Marin Ivezic
Consider a university physics department commissioning a 64-qubit superconducting machine. The hardest component in it arrives on a loading dock. The processor is bought from a specialist supplier as a standalone part, and the university supplies the building, the cooling, the wiring, the software choices, and the engineering hours that turn a chip into a working instrument. Ten years ago that purchase order couldn’t have been written, because nobody sold the item on it.
That is the practical meaning of Quantum Open Architecture, usually shortened to QOA: a quantum computer assembled from separately purchased modules that talk to each other over published interfaces, rather than a sealed product designed end to end by one company. The parts are now genuinely purchasable. Processors, refrigerators, control electronics, calibration software, and cryogenic cabling all have independent suppliers competing for the same order.
For a buyer, the interesting consequence is not that quantum computers got cheaper – the risk moved. In a sealed system, integration risk sits inside the vendor’s margin, priced in and invisible. In a modular system, integration risk sits on your side of the contract unless somebody is paid to take it back. This post is about where that risk lands, what to write into the specification so it lands somewhere you chose, and which costs stay off the quote until the second year.
Four buying positions, not two
The old choice was binary and both branches were unattractive. Buy a sealed machine from a full-stack vendor and you own an appliance you can’t open, on the vendor’s upgrade schedule. Build everything yourself and you spend five years reproducing a dilution refrigerator that a Finnish company already sells.
Modular supply opens two positions in between, and the difference between them is the whole procurement question.
Position one, the sealed system. One supplier, one warranty, one number to call. You get predictability and you give up the ability to swap any part of it. Cloud access to somebody else’s hardware is a variant of this position with a shorter commitment.
Position two, the integrated modular system. Multiple component suppliers, one prime contractor who commissions the machine, guarantees a performance figure, and manages the suppliers behind the scenes. You pay an integration fee that looks like overhead on the quote and is in fact the transfer of technical risk.
Position three, self-integration with support. You buy the modules directly, you own the integration, and you buy engineering support by the day from the component suppliers when something behaves oddly. Most university deployments sit here.
Position four, full in-house build. Still the right answer for a group whose research question is the hardware itself, and the wrong answer for almost everyone else.
Most disappointing quantum procurements we hear about are position-three purchases written as though they were position-two purchases. The buyer priced the components, budgeted nothing for commissioning, and discovered in month four that no supplier had agreed to make the machine work as a whole.
What you are actually buying, layer by layer
A modular system has roughly five purchasable layers. Each one is a separate negotiation with its own failure mode, and each one has a question that belongs in the specification rather than in a later email.
The processor
The quantum processing unit, or QPU, is the chip carrying the qubits themselves. In superconducting systems these are usually transmons, the dominant qubit design, built as small circuits on a chip that behave quantum-mechanically once they are cold enough. Suppliers sell them by qubit count, but the count is the least useful number in the datasheet.
Two figures matter more. Coherence time describes how long a qubit holds its state before ambient noise destroys it, and gate fidelity describes how close an operation on one or two qubits comes to the operation you asked for. A 41-qubit chip with poor two-qubit fidelity will run shorter useful circuits than a 17-qubit chip with good fidelity, and it will cost more to cool and control.
The specification question: are the quoted fidelities measured on the individual unit you are buying, or on a representative device from the same fabrication run? Those are different commitments, and only one of them is testable on delivery.
The cold chain
Superconducting qubits work at a few thousandths of a degree above absolute zero, which is what a dilution refrigerator delivers. Bluefors turned this layer into a purchasable product long before the rest of the stack followed, and buying rather than building a refrigerator is now so normal that buyers forget it’s a decision.
The refrigerator is only part of the item. The cold chain also includes the wiring that carries radio-frequency (RF) signals from room temperature down to the chip, the filters and attenuators along that path, the packaging that holds the chip, and the mechanical fit between all of them. Every one of those introduces heat or noise if it is chosen badly, and the wiring budget scales with qubit count in a way the chip price does not.
The specification question: how many wiring lines does the refrigerator support, how many does your intended processor need, and what does adding the next fifty cost? A machine that can’t be upgraded because its cold chain is full is a modular system in name only.
The control plane
The control electronics generate the microwave pulses that perform operations on qubits and read the results back out. This layer is where the promise of modularity is either delivered or quietly withdrawn, because it sits between the two things a buyer most wants to swap.
Suppliers differ on philosophy rather than on features. Some optimise for low latency, meaning the time between measuring a qubit and acting on that measurement, which is what error correction and feedback algorithms depend on. Some optimise for channel density and rack economics, which is what matters when a five-qubit prototype has to become a fifty-qubit system without turning into a cabling problem. Some optimise for instrument-grade precision on a small number of channels, which suits an experiment better than a service.
