Marin Ivezic
A group at the University of Naples Federico II brought up a 64-qubit superconducting quantum computer. The processor arrived as a purchased component: a 64-qubit chip from QuantWare, a Dutch supplier, bought rather than fabricated in-house. The team spent its own effort on cooling, control, calibration and the problems the machine was meant to run. Buying the processor compressed the schedule, and the time saved went into the rest of the system.
A quantum processing unit, or QPU, is the chip that holds the qubits and the wiring between them. Ten years ago you could not order one. If you wanted qubits, you fabricated them, which meant a cleanroom, a fabrication team and several years. The fact that a university can now issue a purchase order for the hardest part of the machine is the single change behind everything called quantum open architecture.
Every qubit count in this article is a count of physical qubits, meaning the raw devices on the chip. Logical qubits are error-corrected units assembled from many physical ones, and the algorithms that get quoted in board papers need thousands of them. None of the systems described here has any. A 64-qubit machine is a research instrument and a training platform, not a production computer, and the modularity argument has to be made on those terms or not at all.
What modularity is actually a decision about
Open architecture gets discussed as a philosophy. For anyone specifying a system it is narrower than that: a decision about where to put the interfaces, and a decision about who is accountable on each side of them.
A vertically integrated machine has interfaces too. They’re just internal, undocumented and invisible to the buyer, which is convenient until you want to change something. An open build makes those interfaces contractual. That is the whole trade. You gain the ability to source each part from whoever is best at it, upgrade parts on separate cycles, and audit what you’re running. You take on the obligation to specify what crosses each boundary, verify it, and own the result when two compliant components refuse to work together.
Four things cross the boundaries in a quantum system, and they behave differently. Heat. Analog signal. Timing. Data. A fifth thing crosses every boundary and appears in no datasheet, which is responsibility. The rest of this article works through them in that order, because that’s the order in which integration projects fail.
The thermal and mechanical boundary
Most quantum processors in commercial use today are superconducting transmons, small circuits patterned on a chip that behave as artificial atoms when cooled far enough. Far enough means roughly 10 millikelvin, a hundredth of a degree above absolute zero, which is colder than deep space and requires a dilution refrigerator: a multi-stage cryostat that uses a mixture of helium isotopes to reach temperatures no other practical machine reaches.
This is the layer where the component market started. From the late 2000s, Bluefors in Finland sold dilution refrigerators as products rather than as one-off lab builds, and the effect on the field was immediate. A group that wanted to do quantum experiments no longer needed a cryogenics specialist on staff. That single unbundling is the reason there are hundreds of quantum labs rather than dozens.
It’s also where the word “open” gets tested hardest, because the interface here is physical. Every control line that enters the fridge carries heat down with it. The mixing chamber at the bottom has a cooling budget of a few hundred microwatts, and the stages above it run from milliwatts to watts. The attenuation, filtering and packaging have to keep each stage inside its own limit. The wiring, the attenuators that reduce room-temperature noise at each stage, the filters, the magnetic shielding and the mechanical mount for the chip are all part of the thermal design. They are not accessories. A supplier like Delft Circuits exists precisely because cryogenic cabling turned out to be a specialism rather than a purchase.
The part that remains least standardized is the mount itself: the package the chip sits in, the connectors it presents, the thermal contact between the package and the cold plate. There is talk of a common quantum socket, in the way that a CPU socket lets you change processors without changing motherboards. It doesn’t exist yet in any enforceable form. Until it does, “our QPU fits any fridge” means “our QPU has fitted the fridges we have tried,” and the difference between those two statements is a mechanical engineering project.
The practical rule for an architect is blunt. If the cold chain is proprietary, the system is proprietary, whatever the software layer claims. Ask which parts of the fridge, wiring and packaging you could replace without redesigning the mount, and ask for the drawings.
