Industries / Technology / Quantum computing
Qubits and hardware
This is the work of making qubits, the physical parts that hold quantum states, and joining many of them in one machine.
7
Milestones achieved, from 1995 to 2025.
3
To watch for: announced, not yet confirmed outside.
2
Open deliverables. Nobody has shipped them.
The road
AchievedWatch forOpen
Achieved
1995
The first quantum logic gate
At NIST in Boulder, Chris Monroe, David Wineland and colleagues used laser pulses to run a controlled-NOT gate, the basic two-qubit operation, on a trapped beryllium ion, using the ion's internal state and its motion as the two qubits. They followed a scheme Ignacio Cirac and Peter Zoller had published earlier that year. The Nobel committee's background paper for Wineland's 2012 physics prize credits the group with the first experimental two-qubit gate and records that Rainer Blatt's group in Innsbruck ran one between two separate ions in 2003.
Credit
Chris Monroe, David Wineland and the NIST Ion Storage Group; Ignacio Cirac and Peter Zoller (theory)
1999
A qubit on a chip
Yasunobu Nakamura, Yuri Pashkin and Jaw-Shen Tsai steered a tiny superconducting circuit, a "single-Cooper-pair box", between two quantum states with a short voltage pulse and read out the result. It showed that a qubit could be built as a solid-state electronic device.
Credit
Yasunobu Nakamura, Yuri Pashkin and Jaw-Shen Tsai
2007
The transmon
Jens Koch, R. J. Schoelkopf and colleagues proposed the transmon, a superconducting qubit reshaped so that stray electric charge barely disturbs it. Later processors use it; for example, the 112-qubit chip that an AI agent calibrated in 2026 (see Control and cooling) is built from transmons.
Credit
Jens Koch, Terri M. Yu, Jay Gambetta, A. A. Houck, D. I. Schuster, J. Majer, Alexandre Blais, M. H. Devoret, S. M. Girvin and R. J. Schoelkopf
2016
Atoms sorted into place
Two separate teams showed how to fill arrays of optical tweezers, tightly focused laser beams that each hold one atom, with no gaps: a camera finds the empty traps and a moving tweezer shuffles atoms into them. One team built defect-free arrays of more than 50 atoms in under 0.4 seconds; the other built arbitrary two-dimensional patterns of up to 50. Moving atoms with tweezers is also how the 2023 logical processor and the 2025 Caltech array below rearrange their qubits.
Credit
Manuel Endres, Mikhail Lukin and colleagues; Daniel Barredo, Antoine Browaeys and colleagues
2021
A Majorana result retracted
In 2018 a team led by Leo Kouwenhoven reported in Nature a conductance signal it read as evidence of Majorana zero modes, the states on which topological qubits would be built. After Sergey Frolov and Vincent Mourik found that the raw data shared with them did not match the published figures, the authors re-analysed the data, found corrections they had not disclosed and a recalibration that moved the signal by 8%, and retracted the paper in March 2021, apologising for "insufficient scientific rigour".
Lessona claim that rests on one group's processing of its own data stays open until the raw data are public and outsiders have checked them.
Credit
Leo Kouwenhoven and co-authors; Sergey Frolov and Vincent Mourik (the check)
2021
Silicon passes 99%
Two groups, one led by Lieven Vandersypen and one by Seigo Tarucha, separately reported in 2021 two-qubit gates with 99.5% fidelity in electron-spin qubits in silicon. Fidelity measures how close an operation comes to the ideal; 99.5% is about one error in 200 operations, better than the roughly 99% that the surface code, the leading error-correcting code, needs. Spin qubits matter because they can be made with chip-industry methods.
Credit
Xiao Xue, Lieven Vandersypen and colleagues; Akito Noiri, Seigo Tarucha and colleagues
2025
6,100 atoms in one array
Manuel Endres's group at Caltech trapped more than 6,100 caesium atoms in an array of about 12,000 tweezers. The qubits' coherence time, how long they keep a quantum state, was 12.6 seconds, a record for this kind of qubit in a tweezer array; single-qubit operations reached 99.98% fidelity, and atoms were moved across the array without losing their state. The paper reports single-qubit operations and transport, not the two-qubit gates a computation also needs.
Credit
Hannah Manetsch, Gyohei Nomura, Elie Bataille, Kon Leung, Manuel Endres and colleagues, Caltech
Watch for
7,500 operations on one chip
IBM says its 120-qubit Nighthawk r2 processor, launched on 31 August 2026, gave accurate estimates from circuits of more than 7,500 operations and runs over 100,000 circuits a second, about 25 times its older Heron processors. The estimates use an error-mitigation method called probabilistic error amplification, which corrects results after the run instead of preventing errors. In November 2026 IBM said its next chip, Nighthawk r3, will be built for 10,000 operations. IBM ran the test itself, and the launch post links no paper or dataset for that run. The site's Note "7,500 operations" covers this claim.
