Preparing a Facility for a Quantum Computer

The word “facility” hides four different projects

Before Germany’s first hybrid supercomputer-quantum system went into the Leibniz Supercomputing Centre (LRZ) in Munich, IQM’s engineers spent more than 25 hours measuring three candidate rooms inside a building that already stood there. They logged floor vibration, magnetic fields and sound across a full day-night cycle. Passing trams showed up in the data. So did music playing in an adjacent room. The survey was later published as arXiv:2509.12949, and the sequence is the useful part: the measurement came first, and the hardware order came after.

The lesson usually drawn from stories like this one is that quantum computers are demanding. True, and not much help to someone holding a floor plan. The more useful lesson is that “quantum computer” names five quite different building projects, and the first thing a facility team needs to establish is which one it has been handed.

The dividing line is modality, meaning the physical approach a vendor uses to build qubits. Superconducting systems put the chip, called the quantum processing unit or QPU, inside a dilution refrigerator: a cooling vessel, commonly called the cryostat, that reaches around 10 to 20 millikelvin. That machine is heavy, it needs chilled water, it consumes scarce helium, and it is exquisitely sensitive to vibration and stray magnetic fields. Everything difficult about quantum facility preparation follows from those four facts.

Take the cryostat away and the project changes character. Neutral-atom systems run at room temperature. Pasqal’s Orion series occupies a dedicated hosting room and draws a few kilowatts steady state, with no cryogenics, no helium inventory and no chilled-water plant. Floor loading and power are within reach of an ordinary equipment room, and the dominant concern shifts to laser safety. Trapped-ion systems also drop the refrigerator, and replace it with an optical table, ultra-high-vacuum chambers that need a long bake-out at high temperature, and five to ten lasers per ion species held to tight thermal stability. The room becomes an optics laboratory with Class 4 laser interlocks rather than a cryogenic hall.

Silicon-spin systems sit in between. They keep the refrigerator, but the higher operating temperature, around 1 kelvin, means some platforms run on helium-4 alone and drop the helium-3 dependency entirely. Structural, vibration and shielding requirements stay close to superconducting; the gas-handling problem shrinks dramatically. Photonic systems run the processor at room temperature and confine the cryogenic requirement to the single-photon detector subsystem, which pushes the difficulty into fiber routing and optical alignment stability instead.

So the honest answer to “can our data center host a quantum computer” begins with a question back. If the answer is neutral atom, remediation may be modest. If the answer is superconducting, we are describing a construction program with a lead time that competes with the hardware itself, and it needs to start earlier than most procurement schedules assume.

The rest of this article assumes the superconducting case, because it sets the ceiling. Anything that satisfies a transmon installation will satisfy the others.

What the survey measures, and why it runs a full day

A site survey that runs four hours on a quiet Saturday will tell you almost nothing. Most of the interference that damages qubit performance is periodic. Heating and ventilation plant cycles. Freight elevators run on Monday mornings. Commuter rail follows a timetable. Cleaning crews push heavy equipment through corridors at night. Structural resonances shift as the building warms and cools across the day. A survey has to be long enough to catch a full diurnal cycle plus at least one weekday peak, which is why 25 hours of continuous measurement has become the working floor rather than a cautious maximum.

The instrument set covers five categories, and each one produces a pass or fail against a stated threshold rather than an impression.

Vibration. Accelerometer data across the spectrum, with the critical band below 10 Hz where traffic, foot traffic and compressors dominate. The published reference point is the VC-A criterion, a vibration standard originally written for electron microscopy and microsurgery suites. State-of-the-art quantum installations aim at VC-A or tighter.

Electromagnetic interference, usually shortened to EMI. A spectrum sweep across the 4 to 8 GHz band where superconducting qubits operate, plus separate DC and AC magnetic field mapping.

