, Maybell Quantum in the United States, and Oxford Instruments in the United Kingdom serve the Western market from three allied countries. The geography looks comfortable. The queue does not. All three are capacity-constrained, delivery depends on configuration, and when several national programs order in the same quarter the queue stretches for everyone behind them. The cryostat is the longest-lead item in a superconducting build, and a team that orders it after the facility is ready has already lost the months it spent waiting.
Rung three: nothing to buy instead
Superconducting flex cabling carries signals from room temperature down to the cold stage on thin polyimide ribbon with superconducting traces on it, instead of on bundles of coaxial line. The advantage is channel density and heat load: flex gets far more lines into the same cryostat without dumping thermal load onto the mixing chamber. Delft Circuits in the Netherlands is the only company selling this as a commercial product. Coaxial cabling is the fallback, and it works, up to a ceiling of a few hundred channels. Any superconducting system aiming past that ceiling has a single supplier and no second source. This is the narrowest point in the Western quantum hardware supply chain, and a team planning a 200-qubit-class machine for 2027 should be talking to that supplier now rather than when the processor order lands.
Helium-3 is the working fluid that makes a dilution refrigerator work at all. It is produced almost entirely from the radioactive decay of tritium in nuclear stockpiles, which means the supply is a by-product of weapons programs rather than a market responding to demand. Quantum computing competes for it against medical imaging, neutron detection, and fusion research. Helium-4 does not substitute, because the dilution cycle depends on the specific behaviour of the two isotopes mixing. No procurement strategy eliminates this dependency for a superconducting build. Two things reduce exposure: negotiate multi-year helium-3 supply as part of the cryostat contract rather than as a consumable afterthought, and commission recovery infrastructure from day one, so the initial charge is recycled rather than vented. Every unplanned warm-up that loses helium-3 is a supply chain event.
The third item on this rung is the one buyers rarely map, because it hides inside a component they think they have diversified. Every commercial control platform on the market is built on FPGAs from AMD/Xilinx or Intel/Altera, the reconfigurable chips that generate and time the pulses. Switching control vendors does not switch this dependency, because all three vendors share it. The quantum industry is also a rounding error in FPGA demand next to defence, telecoms, and AI, so allocation decisions are made elsewhere and communicated late. The only mitigations are early ordering and written allocation confirmation, with strategic allocation discussions available to sovereign programs and to almost nobody else.
The ladder moves
A rung assignment is not a property of the component. It is a property of the component in your jurisdiction, and export controls can promote a part from rung two to rung three overnight.
The Wassenaar Arrangement is the multilateral export-control regime whose member states use its lists to set their own national licensing rules, and its 2023 List of Dual-Use Goods and Technologies added quantum computing control entries that member states rolled into national lists through 2024. The US Export Administration Regulations control certain quantum technologies, FPGAs, and cryogenic CMOS components. The EU Dual-Use Regulation 2021/821 carries quantum items on its control list. The US Bureau of Industry and Security lists Origin Quantum and QuantumCTek on its Entity List, which settles the question of PRC-origin hardware for any defence, government, or regulated-sector program.
The practical consequence is narrow and checkable. A buyer outside the United States or the EU sourcing a Dutch processor needs written confirmation that the Dutch authority will license the shipment to their jurisdiction before the purchase order goes out, not after. A buyer in a country without an established FPGA supply relationship needs the same confirmation from the chip vendor. These checks cost days. Discovering the answer during month five of a build costs quarters.
There is a second way the ladder shifts, and it is geographic rather than legal. Look at the superconducting stack again: the processor, the control electronics, and the cabling can all come from Dutch companies inside roughly one hour’s drive of each other. Three layers, one policy jurisdiction, one regional risk. Nothing about that is a criticism of the suppliers, who are excellent. It’s a mapping exercise most programs never run, because they diversify by layer and forget to diversify by country. Qualifying one non-Dutch alternative at each of those three layers is cheap insurance against a correlated failure.
Buy in reverse ladder order
The ladder implies a purchase sequence, and it runs opposite to how most teams instinctively buy. The instinct is to start with the processor, because that’s the interesting part. The schedule says start with rung three, then rung two, then rung one, and run them in parallel wherever cash flow allows.
Order the cryostat first and confirm the delivery slot in writing. Place the control electronics order with FPGA allocation confirmed, not assumed. Open the cabling conversation before the channel count is final, so the supplier can plan capacity against your window rather than fitting you into a batch. Rung-one components can be selected late, because selecting them late costs weeks and nothing else.
This is also the argument for open architecture, meaning a system assembled from independently sourced layers rather than bought sealed from a single vendor. A vertically integrated system removes every procurement decision, which sounds like a relief until the vendor has a production problem, a policy change, or a strategic pivot. Then the entire machine sits on rung three, because the alternative to your supplier is starting the program over. Assembling from qualified layers keeps most of the stack on rung two, where a supplier failure costs re-integration effort you can scope and budget. It also tells you what is actually inside the machine, which matters when someone asks whether the build meets a sovereignty requirement.
Modality decides which ladder you climb
The choice of qubit modality is usually argued on physics: gate fidelity, connectivity, coherence time, algorithmic fit. It is also a supply chain decision, and the three main options carry very different rung-three exposure.
Superconducting has the deepest dependency graph. It needs the cryostat, the helium-3, the millikelvin cabling, the commercial control stack, and the FPGAs underneath it. Every rung-three item in this article applies. In exchange, it has the most developed multi-vendor market and the most experience assembling machines from parts.
Neutral atom drops the entire cryogenic chain. No dilution refrigerator, no helium-3, no millikelvin cabling. Systems from Pasqal run in a standard rack at room temperature, and the supply chain moves to lasers from suppliers such as Toptica and M Squared, plus ultra-high-vacuum components from a broad industrial base. The trade is at the integration layer: these machines ship largely as appliances, so a buyer swaps a long rung-three list for a single-vendor relationship covering the whole system.
Silicon spin removes the helium-3 exposure by operating at around one kelvin rather than tens of millikelvin, which brings a helium-4-only cryostat into reach. Fabrication moves toward existing semiconductor plants, so the constraints become fab access and lithography time rather than quantum-specific bottlenecks. The vendor base today is thin, with Diraq in Australia, SemiQon in Finland, and Intel in the United States. It’s a bet on a supply chain that barely exists yet and has the clearest route to volume manufacturing.
None of these is the correct answer. If a program carries a sovereignty mandate that restricts sourcing, map the available vendors per modality before committing to the physics, because the modality that best fits the computational goal may not be the one the procurement rules allow.
What to do before the first purchase order
Three actions cover most of the exposure. Grade every line item on the bill of materials by replacement cost in weeks, and write the number down next to the price. Qualify a second vendor at every rung-two layer, treating the qualification engineering as an insurance premium rather than as waste. Order the rung-three items first and get the confirmations in writing.
That work takes a few weeks and it protects a schedule measured in years. Programs that skip it do not fail loudly. They discover in month five that first signal has moved to month fourteen, and by then every remedy is expensive.
Quantum Academy’s engineering and procurement training covers this material as a working method, with the component maps, vendor qualification criteria, and lead-time sequencing applied to real build plans rather than described in the abstract. You can see the current programs at quantumacademy.com/.
For the geopolitical and policy analysis behind these dependencies, including the chokepoint and export-control detail this article summarises, see the original supply chain concentration analysis on PostQuantum.com.