Marin Ivezic
A superconducting transmon qubit, the design used by most of the large superconducting programs, operates at 10 to 20 millikelvin. The cosmic microwave background, the residual heat of the early universe, sits at 2.7 kelvin. The coldest stage of a dilution refrigerator is therefore roughly 150 times colder than deep space, and it has to hold that temperature continuously while several hundred cables carry microwave control pulses down to a chip whose quantum state survives for tens of microseconds.
This is the part of a quantum build that surprises first-time buyers most. A dilution refrigerator isn’t lab equipment you rack-mount and forget. A mid-range system weighs about three-quarters of a metric ton, takes the better part of a year to arrive, will outlive several generations of the processors it hosts, runs on a helium isotope produced by the decay of nuclear weapons material, and will take an entire quantum operation offline for a week if the building loses power.
We put cryogenics early in our engineering curriculum because of what it constrains. Three properties of the cold environment set the ceiling on everything downstream: how long the refrigerator takes to procure, whether its working fluid can be sourced and recovered, and how many signal lines can be pushed into it before the cooling budget runs out. None of the three is fixed by the quality of the quantum processing unit (QPU) you buy.
Why cooling comes first in the sequence
The common sequencing error looks reasonable from the outside. A team selects a qubit modality, specifies a processor, chooses control electronics, and then goes looking for a refrigerator to put it all in. By that point the refrigerator has become a fitting exercise, and the fitting usually fails in one of two ways: the delivery date pushes the whole program back by two quarters, or the chosen system has no room for the wiring the roadmap will need in three years.
Reverse the order. The cryostat has the longest lead time of any component in the build, the highest single capital cost, and an operating life measured in a decade rather than a refresh cycle. It is the one item nobody replaces to chase a specification bump. Select it against the system you intend to run in five years, then fit the processor and the electronics to it.
One caveat before going further. All of this applies to superconducting and silicon-spin builds, and partially to photonic builds through their single-photon detector subsystems. Neutral-atom machines and most trapped-ion machines need no dilution refrigeration at all. That two of the five major modalities sidestep this entire chapter is itself a procurement argument, and we return to it at the end.
What a dilution refrigerator actually does
A dilution refrigerator is the only technology that delivers continuous cooling at millikelvin temperatures. It works on a property of helium-3 and helium-4 mixtures: below about 870 millikelvin, the mixture separates into two layers, a concentrated phase that is almost entirely helium-3 and a dilute phase that holds roughly 6.6 percent helium-3 in helium-4. Moving a helium-3 atom across that boundary into the dilute phase absorbs energy from its surroundings, in the same way evaporation cools a surface. Pump the helium-3 around a closed circuit through that boundary and the cooling continues indefinitely.
Getting to the starting line is a separate job. A pulse-tube cryocooler, a mechanical closed-cycle cooler with no moving parts in the cold zone, brings the system from room temperature down to about 3 kelvin. The dilution circuit takes over from there. The coldest stage, the mixing chamber, is where the phase boundary sits and where the QPU is mounted, which is why engineers usually call it the MXC plate and quote payload capacity against it.
The market that supplies these systems is small and consolidated. Bluefors in Helsinki is the volume leader, with a product line that runs from single-cryostat research systems up to industrial platforms tall enough to walk into. Oxford Instruments NanoScience in the UK builds modular systems designed around fast sample exchange, which suits groups swapping processors frequently. Maybell Quantum in Denver has built a US-based alternative with integrated vibration isolation. Leiden Cryogenics serves research groups that need geometries nobody else will build. Buyers have less negotiating room here than they expect, and lead times move with the order book rather than with the urgency of the buyer.
Two isotopes, two supply chains
Engineers new to this field routinely conflate helium-3 and helium-4. They are chemically the same element and commercially unrelated.
Helium-4 is the common isotope. It comes out of natural gas wells as a byproduct, cools MRI magnets, and costs single-digit dollars per liter. Its supply is volatile and periodically disrupted, and those disruptions make the news.
