Marin Ivezic
The trick that worked last time
In September 2011, the U.S. Government Accountability Office told Congress how the previous helium-3 crisis ended. Report GAO-11-753 described federal demand running near 80,000 liters in 2008 against a supply that came nowhere near meeting it, and it described the fix. Boron-10 lined proportional counters had passed field testing, and the first radiation portal monitors built around them could be deployed in fiscal 2012. The largest consumer of helium-3 on Earth redesigned the detector until the gas was no longer part of the design.
Quantum computing can’t copy that move, and the reason is physical rather than commercial. Those detectors used helium-3 for its neutron capture cross-section, a property that boron-10 and lithium-6 also have. A dilution refrigerator uses helium-3 for its quantum statistics. Below about 0.87 kelvin, a mixture of helium-3 and helium-4 separates into two layers, one rich in helium-3 and one dilute. Pumping helium-3 atoms across that boundary absorbs heat, roughly the way evaporation cools a surface, and it keeps absorbing heat down to a few millikelvin where nothing else does. Helium-3 is the only stable isotope with a fermionic nucleus that stays fluid at absolute zero, and no other element in the periodic table can stand in for it.
So instead of one clean substitution, the field is running every other play at once. Harvesting more tritium, breeding tritium deliberately, filtering helium-3 out of ordinary helium, moving processors to warmer operating points, reviving magnetic cooling from the 1930s, building refrigerators onto the chips themselves, and mining the Moon.
None of these replaces helium-3 on the timeline that superconducting and spin-qubit hardware is scaling on. The realistic outcome is a portfolio, and this is a cost and logistics problem for two modalities rather than a constraint on the field. Trapped ions, neutral atoms, and photonic processors route around the millikelvin stage entirely. Anyone selling a quantum timeline built on a gas shortage is describing one supply chain, not an industry.
Three exits, and why the arithmetic differs
Every response to a scarce input falls into one of three categories. Make more of it, need less of it, or replace the process that consumes it.
For helium-3 the three separate unusually cleanly, because the gas is a working fluid rather than a fuel. A dilution refrigerator recirculates its charge indefinitely. Demand comes from new machine builds plus losses during service, not from consumption per computation, and a mid-size system charge runs to tens of liters. That distinction changes how you grade the exits. Supply-side routes compete against demand that scales with the installed base, while demand-side routes attack the requirement for an installed base of dilution refrigerators in the first place.
We teach this framing before any of the technology, because engineers who skip it tend to evaluate the routes against each other when the routes are aimed at different terms in the same equation.
Making more helium-3
Harvesting tritium that already exists. Helium-3 on Earth is a decay product of tritium, which has a 12.3-year half-life. One number governs every supply-side plan, and it deserves stating plainly: each kilogram of tritium in storage evolves about 420 liters of helium-3 per year at standard conditions. Not per half-life. Per year, every year, declining by about 5.6 percent annually unless the inventory is topped up. Whoever holds tritium holds an annuity.
Two civilian sites collect that annuity today. Ontario Power Generation’s Darlington Tritium Removal Facility (TRF) has been stripping tritium from CANDU heavy water since 1990, and its subsidiary Laurentis Energy Partners began extracting helium-3 from the stored inventory in 2021, the first civilian source outside weapons programs. South Korea runs the second facility at Wolsong. Romania’s Cernavoda station is building a third, with commissioning tests scheduled for 2026, and the Romanian design work has explicitly contemplated helium-3 as a revenue line rather than a byproduct.
This route already works, uses nuclear infrastructure that exists, and gets extended for decades by the CANDU refurbishment wave. It also has a hard ceiling. Civilian tritium production is a few kilograms a year, which caps new helium-3 evolution in the low thousands of liters annually against a five-figure shortfall. The tritium now has a competing buyer, too. Fusion developers want the same kilograms as startup fuel, and every gram burned in a tokamak is 7.4 liters of future helium-3 leaving the cryogenics supply chain. Watch for commercial extraction at Wolsong and Cernavoda on the Laurentis model, and for any agreement that taps the helium-3 accumulated in Darlington’s stored tritium.
