How do you build a qubit from circuits, ions and atoms?

Each qubit type gets the same eight lines:

1. What the qubit physically is.
2. How gates work.
3. How you read it out.
4. What keeps it isolated.
5. Its errors.
6. How it scales.
7. Who builds it.
8. Key numbers, dated.

Take a loop of metal, such as aluminium, a fraction of a millimetre across, printed on a chip. Cool it to about a hundredth of a degree above absolute zero and it superconducts: current flows with no resistance at all. The whole circuit then behaves like one big quantum object, with energy levels like an atom's (chapter 1). The 2025 Nobel Prize in Physics went to John Clarke, Michel Devoret and John Martinis for showing exactly this in the 1980s.

A special component called a Josephson junction, two pieces of superconductor separated by a thin insulating layer, makes the levels unevenly spaced. That lets engineers use just the bottom two as 0 and 1. The most common design, the transmon, came from Yale in 2007.

• Gates: short microwave pulses, tens of billionths of a second long, sent down wires into the fridge. Two-qubit gates use small couplers between neighbouring qubits.
• Readout: each qubit sits next to a tiny microwave resonator whose note shifts slightly depending on whether the qubit is 0 or 1.
• Isolation: a dilution refrigerator, the golden "chandelier" you see in photos. The chip itself sits at the bottom, about the size of a fingernail.

Strengths: very fast gates, and chips made with tools the chip industry already has.

Weaknesses: qubits keep their state for only tens to hundreds of millionths of a second; each qubit only talks to its neighbours; and every qubit needs wiring down into the fridge, which gets hard at large scale.

Google, IBM, Rigetti, IQM in Finland, OQC in the UK, Alice & Bob in France, and USTC in China, among others.

• Google Willow (December 2024): 105 qubits; the chip that showed error correction below threshold (chapter 8).
• IBM Nighthawk (November 2025): 120 qubits. IBM's roadmap aims for Starling in 2029: 200 logical qubits running 100 million operations.
• Best two-qubit accuracy reported: 99.93% (an IQM-led team, 2025).
• DARPA's final stage (October 2026): the US defence agency DARPA picked IBM for hands-on hardware testing in its benchmarking programme (chapter 11).

Take a single atom, such as ytterbium or barium, and knock off one electron so it has an electric charge. Now it's an ion, and electric fields can hold it still in a near-perfect vacuum, floating above a small chip called a trap. Two of the ion's energy levels form the qubit.

Every ion of the same element is identical: nature makes them perfect copies, which is one reason ion qubits are so accurate.

• Gates: carefully tuned laser or microwave pulses. The ions in a trap push on each other and share a gentle wobbling motion, and gates use that shared wobble to link any pair. So every qubit can talk to every other qubit directly.
• Readout: shine a laser on the ion. If it's in one state it glows; in the other it stays dark. A camera counts the glows.
• Isolation: ultra-high vacuum and careful shielding. Some systems also cool the trap, but not to the extremes superconducting chips need.
• Moving ions: newer designs shuttle ions between zones of the chip, like trains through junctions, to scale up.

Strengths: the highest accuracy of any type so far, very long-lived qubits (one lab estimates a single ion can hold its state for over an hour), and every qubit connects to every other.

Weaknesses: gates are much slower than superconducting ones, and it's hard to fit more than a few hundred ions in one machine without linking machines together.

Quantinuum (US and UK), IonQ (US, which bought Oxford-based Oxford Ionics in 2025), and AQT in Austria.

• Quantinuum Helios (November 2025): 98 qubits, every pair connected, two-qubit accuracy 99.921%. It also ran 48 fully error-corrected logical qubits (chapter 8).
• IonQ Tempo (2025): 100 qubits. IonQ also reported two-qubit accuracy above 99.99% in October 2025 (company claim with a preprint).
• Roadmaps: Quantinuum aims for a fully fault-tolerant machine, Apollo, in 2029; IonQ aims for millions of physical qubits by 2030.
• DARPA's final stage (October 2026): the US defence agency DARPA picked IonQ for hands-on hardware testing in its benchmarking programme (chapter 11).

Here the atoms keep all their electrons, so they have no charge and electric fields can't hold them. Instead, tightly focused laser beams, called optical tweezers, grip each atom like a pair of tweezers made of light. Hundreds or thousands of tweezers make a grid of atoms, such as rubidium or caesium, in a vacuum chamber.

• Gates: single-qubit gates use lasers. For two-qubit gates, a laser kicks an atom into a hugely swollen state called a Rydberg state, where its outer electron orbits so far out that the atom feels its neighbours from micrometres away. That interaction links two atoms.
• Moving atoms: the tweezers can carry atoms around the grid mid-calculation, so any atom can be brought next to any other.
• Readout: atoms glow or stay dark under a laser, and a camera photographs the whole grid at once.

Strengths: the largest arrays of any type, long-lived qubits, and the flexibility of moving atoms around.

Weaknesses: each round of operations is slow, and atoms sometimes fall out of their tweezers and have to be replaced.

QuEra (US, with Harvard and MIT), Atom Computing (US), Pasqal (France) and Infleqtion (US).

• Largest array: 6,100 atoms held at once, at Caltech (September 2025).
• Error correction: Harvard and QuEra ran 48 logical qubits in December 2023, and in November 2025 showed every ingredient of fault tolerance working together on 448 atoms (chapter 8).
• Best two-qubit accuracy: 99.56% in a commercial system (Atom Computing, 2024); around 99.85% in a 2026 Harvard preprint.
• Roadmaps: QuEra aims for 256 logical qubits in 2028. Pasqal's 2024 target of 10,000 qubits by 2026 has moved to 2029, a reminder that roadmaps slip (chapter 11).
• DARPA's final stage (October 2026): the US defence agency DARPA picked Atom Computing for hands-on hardware testing in its benchmarking programme (chapter 11).