The analog and radio-frequency boundary
Above the fridge sits the control electronics: instruments that generate shaped microwave pulses to drive the qubits, and measure the response of a readout resonator to determine what state a qubit ended in. Qblox, Quantum Machines, Zurich Instruments and Keysight all sell into this layer with different philosophies, some optimising for latency, some for modular scaling, some for measurement precision.
Two definitions worth fixing here. Gate fidelity is how closely a physical operation matches the intended one, usually quoted as a fraction of a percent of error per gate. Crosstalk is the unwanted effect of driving one qubit on its neighbours, and it’s a systems problem rather than a chip problem, because it depends on the pulse shapes, the wiring and the chip geometry together.
The interface at this boundary is a list that most datasheets only partly answer: signal levels, impedance, connector types, the attenuation plan down the fridge, the shared timing reference, the readout chain including any cryogenic amplifiers, and the software format in which pulse sequences are uploaded. Two vendors who both advertise support for superconducting qubits can satisfy all of that on paper and still need weeks of joint bench work before the first two-qubit gate hits its specified fidelity.
A pre-validated pairing is worth more at this boundary than anywhere else, because this is where the bench weeks get spent. That is also why the strongest evidence a vendor pair can offer isn’t a compatibility statement. It’s a named system, running, that uses both.
The timing boundary
The boundary that decides whether a modular machine has a future is the one nobody puts in a diagram, because it isn’t a cable. It’s a latency budget.
Coherence time is how long a qubit holds its state before noise destroys it, typically tens to a few hundred microseconds for transmons. Anything the classical side wants to do in response to a measurement has to happen inside that window. Mid-circuit measurement means reading a qubit partway through a computation; feedback means acting on that reading, conditioning the next operation on the result. Both are ordinary requirements now rather than exotic ones.
Quantum error correction raises the stakes. It works by repeatedly measuring extra qubits to produce a syndrome, a pattern of measurement outcomes that indicates where an error probably occurred without revealing the computation itself. A decoder, which is classical software, turns that pattern into a correction. Syndromes arrive continuously, and if the decoder can’t keep up, the backlog grows without limit and the correction stops working. The requirement is a hard real-time one, and it runs across the whole chain: chip, readout, digitiser, decoder, control processor, back to the pulse generator.
Latency is therefore not a property of any single box you can buy. It’s a property of the assembled system, and every additional vendor interface is a place where microseconds get spent. The Israeli Quantum Computing Center was built with a supercomputer and GPU hardware coupled directly to the quantum control system rather than reached over a general-purpose network, precisely because a hybrid loop that goes out to a datacentre and back is too slow to be part of a quantum circuit.
For anyone specifying a system, this converts into one uncomfortable question to put to every candidate stack: what is the measured round-trip time from a qubit measurement to a conditional pulse on the same chip, and which supplier is accountable for that number? A stack that can’t answer is not ready for error correction work, whatever its qubit count.
The software and data boundary
The top of the stack is the most open and the least consequential, which is worth saying plainly because most of the market’s openness claims stop there.
An SDK, or software development kit, is the library a programmer writes circuits in, such as Qiskit or Cirq. OpenQASM is a text format for describing a circuit at the level of gates. QIR, the Quantum Intermediate Representation, is a lower-level compiler representation that lets different toolchains target different hardware through a common intermediate step. All three are portable, and support for OpenQASM and the major SDKs is widespread. QIR adoption is growing but is not yet universal.
None of that gives you what an integrator needs. The things that determine whether you can operate a machine you didn’t design are pulse-level access, the calibration data format, the ability to script tune-up routines, and permission to read the raw measurement records. A vendor can support every open circuit language in existence and still ship a black box, because the circuit language sits above the layer where the physics is.
There’s a real distinction to keep straight while evaluating this layer. Error suppression means shaping the control pulses so that fewer errors occur, and it’s available today from suppliers such as Q-CTRL whose tools have been applied across hardware from several different makers. Error correction means detecting and fixing errors during a computation using redundant qubits, and it isn’t in production anywhere. Suppression is a software purchase that improves a machine you already have. Correction is an architecture requirement that constrains the machine you buy next. Vendors sometimes let the two blur.
