What It Takes to Build a Quantum Computer

Picture a build the way several recent deployments have actually gone, composited from a few of them: parts from five suppliers across three continents, the processor and control electronics from one, the dilution refrigerator from another, the calibration software from a third. A team specifies the components, buys them, and makes them work together in a room that had to be surveyed before anything was ordered. Nobody on that team invents a new kind of qubit.

What deserves attention in a project like that is the schedule, and the schedule is set by cryogenics, cabling, and calibration rather than by qubit count or any open question in physics. That is the character of quantum hardware work now, and it is why the hiring is going where it is going.

This piece maps the work: the six engineering trades a build actually draws on, how the choice of qubit technology shifts the load between them, and what happens during the weeks between a cold refrigerator and a machine that computes.

The physics is no longer the limiting factor

Single-qubit gate fidelities above 99.9% and two-qubit fidelities above 99.5% are ordinary results on superconducting hardware today. Google’s Willow processor showed error rates falling as the error-correcting code was made larger, at code distances 3, 5, and 7 (Google Quantum AI, Nature, December 2024). That is the behaviour error-correction theory predicted and the field spent two decades failing to observe. Trapped-ion and neutral-atom platforms have reported error-corrected logical qubits as well, where a logical qubit is one reliable qubit assembled out of many noisy physical ones, with the extra copies used to detect and undo errors as the computation runs.

None of that tells anyone how to get a chip cold, wire it, characterise it, and connect it to a job scheduler. Those problems consume the budget and the calendar, and they are staffed by engineers rather than theorists.

Six trades, one machine

A build team isn’t a room full of quantum physicists. Across the deployments that have published their timelines, the same six kinds of work appear, in different proportions depending on the technology.

Cryogenic engineering

A dilution refrigerator reaches about 10 millikelvin, roughly a hundredth of a degree above absolute zero, by pumping on a mixture of two helium isotopes. It is the largest, heaviest, most expensive, and longest-lead item in any superconducting or silicon-spin build, and it sets the shape of everything around it. Bluefors in Finland dominates the market, with Oxford Instruments in the UK and Maybell Quantum in Colorado as the other established suppliers.

The awkward part is helium-3, the lighter isotope in that mixture. It comes almost entirely from the radioactive decay of tritium held in nuclear weapons stockpiles, so world supply is measured in tens of thousands of litres per year against demand from quantum computing, neutron detection, medical imaging, and fusion research. A large cryostat holds tens of litres of it. It gets recovered after every service event, never vented, and the recovery procedure is a documented operational discipline rather than a formality.

Cooling is slow. An empty system reaches base temperature in one to three days, a fully loaded one takes longer, and an unplanned warm-up above 1 kelvin costs five to ten days of recovery before the machine is useful again. That single fact is why a double-conversion uninterruptible power supply and a reliable chilled-water loop are treated as part of the computer rather than as building services.

Microwave and RF signal integrity

Every superconducting qubit needs several control and readout lines running from room-temperature electronics down to a chip at 10 millikelvin. Each line has to deliver a precisely shaped microwave pulse and carry back a signal of a few photons, while conducting almost no heat downward. The answer is staged attenuation, roughly 60 dB of it distributed across the refrigerator’s thermal levels, so that room-temperature noise is thrown away at each stage along with the heat it carries.

The return path is equally particular. A travelling-wave parametric amplifier, or TWPA, sits at the cold end and amplifies the readout signal while adding close to the minimum noise quantum mechanics permits. Everything downstream of it is easier.

Two numbers govern this trade. T1 is how long a qubit holds an excited state before decaying. T2 is how long it holds a definite phase. Both are measured in microseconds, both are exquisitely sensitive to stray heat, stray light, and stray magnetic field, and both are the first things that go wrong when the signal chain is imperfect.

Cable count is the scaling wall. Conventional coaxial cabling tops out at a few hundred lines per refrigerator before the thermal load and the physical bulk become impossible, which caps qubit count regardless of how good the chip is. Flexible superconducting stripline, of the kind Delft Circuits produces, is the main commercial answer, and anyone designing a new system should specify it at the start rather than retrofit it in two years.

Optics and ultra-high vacuum

Trapped-ion and neutral-atom machines trade the refrigerator for an optical table, and the difficulty moves with it.

A trapped-ion system needs five to 10 stabilised lasers per ion species, each held to sub-megahertz linewidth and tight intensity control, typically using cavity-locked diode lasers or frequency combs. The ions live in ultra-high vacuum near 10⁻¹¹ mbar, which is reached by baking the whole chamber at around 200 °C for days or weeks to drive adsorbed gas out of the metal. A single leak means starting the bake-out again. Optical benches run to a couple of metres on a side and want cleanroom-grade air.