The specification question: what is the measured round-trip latency from readout to conditional pulse, and does the vendor’s roadmap support your qubit count without a forklift replacement? Buyers who treat control electronics as a commodity discover the answer two years later, when the upgrade path turns out to be a rebuild.
The calibration layer
A quantum processor arrives untuned. Getting it to perform at its specification means finding the right pulse shapes, frequencies, and timings for every qubit and every pair of qubits, then finding them again as the device drifts. That work is called calibration or tune-up, it used to consume weeks of expert time, and it is now sold as software by firms such as Q-CTRL and QuantrolOx.
For a buyer this layer is easy to underestimate and expensive to skip. It’s the difference between a machine that hits its numbers in an acceptance test and a machine that still hits them in month nine, when the original engineer has moved on. Automated calibration also travels well across a modular stack, because a good tune-up routine is largely indifferent to which company made the chip underneath it.
The specification question: who runs the tune-up after handover, how long does a full recalibration take, and does the licence cover the hardware you plan to buy next?
The integration and operations layer
The last layer has no hardware in it. Somebody has to design the whole system, commission it, prove it meets its numbers, connect it to your high-performance computing (HPC) environment or your network, apply your access controls and audit logging, and keep it running on a change-management process your organisation recognises.
In a sealed purchase this layer is included and invisible. In a modular purchase it is a line item, an internal team, or an omission. Specialist integrators such as ParTec and TreQ exist because the omission turned out to be common. National facilities have taken the same lesson: Delft’s five-qubit Tuna-5 system and the Israeli Quantum Computing Center in Tel Aviv were both built as multi-vendor platforms with a named party responsible for making the parts agree.
The specification question: who signs the statement that the system is complete, and what happens to their fee if they can’t?
The contract is where modularity succeeds or fails
Component selection is the part buyers enjoy and the part that matters least. Four clauses decide whether a modular purchase behaves like an asset or like a research project.
Acceptance criteria
Write the acceptance test before you choose the vendors, not after. An acceptance test that says “64 qubits, operational” is unenforceable. A usable one names the metric, the measurement method, the sample, and the repetition.
We teach candidates to write acceptance criteria in four parts: what is measured, how it is measured, across how much of the machine, and how many times. Applied to a processor, that produces something like a median two-qubit gate fidelity across every connected qubit pair, measured by randomized benchmarking, which is a standard procedure that runs long sequences of operations designed to cancel out and infers the error rate from how far the result drifts, repeated on three separate days without recalibration between runs. Every clause in that sentence closes a dispute that would otherwise happen after payment.
The warranty boundary
In a multi-vendor system, faults don’t respect supplier boundaries. A qubit that decoheres early might be a chip defect, a noisy pulse, a warm wire, a ground loop, or vibration from a pump. Each of those belongs to a different company, and each company can honestly say the fault is not theirs.
There are two workable answers and one common non-answer. The first workable answer is a prime contractor who accepts a whole-system service level agreement (SLA) and then argues with the component suppliers on your behalf. The second is an in-house team with the instrumentation and the authority to diagnose to the layer, so that warranty claims arrive with evidence attached. The non-answer is a set of five component warranties and the hope that fault ownership will be obvious. It rarely is.
Upgrade rights
The reason to buy modular is to replace one part later without replacing the rest. That reason only survives if the contract says so. Ask what happens when you install a processor from a different supplier: does the control software licence still apply, does the calibration package support it, does the refrigerator’s chip mount fit, and does anybody’s warranty survive the change?
Pre-validated reference designs exist partly to answer this. Multi-vendor packages such as QUB, assembled from a processor supplier, a control supplier, and a calibration supplier who have already tested their parts together, exist to sell interoperability as a product rather than as a hope. Whether you buy one or not, the underlying question belongs in your contract.
Exit and spares
Component suppliers in this market are young, and some will be acquired or will fail. That’s not a reason to avoid them – it’s a reason to ask three questions before signing: what spares are held and where, whether firmware and calibration data are escrowed in a form somebody else could use, and whether the interface documentation is yours to keep. A supplier confident in their interfaces won’t object to any of the three.
A worked comparison
The following is an illustration rather than a real tender, and the numbers are chosen to make the arithmetic visible.
Two bids arrive for an on-premises 17-qubit superconducting system, to be installed in an existing facility and connected to a small HPC cluster.
Bid A quotes the components and a two-week commissioning visit. It is the cheaper of the two by roughly a fifth. Acceptance is defined as “system demonstrates two-qubit gates on all qubit pairs”. Warranty is per component, twelve months. Calibration software is a one-year licence tied to the processor model. Integration with the HPC cluster is listed as customer scope.