Breeding tritium on purpose. This is the route almost nobody has modelled, and the infrastructure for it already runs. The United States breeds tritium deliberately by irradiating lithium-6 inside Tennessee Valley Authority reactors at Watts Bar, in tritium-producing burnable absorber rods (TPBARs) that ship to the Savannah River Site for extraction. Helium-3 recovery is already an explicit mission there, collected when aging weapons reservoirs are serviced.
The decay arithmetic does the rest. A gram of tritium yields about 7.4 liters of helium-3 once it has fully decayed, roughly half of that inside the first 12.3 years. A civilian program breeding one kilogram a year would reach an evolution rate near 7,500 liters a year at steady state, which would rival today’s entire civilian output. Ten years in, though, the same program delivers about 43 percent of that. Adding 10,000 liters a year within a decade needs something like three kilograms of tritium bred annually, which would commit a large share of the existing irradiation capacity to the job.
Two things block this, and neither is technical. The rods compete directly with the weapons mission for reactor positions, and no private buyer can finance a 12.3-year half-life on venture timelines. The plausible trigger is a fusion-fuel expansion with helium-3 recovery attached contractually as the co-product that improves the program’s economics. The fusion lobby may turn out to be the accidental patron of quantum cryogenics.
Separating helium-3 from ordinary helium. The least glamorous supply route may also be the quickest. Ordinary helium isn’t isotopically pure. Atmospheric helium runs about 1.4 parts per million (ppm) helium-3, and helium recovered from natural gas typically carries 0.05 to 0.3 ppm, poorer than air because crustal helium is mostly radiogenic helium-4. Dilute, yes, and it’s already flowing through the world’s liquefaction plants and being sold at helium-4 prices.
The physics has been demonstrated at laboratory scale for generations. Below the 2.17 kelvin lambda point, helium-4 becomes a superfluid and passes through porous superleaks that normal-fluid helium-3 cannot follow, and cryogenic distillation, heat flush, and adsorption all separate the isotopes by other means. What changed is the price. Analyses from the last decade concluded that separation needed helium-3 prices several times higher than the market then paid, and the market has since moved into that range for purified material.
At 0.2 ppm feed and 50 percent recovery, netting 1,000 liters of helium-3 a year means processing about 10 million cubic meters of helium, a mid-single-digit share of world production, through one add-on train. That isn’t absurd. It’s a finishing column bolted to the liquid stream of a large liquefier. It does mean the route runs through the handful of companies that own liquefaction capacity, none of which has announced an isotope program, and it inherits every bit of helium-4’s own supply fragility.
Fusion, and two wildcards. Deuterium-deuterium fusion produces tritium and helium-3 in roughly equal branches, and Helion plans to breed its own helium-3 that way rather than buy it. Sustained net-energy D-D operation has not been demonstrated by anyone, and the deuterium-helium-3 regime needs temperatures several times higher than deuterium-tritium. Until that changes, fusion is a competitor for helium-3’s parent isotope rather than a producer of the gas.
Russia historically supplied a meaningful share of Western helium-3 and has been effectively excluded since 2022, so a thaw would reset the supply arithmetic overnight and collapse the prices that justify every capital-intensive alternative above. Lunar extraction has real contracts and real excavation prototypes, and every step past excavation remains unproven, with no commercial volumes plausible before the early 2030s.
Needing less helium-3
Inventory discipline. The cheapest liter is the one that never leaks. Because dilution refrigerators recirculate, industry demand decomposes into new-system charges plus service losses, and both respond to unglamorous engineering. Leak-tight gas handling, closed-loop recovery during maintenance, reclamation from decommissioned systems, and circuit designs that need less gas per watt of cooling power. None of this makes news. All of it compounds, and for anyone operating fridges today it’s the only lever with an immediate payback.