Neutral-atom systems use a trap laser near 1064 nm, shaped by spatial light modulators or acousto-optic deflectors into an array of optical tweezers, each holding one atom. A magneto-optical trap, which combines laser beams with a magnetic field gradient to cool and gather atoms, loads the array, and a sensitive camera images the atoms by their fluorescence. Separate lasers drive the atoms into high-lying Rydberg states to make them interact, which is how the two-qubit gate happens.

If you are an RF engineer, trade two is close to work you already do. If you have run a surface-science or accelerator vacuum system, trade three is too.

Calibration and control

When a chip comes out of a cool-down, it is uncharacterised. Somebody has to find every readout resonator, find every qubit frequency, drive Rabi oscillations to fix the pulse amplitudes, measure T1 and T2, calibrate single-qubit and then two-qubit gates, tune the readout discrimination, and benchmark the result. This is the trade that separates a five-month build from a fifteen-month one.

The pulses themselves come from control electronics built around field-programmable gate arrays, or FPGAs, which generate and digitise waveforms with nanosecond timing. Qblox, Quantum Machines, and Zurich Instruments are the established Western suppliers, and all three are modular, so control capacity grows by adding cards rather than replacing the rack.

Automation decides the schedule. With a modern calibration platform of the kind Q-CTRL, QuantrolOx, and Quantum Machines sell, a 5-qubit system can be brought up in under a day and a 20-qubit system in roughly three weeks. Done by hand, the same 20-qubit bring-up runs into months. Calibration doesn’t stop at handover either. A production system recalibrates on a cadence measured in hours rather than weeks, and it keeps doing so for as long as it is in service.

Real-time and HPC systems

Error correction only helps if the correction arrives in time. A surface-code cycle, the repeating round of measurements that detects errors on a two-dimensional grid of physical qubits, takes about a microsecond on superconducting hardware. The decoder, the classical algorithm that turns those measurements into a statement about which errors occurred, has roughly 10 microseconds to answer for a distance-7 code before the backlog becomes unrecoverable. That budget is the reason decoders run on dedicated FPGA hardware, as with Riverlane’s Deltaflow, or on GPUs connected over a purpose-built low-latency link such as NVIDIA’s NVQLink.

Above that sits ordinary scheduling. A quantum processor in a supercomputing centre has to be requestable the same way a GPU partition is, through Slurm or an equivalent, so that a hybrid job can hold classical and quantum resources at once. Pasqal and NVIDIA have demonstrated exactly that. This trade is filled by people from high-performance computing rather than from physics, and the demand for them is growing faster than the supply.

Facilities and site engineering

A superconducting quantum computer is an industrial installation. A loaded cryostat concentrates roughly 750 kg on a small footprint, which most raised data-centre flooring won’t take; the usual requirement is a floor rated to at least 1,000 kg/m² on slab. Vibration matters enough that siting decisions are made around lift shafts, rail lines, and loading docks. Magnetic and radio-frequency interference has to be surveyed before anything is ordered, and standard fluorescent lighting is a common offender.

The services are unusual too. Pulse-tube compressors want chilled water at 15–25 °C, considerably colder than the warm-water loops that modern energy-efficient HPC halls are plumbed for, so a retrofit is often a new loop rather than a tap into the existing one. Add three-phase power with real headroom, helium recovery, and enough ceiling height and floor area for the system you will want in three years rather than the one you are buying now. Over-specifying the room is cheap. Rebuilding it is not.

What changes when the qubit changes

The six trades are constant. Their weighting is not, and modality choice is largely a choice about which trade dominates the project.

Superconducting transmon. A transmon is a superconducting circuit whose two lowest energy levels serve as the qubit, fabricated on silicon or sapphire and operated at 10–20 millikelvin. This has the most developed component supply, with characterised chips available from QuantWare, Rigetti, and others. Cryogenics and the signal chain dominate, facility demands are the highest of any modality, and the operational load includes annual compressor service and continuous drift monitoring. Expect five to nine months from empty room to first qubit signal.

Trapped ion. Room-temperature processor, so no dilution refrigerator, but the optics and vacuum work is unforgiving and the laser subsystem is the long pole. Quantinuum’s QCCD approach, short for quantum charge-coupled device, physically shuttles ions between separate zones for storage, gates, and readout, which buys connectivity at the cost of considerable control complexity. Most systems still ship as integrated appliances rather than as parts, though Infineon’s socketed ion-trap carrier points at a component market forming.

Neutral atom. The easiest modality to put in an existing building. Pasqal’s systems run at room temperature on a few kilowatts with no cryogenics, no helium recovery, and no compressor noise, which removes most of the facility remediation that superconducting forces. Deliveries into GENCI in France, Jülich in Germany, and CINECA in Italy went into conventional data-centre space. The work moves almost entirely into lasers, optics, and imaging.