Bid B quotes the same components through a prime contractor. Acceptance is defined as median two-qubit fidelity of 99.0% across all connected pairs by randomized benchmarking, sustained across three runs on separate days, plus a job submitted from the HPC scheduler and returned to it. Warranty is whole-system for twenty-four months with a named response time. Calibration licence covers any processor from that supplier’s current range. Two named engineers are on site for six weeks.
Bid A is cheaper on the purchase order and more expensive in every other year. Its acceptance clause can be satisfied by a machine that runs gates badly, which means the buyer’s own team discovers the performance shortfall after the payment milestone rather than before it. Its per-component warranty makes every fault a diagnosis exercise before it becomes a repair. Its calibration licence turns the first processor upgrade into a renegotiation. Its HPC integration is real work that somebody must do, and pricing it at zero doesn’t make it free.
The honest version of Bid A costs the difference plus the salary of whoever absorbs the scope. Sometimes that person exists and the maths favours Bid A. The failure mode is choosing Bid A while believing you bought Bid B.
The costs that never appear on the quote
Total cost of ownership (TCO) for a modular quantum system runs well past the hardware, and the items below are the ones we see missed most often in first-time budgets.
- Cryogen and compressor service. Refrigerators need helium, maintenance, and cooling water or chilled air. Service contracts are annual and non-optional.
- Power and heat rejection. The compressor and the control racks draw continuously, and the facility has to reject the heat somewhere.
- Floor, vibration, and electromagnetic environment. A machine installed above a loading bay or next to a lift motor will underperform its datasheet, and remediating the room after installation costs more than surveying it before.
- Wiring replacement. RF lines and connectors degrade with thermal cycling. They are consumables on a multi-year clock.
- Recalibration time. Time spent tuning is time not spent computing. Ask for the expected duty cycle, not just the peak specification.
- Spares inventory. A single failed amplifier can idle the whole system if the replacement has a twelve-week lead time.
- People. One qualified operator is a single point of failure, and two is the practical minimum for anything you intend to offer as a service.
- Export and classification review. Quantum processors and some control electronics sit inside dual-use control regimes in several jurisdictions. Legal review belongs in the project budget, before the order.
None of these is exotic and none of them is quantum-specific in character. They are the ordinary costs of operating scientific instrumentation, and they surprise buyers who approached the purchase as an IT acquisition.
Sovereignty is an integration competence
National programmes are among the largest buyers in this market, and their stated reason for preferring modular supply is usually independence: a country that can assemble a system from parts is not dependent on one foreign vendor’s export policy or product roadmap.
The reasoning holds, with one qualification that changes what should be procured. Owning components is not the same as owning capability. A facility whose parts are all domestic but whose commissioning, calibration, and fault diagnosis are performed by a visiting team under a support contract has relocated the dependency rather than removed it. The scarce competence is integration, held by trained staff rather than by any single vendor’s hardware.
That is why the more serious national facilities are built as open testbeds rather than as delivered products. A testbed that accepts new hardware, and lets outside groups bring devices to run on infrastructure they could not afford alone, leaves behind staff who have commissioned, broken, and repaired the thing themselves. A small multi-vendor machine built before a larger one is attempted buys the same thing at lower cost.
For a buyer with a sovereignty mandate, the procurement consequence is direct: budget for knowledge transfer as a deliverable with an acceptance test of its own. Named local staff, documented procedures, and a supervised recalibration performed by your team rather than the vendor’s.
Who signs the acceptance certificate
Every clause above assumes somebody on the buyer’s side can read a fidelity report, tell a chip defect from a wiring fault, and hold a supplier to a number. That role doesn’t have a settled job title yet, and it isn’t the same as a quantum algorithms role.
The work is closer to systems engineering than to physics. It involves specifying interfaces, running acceptance tests, allocating faults across supplier boundaries, planning upgrade paths, and connecting an instrument to an enterprise environment with its access controls and change management intact. In our systems-integration cohorts, the participants who find this material easiest tend to arrive from HPC operations, RF and instrumentation engineering, or infrastructure programme management, and the quantum-specific content they need is narrower than they expect.
The organisations that are getting good outcomes from modular purchases have that role staffed before the tender goes out. The ones renegotiating in month six generally staffed it after handover, on the assumption that integration was included.
Building the capability
Quantum Academy built the Certified Quantum Systems Integrator (CQSI) program for exactly this gap: the specification, acceptance, and operational side of modular quantum platforms, aimed at engineers and programme leads who will be responsible for a machine rather than for an experiment. It’s a private credential, and what it certifies is competence against a published body of knowledge, not a licence.
Program details, prerequisites, and the current assessment structure are at quantumacademy.com/. For the deeper technical treatment of open architecture and the supplier map behind it, the full analysis on PostQuantum.com covers the stack layer by layer.
If you’re drafting a tender this year, start with the acceptance clause. Everything else in the contract is downstream of the moment somebody has to prove the machine works.