Industries / Technology / Quantum computing
Control and cooling
This is the equipment around the qubits: refrigerators that hold superconducting and silicon chips a few thousandths of a degree above absolute zero, lasers that cool and hold ions and atoms, and electronics that send the pulses that carry out each operation.
4
Milestones achieved, from 1951 to 2010.
2
To watch for: announced, not yet confirmed outside.
2
Open deliverables. Nobody has shipped them.
The road
AchievedWatch forOpen
Achieved
1951
Cooling by mixing helium
Heinz London proposed in 1951 that dissolving helium-3, a rare light form of helium, into ordinary helium-4 could give steady cooling far below one degree above absolute zero, and he set out a definite design with G. R. Clarke and E. Mendoza in 1962. Working refrigerators followed in the mid-1960s at Leiden, at Manchester under Henry Hall and at Dubna under Boris Neganov; sources differ on which came first. The dilution refrigerator is still the only machine that gives continuous cooling to a few thousandths of a degree, and superconducting quantum computers run inside them.
Credit
Heinz London; Das, de Bruyn Ouboter and Taconis (Leiden); Henry Hall (Manchester); Boris Neganov (Dubna)
1978
Ions cooled by laser
Laser cooling slows trapped ions by having them absorb light pushed slightly off their natural colour, so that each absorption takes away a little motion. Proposed for ions by Wineland and Dehmelt in 1975, it was first done in 1978, independently, by David Wineland's group with magnesium ions and by Neuhauser and colleagues, of Toschek's group in Heidelberg, with barium ions. The Nobel committee names it an important step toward controlling single ions, the basis of trapped-ion computers.
Credit
David Wineland and colleagues; Neuhauser, Toschek and colleagues
2007
Refrigerators without liquid helium
Older dilution refrigerators sat in a bath of liquid helium-4 that had to be refilled; cryogen-free ones use an electric cooler called a pulse tube instead. Rob Blaauwgeers and Pieter Vorselman built one for the Low Temperature Laboratory at Helsinki University of Technology, which put it into use in 2007, and founded Bluefors in 2008. An industry guide says Bluefors has since shipped over 1,800 systems and that its KIDE model is used inside IBM Quantum System Two.
Lessondropping liquid helium-4 did not remove the need for helium-3, which is still the working fluid in every dilution refrigerator.
Credit
Rob Blaauwgeers and Pieter Vorselman (Bluefors)
2010
The helium-3 shortage
Most helium-3 comes from the decay of tritium kept for nuclear weapons; in the US this yields about 8,000 litres a year, according to the US Congressional Research Service. After 2001 the US government deployed neutron detectors at its borders that use helium-3, and the stockpile shrank. The federal response began only after the shortage had occurred: an interagency committee rationed the gas, and commercial prices rose from USD 100 to 200 a litre to USD 2,000 or more. The report notes that some cryogenic research can only be done with helium-3.
Lessona supply that is a by-product of another programme can run short without warning, and the response may come only after it has.
Credit
Dana Shea and Daniel Morgan, Congressional Research Service
Watch for
An agent calibrates 108 qubits
Calibration is the daily work of measuring each qubit and tuning the pulses that control it. Huikai Xu, Hai-Feng Yu and colleagues say their Vibe Calibration system, in which language-model agents follow reusable written procedures, calibrated 108 of the 112 qubits on a superconducting processor in 4.7 hours, four to five times faster than by hand, and matched expert results on 14 of 16 qubits checked. Separately, OpenAI said in September 2026 that its GPT-5.6 Sol model, working through Codex, ran routine measurements on a six-qubit chip at MIT but needed experts when signals were weak or noisy.
Achieved whena lab not involved in either study runs the agent on its own chip and reports how many qubits it calibrated and the resulting gate errors.
