Marin Ivezic
Demagnetize a suitable paramagnetic refrigerant from a field of several tesla and, depending on the material and the starting temperature, it can cool below 20 millikelvin. A dilution refrigerator, circulating a mixture of helium-3 and helium-4, holds a processor at roughly the same temperature while depending on an isotope with a constrained supply chain. Magnetic cooling has existed since the 1930s and uses no helium-3 at all. Engineers in our cryogenics sessions ask the obvious question: why is every superconducting quantum computer still bolted to a dilution fridge?
The answer is a lesson in how to read a hardware specification. Temperature is the number the datasheet leads with. Cooling power at that temperature is the number that decides the procurement. This article walks through the physics, the state of the art in 2026, and the questions we teach engineers to ask before they accept a cooling claim.
The specification everyone quotes
A cryostat specification has two halves, and only one of them gets quoted in press releases. Base temperature is how cold the system gets with nothing attached. Cooling power is how much heat the system can absorb continuously while staying at a stated temperature, measured in microwatts.
That heat is not hypothetical. Every coaxial line running from room temperature to the coldest stage carries thermal energy down with it. Attenuators dissipate power. Amplifiers dissipate more. The processor itself contributes. Add qubits and you add lines, and the heat load at the mixing chamber stage, the coldest point in a dilution refrigerator, rises with them.
Dilution vendors publish this number and guarantee it. Bluefors publishes more than 30 microwatts at 20 mK and more than 1,000 microwatts at 100 mK for the XLD1000sl, and more than 3,000 microwatts at 100 mK for its KIDE platform across three cooling units. Oxford Instruments quotes comparable figures for the Proteox series. Dilution cooling power falls roughly as the square of temperature, so a figure quoted at 100 mK tells you very little about performance at 20 mK. A cooling power without a temperature attached isn’t a specification.
How magnetic cooling reaches millikelvin
Adiabatic demagnetization refrigeration, or ADR, exploits the disorder of electron spins. A paramagnetic material contains unpaired electron spins that can point in any direction. The entropy of that spin population, meaning how disordered it is, depends on both temperature and applied magnetic field.
The cycle runs in three steps. First, a strong field is applied while the material sits thermally connected to a bath at a few kelvin, typically supplied by a pulse-tube cooler. The field aligns the spins, their entropy drops, and the heat released by that ordering flows into the bath. Second, a heat switch opens to isolate the material thermally. A heat switch conducts heat well when closed and blocks it almost completely when open. Third, the field ramps slowly down. The spins randomize, entropy rises, and the energy required for that randomization can only come from the thermal energy of the material itself. The material cools.
The effect is immediate and it is finite. Once the spin entropy is exhausted, the material can absorb no more heat and begins to warm. That is the single-shot limit, and it is the reason ADR lost to dilution cooling for quantum computing decades ago.
Single-shot cooling and continuous cooling
Single-shot ADR is a mature product category. Entropy GmbH sells platforms for condensed-matter research, the former HPD product line now sits under Danaher Cryogenics with a two-decade record in national laboratories, and flight-qualified systems from NASA’s Jet Propulsion Laboratory and UCL’s Mullard Space Science Laboratory hold X-ray detectors at 50 mK in orbit. These are excellent instruments, and they are helium-free, mechanically simple and free of any scarce-isotope dependency.
They also cool, hold for a bounded interval, and then warm back up. That suits the duty cycle they serve. A materials scientist measures a sample, warms the stage, swaps the sample and runs again. A space telescope observes, recharges its cooler between windows, and observes again.
A quantum processor has no such duty cycle. Error correction runs continuously, calibration routines fire between computation cycles, and the machine stays between 10 and 20 mK for weeks or months without interruption. A warm-up costs the quantum state, then hours to days of re-cooling, then a full recalibration of every qubit. Single-shot cooling and continuous quantum computation are incompatible operating models, and no amount of improvement to hold time closes that.
Continuous ADR, or cADR, solves this architecturally rather than physically. Two or more ADR stages are chained together through heat switches. One stage cools the load while the other regenerates, re-magnetizing and dumping its heat to the bath. As the active stage approaches its entropy limit, the switches flip and the roles reverse. Cooling never stops. The switches are the hard part, because each one must conduct well, isolate almost perfectly, and survive thousands of cycles without degrading. kiutra, a spin-off from the Technical University of Munich, is the only vendor we know of shipping continuous ADR as a product line.
Inside the 2026 demonstrator
The result that changed the conversation was published in Review of Scientific Instruments from kiutra’s SPROUT project. A four-stage cADR system, built into a 19-inch rack format, held continuous operation below 30 mK and reached 20 mK in single-shot mode, with no helium-3 anywhere in the system. It used both mechanical and superconducting heat switches, and it carried high-density RF wiring supplied by Delft Circuits, enough to operate a 5-qubit superconducting processor.