One-kelvin platforms. Not every quantum workload needs 10 millikelvin, and vendors have started segmenting their product lines accordingly. Helium-4-only systems now ship with hundreds of milliwatts of cooling power in the 1 to 1.2 kelvin band, containing no helium-3 anywhere in the circuit. They target workloads that live comfortably there, notably superconducting nanowire single-photon detector (SNSPD) arrays for photonic processing, and parts of the silicon spin qubit stack. Every system that moves to a helium-4-only platform leaves the helium-3 demand curve permanently.
Qubits that run warm. The most consequential demand-side work is happening inside the qubits. In 2020, Andrew Dzurak’s group at UNSW demonstrated a silicon processor unit cell operating at 1.5 kelvin, and a Delft team showed universal logic in warm silicon qubits in the same issue of Nature. In 2024, Huang and colleagues at Diraq closed most of the remaining quality gap, reporting 99.85 percent single-qubit and 98.92 percent two-qubit gate fidelity above one kelvin, with initialization and readout at 99.34 percent, using an algorithmic initialization protocol that prepares pure states even though thermal energy exceeds the qubit energies.
The motivation is straightforward. Cooling power at one kelvin exceeds cooling power at 20 millikelvin by orders of magnitude, which is why the Diraq work frames warm operation as a scaling prerequisite rather than a curiosity.
The caution is the fidelity tax. A two-qubit gate at 98.92 percent is a genuine achievement at that temperature and still falls below the 99.5 percent region where surface-code overheads, meaning the number of physical qubits you must spend per logical qubit, become tolerable. Warm qubits trade cryogenic scarcity for error-correction overhead rather than eliminating either. Two-qubit fidelity at 99.5 percent or better, above one kelvin, in foundry-fabricated silicon, would widen this exit into a highway and would erase a large fraction of every published helium-3 demand forecast.
The last shortage argues both ways here. Security and safeguards demand was engineered away within about three years of the price signal, and hyperpolarized lung MRI moved from helium-3 to xenon-129, approved by the FDA in December 2022. Markets facing a fifty-fold price increase find substitutes fast. Those substitutions all worked because other nuclei share the property being exploited, and no other nucleus shares this one. Quantum computing can substitute the machine or the operating point. It can’t substitute the isotope.
Replacing the millikelvin stage
Continuous magnetic cooling. Adiabatic demagnetization refrigeration (ADR) cools by aligning magnetic moments in a salt or intermetallic under a strong field, then removing the field so that the material draws heat from its surroundings as the moments randomize. Continuous ADR (cADR) chains stages so the cycle never stops, and it’s the only demonstrated technology reaching quantum-relevant temperatures with no helium-3 at all.
The constraint has never been temperature. It’s cooling power under realistic wiring heat loads. A large dilution unit delivers tens of microwatts at 20 millikelvin, laboratory demonstrations sit well below that, and cycling multi-tesla magnets next to flux-sensitive qubits is a shielding problem that grows with the machine. kiutra’s announced X-Type architecture is specified at 20 microwatts at 20 millikelvin, a 12 millikelvin base temperature, and up to 768 wiring lines, with first customer deployments in 2027. On paper that closes most of the power gap. Independent data from those deployments, at spec, under real loads, is what would turn a demonstrator into a competitor. The thousand-qubit class stays dilution territory either way.
Cooling from inside the chip. The most interesting attack inverts the problem. Instead of a colder box around the processor, put refrigeration into the processor. A Chalmers and University of Maryland team demonstrated a quantum absorption refrigerator built from superconducting circuits, cooling a transmon qubit using nothing but a thermal gradient between two engineered microwave baths. It runs autonomously, with no control pulses and no feedback, and reaches an effective temperature near 22 millikelvin, colder than the surrounding cryostat provides. Finnish groups at Aalto and VTT are pursuing the same goal through electronic tunnel-junction coolers.