Power quality. Total harmonic distortion, or THD, which measures how badly the supply waveform is contaminated by switching loads elsewhere in the building. Also voltage sags and swells, three-phase balance, and the amperage actually available at the proposed panel.

Structural. Load capacity in kilograms per square meter at the exact proposed position, not the average for the floor plate.

Thermal and humidity. Ambient stability in the 20 to 24 degrees Celsius band, relative humidity between 30 and 60 percent non-condensing, and the temperature and flow rate of chilled water that can actually be routed to the room.

The deliverable is a specification document that maps every parameter to pass, fail or marginal, with a remediation plan and cost against each gap. For an existing data center, two to six remediation items is a normal outcome, and lighting replacement, floor reinforcement and a dedicated chilled-water loop are the three that come up most often.

The four sign-offs

Here is the organizing idea that we find most useful when teaching this material: a quantum facility project is not one approval, it is four, and they come from four different disciplines that rarely sit in the same meeting. Naming them early is what keeps the schedule honest, because each one can independently stop a cryostat from being ordered.

Structural

A fully loaded LD or XLD-class dilution refrigerator lands around 750 kg. Larger platforms of the KIDE class reach several metric tons. That mass sits on a footprint close to a single square meter, delivered through three or four feet, and it does not behave like a server rack that spreads its load across a cabinet base.

Work the arithmetic once and the problem is obvious. Raised-floor tiles in a conventional colocation hall are typically rated for 300 to 500 kg/m² distributed. A 750 kg cryostat on roughly one square meter already exceeds that by a factor of one and a half to two and a half before anyone considers the concentrated load under each foot. The Open Compute Project (OCP) specifies a minimum of 1,000 kg/m² in its quantum infrastructure white paper. In practice the answer is either steel pedestals transferring load to the structural deck, or direct slab-on-grade mounting in a purpose-built room, and either way a structural engineer signs the drawing before the purchase order goes out.

The delivery path deserves the same arithmetic. The cryostat arrives crated, and the crate is heavier and larger than the machine inside it. Add a pallet truck and two people and the moving load through a freight elevator approaches a metric ton. Corridor widths, door openings, turning radii and elevator ratings all have to be walked and measured, not assumed from a drawing set. An XLD-class crate does not fit through a standard office doorway, and discovering that on delivery day is an expensive way to learn it.

Electrical

The demand is modest. The LRZ team reports roughly 30 kW peak during cool-down, settling to 10 to 15 kW steady state depending on how many control instruments are racked. That is less than a single high-density classical cabinet. Quality, not quantity, is what the specification is really about.

The working numbers are a dedicated three-phase supply, THD held below 1 percent on every outlet feeding the control racks and cryostat support systems, and an online double-conversion uninterruptible power supply sized for at least 20 minutes at full load.

The UPS sizing confuses people who come from IT, because 20 minutes buys no meaningful uptime. It isn’t meant to. If the pulse-tube cooler stops, the cryostat begins to warm, and once the mixing chamber, the coldest stage where the chip sits, rises above about 1 kelvin, the system cannot simply be restarted. Recovery is measured in days: a controlled warm-up, recovery of the helium mixture into cylinders rather than venting it, diagnosis and repair, a cool-down that itself runs several days, then full recalibration. The UPS exists to buy an orderly shutdown that protects the gas charge, and 20 minutes is enough when the control software can execute that sequence automatically. Where a human has to intervene, 30 minutes gives useful margin.

Read that way, the electrical specification is a risk decision rather than a load calculation. Redundant A and B feeds from separate utility circuits, and a generator with automatic transfer where grid reliability is marginal, are priced against a recovery that costs a week of user access.

Mechanical

Pulse-tube cryocoolers, the first-stage machines that take the cryostat from room temperature down toward 4 kelvin, are driven by compressors that need cooling water at 15 to 25 degrees Celsius, with flow in the range of 10 to 30 liters per minute at 4 to 6 bar.