Helium-3 is the rare isotope and the working fluid every dilution refrigerator depends on. It doesn’t occur on Earth in commercial quantities. Essentially all terrestrial supply comes from the radioactive decay of tritium held in nuclear weapons stockpiles, with a half-life of 12.3 years, predominantly in the United States. Russian material has been effectively out of Western markets since 2022. It sells for orders of magnitude more per liter than helium-4, and purified research-grade material costs more again.
The consequence for a build team is that helium-3 is a procurement problem with no substitution path and no elastic supply response. Tritium decays at the rate it decays. Quantum computing now competes for that output against neutron detection for nuclear safeguards, medical lung imaging, and fusion research, and the quantum share grows every time somebody commissions another fridge.
Three practices follow, and we teach all three as day-one requirements rather than optimizations.
Negotiate the helium-3 charge inside the cryostat contract. Vendors hold allocation arrangements that individual buyers cannot replicate on the spot market. Buying the fridge and then sourcing the fluid separately is the expensive route.
Commission closed-loop recovery before the first cool-down, not after the first incident. Helium-3 is never vented to atmosphere during servicing. Recovery and purification infrastructure is ordinary gas-handling engineering, and it belongs in the initial capital request.
Hold a warm spare charge on site. Mixture contamination and servicing leaks both happen. A sealed spare cylinder turns a procurement crisis into a maintenance afternoon.
The lunar hedge
The industry’s long-range answer to the deficit is lunar. Helium-3 is deposited in lunar regolith by the solar wind, and Interlune, a Seattle company, is developing extraction hardware to harvest it.
Treat that as a strategic hedge. Nothing about it changes a 2027 build plan. Extraction hardware still has to fly, operate in regolith, and scale, and the company’s own public guidance points at the early 2030s for commercial volumes. Track the milestones, plan the reserve as though they won’t arrive, and you’ll be right either way.
The I/O wiring wall
Every superconducting qubit needs several control and readout lines running from room-temperature electronics down to the chip. As a working rule, budget three to five distinct cryogenic signal paths per qubit once tunable couplers and shared readout are accounted for. A 100-qubit machine therefore needs several hundred lines. A 1,000-qubit machine needs several thousand.
Conventional semi-rigid coaxial cable is 2.19 mm across. Stack a few thousand of those and two things break at once. The physical space at each loading port runs out, and the thermal budget runs out first. Coax conducts heat along its length from warmer stages to colder ones, and every attenuator, the component that damps room-temperature noise before it reaches the qubit, dissipates the noise power it removes as heat at whatever stage it is bolted to. A coax-only wiring tree for a large processor can consume the entire cooling capacity of the 100 millikelvin stage on attenuator load alone, leaving nothing for the QPU.
This is the I/O wiring wall, and it is the binding constraint on superconducting scaling independent of processor quality. A vendor can fabricate a thousand-qubit chip. If the cryostat cannot carry the lines without exceeding its cooling budget, the chip cannot be operated.
The commercial answer is flexible superconducting wiring. Delft Circuits’ Cri/oFlex uses niobium-titanium stripline printed on a polyimide ribbon roughly a third of a millimeter thick, carrying eight channels per ribbon with attenuation and filtering built into the flex structure rather than spliced in by hand. Bluefors has its own high-density flex platform for tight integration with its side-loading ports, and MIT Lincoln Laboratory’s flexible ribbon technology covers low-frequency services such as thermometry and DC gate bias. These are complementary rather than competing: a real wiring tree mixes high-frequency flex, low-frequency flex, and a residue of coax.
The specification rule is simple. Do not accept a coax-only wiring design for any system above about twenty qubits. Converting later means a full warm-up, a complete rewire, and a cool-down, which is weeks of downtime plus the risk of disturbing a working machine. The cost difference at build time is small against the total system cost.
Heat is the same problem as space
Cooling power is quoted per temperature stage, and the integrator’s job is to prove that the total heat load at each stage sits inside the available capacity with margin left over. Ask the vendor for the cooling power curve, capacity against temperature at every stage, and run the calculation before the signal chain design is frozen. Each cable segment, each attenuator, each amplifier contributes a number.
Our heuristic in the engineering track: if projected load at any stage exceeds 80 percent of available cooling power on day one, there is no headroom for a processor upgrade or added instrumentation. Two repairs are available. Redesign the signal chain by reducing attenuation or moving components to warmer stages, or specify a larger cryostat. The second is a capital decision, which is precisely why it belongs at the front of the program.