The one genuinely portable asset at this boundary is worth naming. Because error suppression software works across hardware from different suppliers, the calibration and benchmarking work your team does on one machine largely carries to the next. That is the closest thing to a hedge available in this market.
Three ways to buy a modular system
Modularity is not one procurement path. It’s three, and they suit different institutions.
Closed turnkey. One vendor delivers a working machine and owns the performance number. Fastest to production, single point of contact, no integration staff required. You pay for it in three currencies: price, dependence on that vendor’s upgrade roadmap, and the fact that your team learns to operate a machine rather than to build one. For a company that wants a quantum resource and not a quantum capability, this remains the correct answer, and open architecture arguments should not be used to talk anyone out of it.
A published reference design. The middle option, and the newest. Component suppliers have begun to combine QPUs, control hardware and calibration software into pre-validated configurations, sold at small qubit counts and tested together before shipment. The premise is that a buyer gets components that have already been made to work together, while retaining the right to swap any of them later, and the pitch is that this removes the choice between an opaque closed system and a risky self-build.
This is the pattern classical computing already ran. The IBM PC mattered less as a product than as a published specification: once the interfaces between processor, bus, storage and operating system were documented, other manufacturers could build compatible machines, component suppliers could sell into a market rather than to one customer, and buyers could upgrade one part at a time. The compatible-machine market took roughly a decade to mature and it produced its own problems, including a long period where “compatible” meant “mostly compatible.” Quantum reference designs are at the beginning of that curve, not the end. Treat a validated configuration as a shortened integration project rather than as a product.
Bespoke integration. Source each component and assemble it, either with your own staff or through a systems integrator. This is what Naples did, and what the Dutch Tuna-5 project did more deliberately. Tuna-5 is only a five-qubit machine, built at Delft in 2025 under the HectoQubit/2 programme with QuTech and TNO integrating components from QuantWare, Qblox, Orange Quantum Systems and Delft Circuits, and made available over the Quantum Inspire cloud service. Nobody built it for the qubit count. It was built to find out where a multi-vendor stack breaks, in public, at a scale small enough to survive the answer, and the integration problems it surfaced fed straight back into the suppliers’ products.
That is the honest use of a small bespoke build. It buys capability rather than compute. The staff who bring it up learn the failure modes, and those people are the ones who can later specify, accept and operate something larger.
A fourth pattern is emerging around facilities rather than machines. The Israeli Quantum Computing Center at Tel Aviv University, operated by Quantum Machines, runs superconducting and photonic hardware through common control infrastructure and couples it to classical high-performance computing. The design goal is continuous upgradeability rather than a fixed installation. For a chip developer, the value is being able to test a device without first building a million-dollar laboratory. For a national programme, the value is that the facility outlives any particular processor inside it.
What modularity costs
The case for open architecture is stronger when its costs are stated first.
Integration labour is the largest hidden line item. Getting components from different suppliers to behave as one machine takes engineering months, and the work is unglamorous: impedance mismatches, timing alignment, driver-level incompatibilities, thermal anomalies that only appear when everything is cold. The Tuna-5 team’s own account of the project emphasises iteration and testing rather than assembly. Budget for it explicitly or it will appear anyway as schedule slip.
Peak performance usually favours co-design, at first. A vendor that controls the chip, the wiring, the pulses and the firmware can tune across all of them at once. A modular stack starts with small losses at each junction and closes the gap through calibration work. Whether it closes fully is not yet settled by evidence, and anyone claiming otherwise in either direction is selling.
Standards are immature enough to create a second kind of lock-in. Having chosen a particular QPU, cold chain and control combination, you may find the practical replacement options number one or two rather than many, because those are the pairings anyone has actually run. Modularity in principle is not modularity in your building.
Supplier survival is a live risk. Several of the specialists named in this article are young companies. If one fails, the question is whether the interface is documented well enough for someone else to step in. That is a due-diligence item, not a philosophical one: ask for interface documentation in escrow, and ask what a replacement would cost in engineering time.