That temperature is the trap. Modern energy-efficient high-performance computing halls run warm-water cooling loops at around 40 degrees and above, precisely because warm water rejects heat cheaply. A quantum computer cannot share that loop. It needs a dedicated secondary chilling circuit, which means plant space, plumbing routes and a penetration through the shielded room wall.

Two details are easy to miss at design time. Chilled-water lines carry compressor vibration straight into the room through rigid pipework, so the supply and return runs need insulation and vibration isolation as a specified requirement rather than a site decision. And where more than one cryostat is planned, the plant is sized for simultaneous operation of every compressor plus redundancy, not for the first machine.

Gas and life safety

Every dilution refrigerator holds a charge of helium-3 and helium-4. Helium-4 is a commodity. Helium-3 is not: it is produced in tiny quantities as a decay product of tritium, its price per liter runs orders of magnitude above helium-4, and lead times for replacement can stretch into months. An XLD1000sl-class system holds roughly 40 liters of mixture, which makes the gas charge a capital asset that happens to be a fluid.

Closed-loop gas handling is therefore mandatory rather than good practice. Every servicing event that warms the system recovers the mixture into cylinders. Venting it is not a cost event, it is a procurement event with a delivery date attached to it.

The safety case runs alongside the economic one. Both isotopes are asphyxiants, and a sudden release in an enclosed space displaces breathable air. The design requirements are consistent across installations: a secured, ventilated storage enclosure separated from the quantum room, access control on it, oxygen sensors at both floor and ceiling level with audible and visual alarms in the storage enclosure and the quantum room alike, and at least one warm spare helium-3 charge held on site. Ceiling and floor monitoring together, because helium rises but a displacement scenario can strand a low-oxygen layer anywhere.

The parameters with no conventional owner

Vibration and electromagnetic interference are where quantum installations diverge hardest from data center practice, and they share an awkward property: nobody on a normal facilities org chart owns them. No structural engineer has been asked to certify an AC magnetic field below 1 microtesla, and no electrical contractor has been asked to hold a room below the VC-A vibration criterion.

Start with the field thresholds. The OCP specification puts DC magnetic fields below 100 microtesla and AC fields below 1 microtesla near the cryostat. For scale, the Earth’s own field runs roughly 25 to 65 microtesla, so a site can approach the DC threshold before any building infrastructure is considered. Steel structure, reinforced concrete and nearby switchgear all distort the local field.

The AC threshold is the one that produces surprised faces. Standard overhead fluorescent lighting fails it. So do variable-frequency drives in ventilation plant, and some UPS topologies. The fixes are cheap when specified early and disruptive when discovered late: low-noise LED fixtures on DC supplies, ductwork routed away from the cryostat, and the UPS positioned outside the shielded envelope so its switching harmonics stay there.

Shielding then works in layers. Inside the cryostat, the QPU sits within a high-permeability mu-metal can for DC rejection and a superconducting can for high-frequency rejection, and that shielding arrives as part of the signal chain. Room-level shielding still earns its cost, because the drive and readout lines run exposed between the cryostat top plate and the control racks. A full Faraday enclosure around the room, grounded on a star topology from a single reference point, with isolation transformers on the incoming supply and fiber-optic isolation on every digital interface, is the strongest form. It is also expensive, and the LRZ installation shows it is not always necessary: careful site selection plus localized shielding around the cryostat and its immediate racks produced a working system.

Vibration divides into two problems that need separate answers. External vibration below 10 Hz, from traffic, elevators, loading docks and plant, couples through the structure into the machine. The practical rule used in siting is an exclusion zone of about 100 meters from heavy infrastructure, and the same perimeter is applied to strong EMI sources such as cellular base stations and high-voltage switchgear. Internal vibration comes from the pulse tube itself, which pulses helium gas at roughly 1 Hz. Remote-motor compressor configurations move the motor away from the cold head, and some cryostat designs integrate negative-stiffness isolators between the pulse tube and the dilution unit. The isolation slab beneath the machine has to absorb both sources.