The operational calendar
An empty system reaches base temperature in under a day. Loaded with a processor, magnetic shielding, and a full wiring tree, budget several days from room temperature to a stable, calibrated machine. Controlled warm-up for service takes a few days more, and it has to be controlled: venting cold hardware to atmosphere puts thermal shock through connectors, seals, and the chip package.
Routine maintenance is predictable. Pulse-tube cold-head service falls due roughly every 18 to 24 months, compressor adsorbers annually, mixture composition checks annually, leak checks quarterly. Plan one scheduled warm-up and service window per year and treat the associated downtime as a fixed operating cost rather than an exception.
The unplanned warm-up is the event that hurts. A power outage, a compressor failure, or an interruption to the chilled-water loop will warm the cryostat, and once the mixing chamber passes 1 kelvin the recovery path is fixed: recover the helium-3, diagnose, repair, cool down again, then run the full recalibration sequence on the processor. That is a week or more of a facility producing nothing, and it is the reason online double-conversion UPS, redundant cooling loops, and automatic shutdown sequencing belong in the facility specification rather than the wish list.
Spares work on the same logic. Long-lead items are the ones to hold: a pulse-tube cold-head kit, compressor adsorber cartridges, the sealed helium-3 charge, vacuum seals, RF connectors, and a spare cryogenic amplifier, since amplifier lead times run to months. The inventory is insurance priced against idle facility days.
What changes before 2030
Two developments are worth tracking, neither of which should alter a build decision today.
Cryo-CMOS co-integration. Control electronics fabricated to operate at cryogenic temperatures, sitting next to the qubits instead of at the top of a long cable run. Intel’s Horse Ridge line operates at around 4 kelvin and addresses many qubits through few physical channels using frequency multiplexing, and several groups are pursuing millikelvin-stage controllers alongside silicon-spin qubits. If this reaches production maturity, the wiring density problem largely dissolves and the cryostat becomes a thermal enclosure for a combined quantum-classical chip. The helium-3 requirement stays, and the heat budget changes shape because the electronics now dissipate at the cold stage.
Helium-3-free millikelvin cooling. Kiutra in Munich builds continuous adiabatic demagnetization refrigerators, which cool using the magnetocaloric effect of paramagnetic salts and need no helium-3 at all. The technology already suits research systems and small qubit counts. Cooling power at millikelvin temperatures remains below what dilution refrigeration delivers, so industrial-scale continuous operation stays with dilution for now.
There is also a route around the isotope entirely. Silicon-spin processors and superconducting photon detectors can run near 1 kelvin, where helium-4 alone will do, and neutral-atom machines need no cryogenics whatsoever. For defense customers and sovereign programs where helium-3 exposure is unacceptable, modality choice is the supply chain control.
Sequencing the procurement
For a superconducting or silicon-spin build, we teach this order:
- Fix the qubit count you intend to reach in five years, not the count you will commission first.
- Select the cryostat class against that number, with cooling power margin and loading-port capacity to match.
- Specify flexible high-density wiring from day one.
- Negotiate the helium-3 charge, recovery infrastructure, and a spare charge inside the same contract.
- Design the facility, including power continuity and cooling redundancy, around the cryostat’s requirements.
- Only then select the processor and the control electronics.
Teams that follow this order spend more at the start and stop being surprised. Teams that follow the reverse order discover the wiring wall in year three, when the fix costs a rebuild.
Where to learn this properly
Cryogenic infrastructure is one of the places where quantum engineering behaves like ordinary industrial engineering: thermal budgets, lead times, spares, and supply chains, with an unfamiliar isotope in the middle of it. Getting it right is a matter of sequencing decisions correctly and holding the vendor conversation with the right numbers in hand.
The Quantum Academy engineering track covers cryostat selection, signal chain and heat-load budgeting, facility preparation, and the procurement sequence above as applied work rather than theory. Browse the programs at quantumacademy.com/.
For a deeper technical treatment of the helium-3 market and the wiring roadmaps behind these constraints, see the cryogenic infrastructure analysis on PostQuantum.com.