These devices cool qubit populations at picowatt scale, and they do it very well. They don’t cool substrates, attenuators, cabling, or amplifier chains, and they do nothing about thermal photons travelling down the lines, so they can’t replace the platform refrigerator. What they could do is relax its specification. If on-chip machines handle initialization and residual excitation locally, the surrounding bath may not need to be at 10 to 20 millikelvin. Continuous ADR already holds 300 millikelvin in shipping products, and helium-4 systems reach one kelvin with no helium-3 at all. A multi-qubit processor running error correction with on-chip cooling in a bath above 100 millikelvin would be the result to watch for.
What doesn’t work. Four routes still arrive regularly in our inbox and none of them go anywhere. Helium-4 substitution fails at the level of quantum statistics, since the two isotopes do physically different jobs. Pomeranchuk cooling consumes helium-3 itself, which makes it a customer rather than an alternative. Laser cooling of solids stalls near 100 kelvin, four orders of magnitude above where qubits operate. Nuclear demagnetization reaches microkelvin and needs a millikelvin precooler underneath it, so it extends the temperature floor, which was never the binding constraint, and adds nothing to cooling power, which is.
The scoreboard
| Route | Where it stands | What would change the grade |
|---|---|---|
| CANDU tritium harvesting | Operating at Darlington, Cernavoda arriving | Commercial helium-3 sales from Wolsong and Cernavoda; a deal on Darlington’s accumulated inventory |
| Purpose-bred tritium | Infrastructure exists, no civilian program | A long-dated offtake attached to expanded TPBAR irradiation |
| Fusion D-D breeding | Undemonstrated | Sustained net-energy D-D operation |
| Isotope separation from helium | Physics proven, no industrial column built | A pilot column at a flagship liquefier with published recovery data |
| Lunar extraction | Prototypes and contracts, demo pending | The extraction-and-return demonstration succeeding |
| Recovery and lean inventory | Available now | Nothing. Do it regardless |
| One-kelvin helium-4 platforms | Shipping | Broader SNSPD and spin-qubit adoption |
| Warm qubits | 98.92% two-qubit fidelity above 1 K | 99.5% two-qubit fidelity above 1 K in foundry silicon |
| Continuous ADR | Shipping at 300 mK, millikelvin systems announced for 2027 | Independent customer data at the announced X-Type numbers |
| On-chip cooling | Single-qubit reset to 22 mK | Multi-qubit error correction in a bath above 100 mK |
What to specify now
If you run or buy cryogenic hardware, your helium-3 strategy for the next three years is written entirely in the demand column. Put charge volume, guaranteed loss rate during service, and recovery provisions into the specification of every refrigerator you procure, and treat them as commercial terms rather than datasheet trivia. Qualify a helium-4-only platform for every workload that tolerates one kelvin, because those workloads leave the problem permanently. Open the supply relationship, whether with a harvesting operator or a national isotope program, before the year you need the gas.
Then track the cADR roadmap the way you’d track a competitor’s. If the announced 20-microwatt specification holds under customer wiring loads in 2027, your next characterization cryostat has no reason to contain helium-3 at all.
For anyone shaping national programs, two files on this list get almost no attention and are the closest to delivery. Civilian tritium breeding as a co-product of fusion fuel policy, and separation capacity at allied helium liquefiers. Both are conventional industrial programs hiding behind an exotic-sounding isotope, and both are years nearer than lunar extraction.
Where this fits in the engineering stack
Cryogenics is where a lot of quantum hardware programs discover that their real constraints are procurement constraints. The physics of a dilution refrigerator takes an afternoon to understand. Reading a cooling-power specification against a wiring budget, grading a supply route by what it can actually deliver and when, and writing a fridge contract that survives a shortage take rather longer, and they’re the skills that separate a program that scales from one that waits.
That’s the layer our engineering programs are built around, from cryogenic infrastructure through the modality trade-offs that decide whether you need a millikelvin stage in the first place. You can see the full range at quantumacademy.com/. For the deeper technical treatment of each supply route, including the sourcing behind the numbers here, the original analysis on PostQuantum.com goes considerably further into the physics.