Designing for the second machine

The most common expensive mistake in quantum facility work is building precisely for the system arriving this year.

QPUs turn over on two to five year cycles. Control electronics are upgraded more often than that. The building is the slowest-moving element in the deployment, and it should be specified to support a decade of hardware change without structural work. Five headroom decisions cost little at construction and a great deal afterwards:

1. Power. A single 20-qubit system draws 10 to 15 kW steady state. Specifying around 50 kW of three-phase capacity accommodates a second cryostat, more control racks and a GPU node for real-time error-correction decoding. Cable trays and panel boards are cheap during fit-out.
2. Cooling. Size the chilled-water plant for two compressors even when installing one. Adding capacity later means plumbing work in the quantum room, which means warming the cryostat that is already there.
3. Floor area. Double the shielded footprint at design time if a second machine is credible. Building a second shielded room next to an operating one is a construction project with a live neighbour. Two cryostats also buy redundancy: LRZ runs two systems, so maintenance on one does not take user access to zero.
4. Fiber and crane. Pull more fiber than the current design needs, since GPU-coupled decoding already wants very high bandwidth between racks and compute, and future QPU-to-QPU optical links will want dedicated paths. Rate the overhead hoist for the heaviest cryostat class that might plausibly arrive, not the one on order.
5. Gas handling. Size recovery and storage for the total helium-3 inventory across all planned systems. Cylinder storage space is inexpensive; improvising it later is not.

Where the standards sit

Two documents belong in a hosting contract, and they do different jobs.

The general baseline is EN 50600, published internationally as ISO/IEC 22237, which covers the building shell, power distribution, cooling, physical security and cabling for data centers. A clean audit against it produces a sound facility. It says nothing about millikelvin cryogenics, sub-microtesla shielding or vibration criteria.

The quantum-specific layer comes from OCP. Its Future Technologies Initiative launched a Quantum Information Infrastructure workstream in late 2025, bringing hardware builders, data center operators and HPC centers together. Its first deliverable is a white paper on integrating quantum processing units into data center infrastructure, accompanied by a QPU deployment checklist, with the IQM and LRZ installation as the first published case study. OCP has signaled an intention to extend its existing readiness self-assessment program with quantum-specific facility parameters.

For anyone writing a specification, the practical move is to cite EN 50600 or ISO/IEC 22237 as the floor, then attach the OCP checklist as the quantum layer above it. That gives procurement a community-validated set of thresholds to hold a landlord or a contractor to, instead of a set of numbers assembled from vendor emails.

PostQuantum.com’s technical treatment of facility preparation goes deeper into the reference floor plan and the per-modality integration sequences for teams that need the engineering detail.

Building the competence before the hardware

None of this is exotic engineering. It is ordinary structural, electrical, mechanical and safety work held to thresholds that ordinary data center practice never has to meet, coordinated by someone who understands why each threshold exists.

That coordination role is the scarce one. Vendors will supply their own facility requirement documents, and those documents are accurate and narrow: they describe what their machine needs, not how a survey is run, not which of four sign-offs is on the critical path, and not what a landlord should be contractually obliged to deliver. Organizations that get this right have someone in-house who can read a site survey report, argue with a structural engineer about point loading, and tell the difference between a remediation item and a reason to choose a different modality.

Quantum Academy’s engineering programs cover the cryogenic infrastructure, the signal chain and the procurement decisions that this article sits underneath, with worked survey and specification exercises rather than an overview. They are listed at quantumacademy.com/. Engineers mapping this against a career direction rather than a current project will find the adjacent paths set out at QuantumCareers.com.

The scheduling advice is simpler than the engineering. Start the facility work in parallel with hardware procurement, and start the survey before either. A building can be remediated on a known timeline. A delivery date cannot be moved backwards to accommodate a floor that will not hold the weight.