Claimed by
Xu, Yu and colleagues, June 2026 (preprint); OpenAI with MIT's Engineering Quantum Systems group, September 2026
New sources of helium-3
Bluefors agreed in September 2025 to buy up to 10,000 litres of helium-3 a year for 2028 to 2037 from Interlune, which plans to harvest it on the Moon. In July 2026 Interlune said its Cold Capture process had made 99% pure helium-3 from ordinary Grade A helium on Earth and could, if used widely, triple the US supply.
Achieved whenhelium-3 from either source is delivered to a refrigerator maker and the amount is confirmed by the buyer or a government agency.
Claimed by
Bluefors and Interlune, September 2025; Interlune, July 2026
Open
01
Agent-tuned pulses in the cloud
Using the pulse-level access that Amazon Braket offers on Rigetti's 108-qubit Cepheus processor, have an AI agent calibrate its own single-qubit gate pulses from scratch on a few qubits: find each qubit's frequency, set the pulse strength, and check the result. Then measure the agent's gates by randomized benchmarking (see Benchmarks) against Rigetti's default gates on the same qubits.
The briefCredit · unclaimed02
Robot wiring for cryostat rigs
A robot arm that connects and tightens the small threaded coaxial connectors, called SMA connectors, on a bench mock-up of a cryostat plate, with a network analyser checking each line afterwards. No cryostat is used, only a plate with the same connector layout.
The briefCredit · unclaimed
Open 01 · the brief
Agent-tuned pulses in the cloud
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
Using the pulse-level access that Amazon Braket offers on Rigetti's 108-qubit Cepheus processor, have an AI agent calibrate its own single-qubit gate pulses from scratch on a few qubits: find each qubit's frequency, set the pulse strength, and check the result. Then measure the agent's gates by randomized benchmarking (see Benchmarks) against Rigetti's default gates on the same qubits.
Why it is still open
The agent calibrations reported so far ran in the authors' own labs with full control of the hardware. Other work tests agents in simulation: a July 2026 paper built a physics-based simulator of transmon chips for this purpose. A November 2026 search found no report of an agent calibrating pulses on a public cloud machine, where an outsider gets only the access the provider allows.
Built with what exists
The Amazon Braket SDK and its pulse-level controls on Cepheus-1-108Q, with AWS Cloud Credits for Research; a coding agent such as Claude Code or OpenAI's Codex; standard randomized-benchmarking code.
First step
Run one qubit-frequency measurement on Cepheus by hand through Braket's pulse controls, then give the agent the same task and compare the two results.
Done looks like
On 5 qubits, the agent's single-qubit gates reach an error per gate no more than twice that of Rigetti's default gates measured the same way on the same qubits, by randomized benchmarking on 3 different days, with every agent step logged and published.
Open 02 · the brief
Robot wiring for cryostat rigs
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
A robot arm that connects and tightens the small threaded coaxial connectors, called SMA connectors, on a bench mock-up of a cryostat plate, with a network analyser checking each line afterwards. No cryostat is used, only a plate with the same connector layout.
Why it is still open
A 100-qubit superconducting machine needs 300 to 500 signal lines and a 1,000-qubit machine 3,000 to 5,000, according to an industry guide, which also lists SMA connectors as spare parts because tightening them with a torque wrench wears them out. A 2024 Chalmers master's thesis, written with Saab, found that a collaborative robot could do most of the connecting and disconnecting in a radar test calibration, though some connector types were hard to handle. A November 2026 search found no robot used for cryostat wiring.
Built with what exists
A collaborative arm such as a Universal Robots UR5e, or a low-cost SO-101 arm with Hugging Face's LeRobot software for early trials; a torque-limiting SMA driver held by the arm, so the arm only aligns and turns; a camera; a low-cost vector network analyser. A learned policy trained from human demonstrations is one route (the Note "Five demonstrations per task" covers NVIDIA's GR00T 2, and the Note "No download" covers LeRobot v0.7.0).
First step
Build a flat plate with 16 SMA bulkhead connectors and record 20 human demonstrations of the connect-and-tighten motion.
Done looks like
100 SMA connections on the mock plate, filmed, with a published tally of how many reached the set torque and passed the analyser check; a second person repeats 20 on their own rig.