Achieved whenIBM publishes the circuits and target values, and a group outside IBM reproduces the estimate on a rented r2.
Claimed by
IBM, August and November 2026
Topological qubits, second try
Microsoft says its Majorana 2 processor keeps topological qubit states for more than 20 seconds, over 1,000 times longer than its earlier devices, after it replaced aluminium with lead in the chip. In June 2026 Nature published a peer-reviewed critique by Henry Legg of the University of St Andrews arguing that Microsoft's February 2025 Nature paper rested on coding errors and a flawed tune-up procedure.
Achieved whena lab outside Microsoft measures the same signature in its own device, or Microsoft releases its raw data and an outside analysis agrees with it.
Claimed by
Microsoft, June 2026
Four nines on ions
IonQ says it reached two-qubit gate fidelity above 99.99%, one error in 10,000, using electronic rather than laser control of trapped ions, ahead of the 99.97% that Oxford Ionics, now part of IonQ, reported in 2024. The figures come from IonQ's own technical papers.
Achieved whenthe papers pass peer review and an outside party measures 99.99% on IonQ hardware.
Claimed by
IonQ, October 2025
Open
01
Agent-designed qubit chip
Give an AI coding agent target values for a small superconducting chip, for example four transmon qubits with set frequencies and couplings, each with a readout resonator, and have it write the layout, run the electromagnetic simulation and revise until the simulated values hit the targets, with no human edits. A second person then reruns the simulation in a different solver. A later stage sends the layout to a fabrication service and measures the chip.
The briefCredit · unclaimed02
Announced versus measured
A public table of every processor announced since 2023 with 100 or more qubits. For each, list the error figures published, who measured them, whether they cover the whole chip or a few qubits, and whether anyone outside the maker has measured the machine.
The briefCredit · unclaimed
Open 01 · the brief
Agent-designed qubit chip
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
Give an AI coding agent target values for a small superconducting chip, for example four transmon qubits with set frequencies and couplings, each with a readout resonator, and have it write the layout, run the electromagnetic simulation and revise until the simulated values hit the targets, with no human edits. A second person then reruns the simulation in a different solver. A later stage sends the layout to a fabrication service and measures the chip.
Why it is still open
Parts of the work are automated without agents. SQuADDS, an open database from Eli Levenson-Falk's group, gives starting designs, many of them checked against measured devices, that the open-source Qiskit Metal package generates and finite-element solvers simulate. A November 2026 search of arXiv for language-model agents in superconducting chip layout found no such paper.
Built with what exists
Qiskit Metal and SQuADDS for layouts; scqubits for qubit energies; Palace, an open-source 3D electromagnetic solver from AWS, or Ansys HFSS for simulation; a coding agent such as Claude Code running a model such as Claude Fable 5.2 (the Note "Three working devices" covers the same question for circuit boards).
First step
Install Qiskit Metal and SQuADDS, pick one experimentally checked SQuADDS design, and ask the agent to reproduce its simulated frequencies from the target values alone.
Done looks like
For 10 different target sets, the agent's layouts land within 2% of each target qubit frequency and 10% of each target coupling in simulation, with no human edits; a second person reproduces 3 of the 10 in a different solver.
Open 02 · the brief
Announced versus measured
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
A public table of every processor announced since 2023 with 100 or more qubits. For each, list the error figures published, who measured them, whether they cover the whole chip or a few qubits, and whether anyone outside the maker has measured the machine.
Why it is still open
Qubit counts and measured quality are reported apart. IBM says it has demonstrated chips with up to 1,121 qubits, while the Heron r3 it rents has 156; IonQ's 99.99% figure is a two-qubit gate measurement, not a figure for a whole processor. Metriq, an open benchmarking platform from the Unitary Foundation, has published results from more than ten machines across several makers. A November 2026 search found no table that links each announced chip to an outside measurement.
Built with what exists
Makers' releases and arXiv papers; Metriq's open dataset; IBM's and Amazon Braket's public device pages; a spreadsheet and a small script; an AI research agent to draft entries, each checked by a person against its link.
First step
List the processors of 100 or more qubits announced in 2025 and 2026 by IBM, Google, IonQ, Quantinuum, Rigetti, USTC and the Tianyan Quantum Group, and fill in what each has published.
Done looks like
At least 25 processors, every figure linked to its source; a second person checks 10 rows and agrees; the table is updated monthly for six months.