Photonic. The exception to the component model. PsiQuantum fabricates its chips at GlobalFoundries and Xanadu at its own facility, and buying a photonic machine in pieces isn’t currently practical. The processor runs warm, but the detectors do not: superconducting nanowire single-photon detectors, or SNSPDs, need cooling to somewhere between 0.8 and 4 kelvin, which is a different and much lighter cryogenic problem than a dilution refrigerator.

Silicon spin. Qubits made on the same 300 mm lines that produce classical processors. Diraq and imec have reported above 99% two-qubit fidelity on devices selected at random across a wafer from an industrial fab, which is the result the whole thesis rests on. Operating temperature near 1 kelvin allows helium-4-only cryostats and removes the helium-3 dependency entirely. Intel has distributed its 12-qubit Tunnel Falls chips to university groups. The supply base is younger than superconducting, and the long-run cost argument is the strongest of any modality.

Bring-up

Between a cold refrigerator and a working computer sits a sequence that runs roughly the same way everywhere: resonator spectroscopy, qubit spectroscopy, Rabi oscillations, coherence measurement, single-qubit gates, two-qubit gates, readout optimisation, then system-level benchmarking and cross-talk mapping. On a well-automated 20-qubit system this is a few weeks of mostly unattended work punctuated by human decisions.

Articles that describe only the happy path are not much use to anyone who has to do it, so here is a failure that first-time teams meet almost every time.

Worked example: the chip that lost its coherence

A vendor ships a 21-qubit processor with a test report showing T1 averaging near 80 µs. You install it, cool down, measure, and get 25 µs. Nothing is broken. The chip was characterised in the vendor’s refrigerator, on the vendor’s wiring, behind the vendor’s shielding, and yours is different.

Work the diagnostic in order. First, leave the chip at base temperature for 24 hours and measure again. If T1 climbs, the cause is almost certainly quasiparticle poisoning, meaning stray infrared photons are breaking apart the superconducting pairs that carry the qubit’s state; the fix is to add or reseat infrared-absorbing filters at the cold plate and mixing-chamber stages. Second, check the magnetic shielding: the mu-metal can and the superconducting niobium shield have to sit properly around the sample holder, and a shield that’s present but not seated does nothing. Third, check that the final attenuator is bolted to the mixing-chamber plate with real thermal contact rather than hanging on its own cable, because an attenuator floating thermally radiates heat straight at the chip it is supposed to protect.

Two other patterns are worth recognising. Readout fidelity under 95% usually traces to insufficient TWPA gain, poor separation of the two states in the measured signal, or thermal population of the excited state, which points back at the mixing-chamber temperature under full wiring load. Two-qubit gate fidelity that lags while single-qubit fidelity is fine usually means leakage into a third energy level, or residual coupling between neighbours that the coupler bias has not cancelled.

The general lesson holds across every modality. Vendor performance figures are device-typical results from the vendor’s own setup, and no supplier publishes guaranteed per-chip minimums. You’re measuring your own hardware, in your own refrigerator, and you keep measuring it, because the numbers drift.

What to learn, and in what order

Most of the six trades are classical engineering. An RF engineer, a vacuum technician, a controls engineer, and a data-centre engineer already hold the majority of what a build team needs, and the gap is a specific and learnable fraction on top. That is the crossover we see most often, and it is faster than retraining a physicist in facilities engineering.

The order we would suggest for someone entering this work:

1. The physical picture and the vocabulary. What a transmon or a trapped ion physically is, what a gate fidelity number does and does not promise, and the difference between physical and logical qubits stated precisely enough that vendor material stops being persuasive by default.
2. The thermal and signal chain. Reading a wiring diagram for a dilution refrigerator, understanding why the attenuation is staged, and knowing what each cold component is for. This is where most integration errors are made and diagnosed.
3. Calibration. More operational hours go here than anywhere else, and it is the trade where automation has changed the job description most in the last three years.
4. Facilities. Move this to the top of the list if a site decision is coming, because facility mistakes are the only ones on this list that cost a year to undo.
5. HPC integration and error correction. Last, and increasingly the differentiator for anyone aiming at systems built for the end of the decade rather than for today.

A word on security while the building is still empty, since it’s cheaper there than later. Network segmentation between the control plane, the user-facing service, and the corporate network belongs in the design, along with firmware integrity verification on the control hardware. Calibration telemetry, which you are collecting anyway, doubles as a tamper signal, because unexplained drift on a specific channel is a fault report and an alert at the same time.

Where to start

The people building these machines aren’t, for the most part, doing new physics. They are doing careful cryogenic, microwave, optical, and systems engineering under tighter tolerances than most of them have worked to before, and the shortage is in that combination rather than in theory.

For the deeper technical treatment of the component supply chain behind all of this, the component-level analysis on PostQuantum.com goes considerably further into vendors and specifications.

Quantum Academy’s quantum engineering programs are listed at quantumacademy.com/.