That last detail is what separates this from earlier demonstrations. Reaching millikelvin in a bare cryostat is a physics result. Reaching it with the wiring, shielding and thermal staging a real processor requires is an engineering result, and it is the first time continuous magnetic cooling and superconducting qubits have shared a cryostat architecture.
Cooling power at temperature
Return to the two-part specification. Temperature: solved. Continuity: solved at small scale. Cooling power at millikelvin, with full wiring installed: not published.
The SPROUT paper does not quantify cooling power at sub-30 mK operation, and 5 qubits present a heat load orders of magnitude below the systems that consume helium-3 today. kiutra’s commercial platforms reflect the same boundary. The L-Type Rapid runs continuously at 300 mK and reaches lower temperatures only in single-shot mode, and the company positions its current products as characterization and testbed systems rather than production hosts. Its LEMON project, funded by the European Innovation Council, targets the modular scale-up.
So the honest comparison cannot be made yet, and that itself is the finding. One side of the market publishes guaranteed microwatts at 20 mK and has done so for years. The other side has published a temperature. Anyone selling continuous magnetic cooling today as a drop-in replacement for a production dilution refrigerator is extrapolating from a demonstrator to a product line without the connecting data. Anyone still calling magnetic cooling a lab curiosity is behind the evidence in the other direction.
Three constraints on the path to a product
The magnet
ADR cycles a field of several tesla on and off, and superconducting qubits are acutely sensitive to stray magnetic fields. Field shifts qubit frequencies, trapped flux causes loss, and a strong enough field quenches the superconducting film outright. Shielding and geometric separation handled this at 5 qubits. At larger counts the shielding mass grows and the thermal links between the shielded module and the qubit stage introduce heat of their own, which comes out of the cooling budget the system is meant to provide. Silicon spin qubits are less exposed here, since they already operate in a deliberately applied field.
The refrigerant
Conventional paramagnetic salts contain water of crystallization and degrade under vacuum and thermal cycling. The push toward quantum applications has produced water-free alternatives. KBaYb(BO₃)₂, a geometrically frustrated ytterbium borate reported by Tokiwa and colleagues in Communications Materials in 2021, reaches at least 22 mK and survives bakeout for cryostat integration. NH₄GdF₄, reported in the Journal of the American Chemical Society in 2025, shows a magnetic entropy change of up to 38.2 J·kg⁻¹·K⁻¹, well above the long-standing benchmark material. A 2025 review in Accounts of Chemical Research from Xiamen University sets out the central trade-off: materials that order at very low temperatures have weak magnetic interactions and therefore little entropy to spend, while materials with large entropy change tend to order higher and raise the temperature floor. These are research-grade crystals. Growing them in kilogram quantities with reproducible entropy profiles is a separate problem, and it hasn’t been solved at production scale.
The duty cycle
A dilution refrigerator runs for months between maintenance events. A cADR system cycles its stages continuously, and every cycle stresses the heat switches, the magnets and the thermal joints. No multi-thousand-hour reliability dataset has been published for continuous magnetic cooling under processor load. Dilution refrigeration has decades of accumulated field time behind it, and that record is itself part of what a buyer is purchasing.
Two technologies that change the arithmetic
Cryogenic CMOS attacks the problem from the opposite end. Rather than raising the cooling power of the refrigerator, it lowers the heat load by moving control electronics into the cold environment next to the qubits, cutting the cable count from room temperature. Intel’s Horse Ridge line, SEEQC’s single-flux-quantum logic and Diraq’s silicon-spin integration all pursue this. It helps both cooling technologies, and it helps magnetic cooling disproportionately, because magnetic cooling’s deficit is power rather than temperature.
Better helium-3 management is the unglamorous competitor to both. Closed-loop recovery during servicing, reduced-charge dilution circuits and recycling from decommissioned systems extend the effective supply with no new technology and on a shorter timeline than any scale-up programme.
Five questions for a vendor
When a cooling claim crosses your desk, these five answers separate a demonstrator from a machine.
- What is the continuous cooling power, in microwatts, at 20 mK?
- Is that figure continuous or single-shot at that temperature?
- How many RF lines does the wiring support, and was the cooling power measured with them installed and driven?
- What peak field does the magnet cycle, and what is the measured stray field at the sample stage?
- How many thermal cycles have the heat switches accumulated, under what load, and with what measured degradation?
A vendor who answers all five is describing a product. A vendor who answers only the first, or answers it at 100 mK, is describing a temperature.
Where to take this next
Cryogenic infrastructure is one of the places where quantum engineering decisions turn on specification sheets rather than physics arguments, and reading those sheets accurately is a learnable skill. Our engineering programs are built around exactly this kind of evaluation work, and the current catalogue is at quantumacademy.com/.
For the deeper technical treatment of this comparison, including the vendors and the timeline, Marin Ivezic’s analysis at PostQuantum.com is the companion piece to this one.