Scaling is unproven. Integrating five or fifty qubits from modules has been done. No one has yet integrated around a thousand high-fidelity, gate-based qubits in a modular system. Cabling density, thermal load and control channel counts all grow faster than qubit count, and the argument that modular scaling works by adding racks is a hypothesis rather than a result.
What belongs in the contract
Total cost of ownership, in this context, means the purchase price plus the cryogen and electricity consumption, the maintenance contracts, the calibration labour, the facility work, and the engineering time to keep a multi-vendor stack current. On a research instrument the purchase price is frequently the smaller half.
Six things are worth writing down before anything is signed, because none of them is settled by a standard.
Acceptance criteria as measured numbers. Not “64 qubits” but median and worst-case two-qubit gate fidelity across a named benchmark, readout fidelity, coherence times measured in your facility rather than the vendor’s, and the measurement procedure itself. A specification that doesn’t name the benchmark hasn’t specified anything.
Ownership of the performance number. In a single-vendor purchase this is automatic. In a modular one, someone has to hold it, and that someone is either your own team or an integrator acting as prime contractor. Decide which before the first component arrives, because the alternative is a fidelity shortfall with three suppliers each pointing at the other two.
Interface documentation as a deliverable. Mechanical drawings for the mount, the attenuation and filtering plan, connector and signal specifications, timing reference requirements, calibration data schemas, and the pulse-level API. Every one of these is what makes a future swap possible, and every one is easier to obtain before payment than after.
Spares, obsolescence and support escalation. Lead times on cryogenic components are long. Establish what happens on a failed amplifier at three in the morning, and which supplier answers.
The upgrade path, in writing. Which components can change without disturbing the others, and what the vendor commits to about backward compatibility of the mount and the control interface.
Export control and site security. Cryogenic and control hardware can fall under control regimes that govern the machine as assembled, not only the individual parts. Establish this early with counsel, because it can determine which suppliers are available to you at all.
The discipline this model requires
The unbundling of the quantum computer has produced a role that didn’t exist five years ago. Firms such as ParTec in Germany and TreQ in the Netherlands now sell integration itself, and the work divides into two jobs that are usually confused.
Platform integration is assembling and commissioning the machine: the cold chain, the control plane, the firmware, the interfaces, the calibration, the acceptance tests. Enterprise integration is everything required to make that machine part of an organisation: identity and access control, segmentation between the control network and general IT, scheduling alongside classical high-performance computing, logging and audit, change control, patching, uptime commitments and a cost model that survives the first year.
Most quantum projects that stall have completed the first and never planned the second. A commissioned machine with no operational model is a demonstration, and demonstrations don’t get renewed budgets.
This has a consequence for sovereignty arguments that gets lost in the component-counting. Owning parts is not owning capability. A country or an organisation is only independent to the degree that it has people who can commission, secure, maintain and evolve the platform. That is a training problem before it is an industrial policy problem, and it’s the reason the small national builds keep being justified as capability projects rather than compute projects.
Where to build the skill
Quantum open architecture will succeed or fail on the availability of people who can hold an entire stack in view at once: cryogenics well enough to read a heat budget, radio-frequency engineering well enough to specify an interface, classical real-time systems well enough to argue about latency, and enterprise operations well enough to run the result on a Tuesday. That combination is rare and it isn’t produced by physics training alone.
Quantum Academy’s systems integration certification path was built for exactly that gap. It covers the module boundaries described here, the acceptance and benchmarking practices that make a multi-vendor build contractible, and the operational work that turns a commissioned machine into infrastructure. It’s aimed at architects, integration engineers and technical procurement leads rather than at researchers, and it assumes no prior quantum hardware background. You can review the programme structure and current intake dates at quantumacademy.com/.
For the deeper technical treatment of the stack itself, including the component market in more detail than a buyer’s article can carry, Marin Ivezic’s analysis of quantum open architecture on PostQuantum.com is the companion piece to this one.