A quantum computer is not one technology but a contest among several. Companies and laboratories around the world are betting on very different ways of making qubits, and it is still unclear which will win, or whether different approaches will suit different jobs. In 2000 the physicist David DiVincenzo listed what any approach needs: qubits that can be made in large numbers, set to a known starting state, kept coherent long enough, operated on with a full set of gates, and measured reliably. Every design trades these off differently.

Superconducting circuits

Superconducting qubits are tiny electrical circuits that behave quantum mechanically when cooled to about a hundredth of a degree above absolute zero, inside a machine called a dilution refrigerator. That is colder than deep space. Google and IBM lead this approach. Its strengths are fast gates and the ability to make chips with methods borrowed from the electronics industry. Its weaknesses are short coherence times and a tangle of wiring that grows with every qubit.

Trapped ions and neutral atoms

Trapped-ion qubits are single charged atoms held in place by electric fields inside a vacuum and controlled with lasers or microwaves. Companies such as Quantinuum and IonQ use them. Because every atom of an element is identical, ions make very clean qubits with some of the lowest error rates anywhere, though their gates are slower and large traps are hard to build. A related approach uses atoms with no charge. Neutral atom machines hold uncharged atoms in arrays of tightly focused laser beams called optical tweezers, which can arrange hundreds or thousands of atoms and even move them around mid-calculation. QuEra, Pasqal and Atom Computing are among the companies pursuing this route.

Light, silicon and more exotic bets

Photonic machines, built by firms such as PsiQuantum and Xanadu, encode qubits in particles of light, which barely interact with their surroundings and can travel through fiber; the difficulty is making photons interact with each other and not lose them. Spin qubits store information in single electrons trapped in silicon, hoping to use the chip industry's existing factories. Microsoft is pursuing topological qubits, which would store information in a way naturally protected from noise; in February 2025 it announced a chip based on this idea, but many physicists say the evidence that it works as claimed is not yet convincing. D-Wave sells quantum annealers, special machines for optimization problems, and debate continues over whether they beat classical methods.

Why errors are so hard

Ordinary computers protect data with redundancy: store three copies of a bit and take a vote. Quantum computers cannot do that. The no-cloning theorem, proved in 1982, shows that an unknown quantum state cannot be copied, so quantum computers cannot protect data simply by keeping backup copies. And looking at a qubit to check it destroys its superposition. Worse, a qubit can suffer not only bit flips but also phase errors, which spoil the minus signs that interference depends on.

Quantum error correction

In 1995 Peter Shor found a way around these obstacles, and others soon improved on it. Quantum error correction spreads one logical qubit across many physical qubits and measures only parity checks that reveal errors without revealing the stored information. It is like asking whether two hidden cards match in color without asking what either card is. The most studied scheme, the surface code, arranges qubits on a grid. The crucial result is the threshold theorem: if physical error rates fall below a certain threshold, adding more qubits makes a logical qubit more reliable rather than less. Google reported in December 2024 that its Willow chip had crossed this line: each step up in grid size roughly halved the logical error rate. The cost is steep. Estimates vary, but a single reliable logical qubit may need hundreds to a thousand or more physical ones.

The NISQ era

In 2018 Preskill named the current period. His term NISQ, for noisy intermediate-scale quantum, describes today's machines: tens to a few thousand qubits, too noisy and too small for full error correction. Whether such machines can do anything useful before fault tolerance arrives is one of the field's biggest open questions. Several companies publish roadmaps aiming for useful error-corrected machines around 2029 to 2030, targets that should be treated as goals rather than forecasts.

Worked example: two machines on paper

Suppose Machine A has 1,000 qubits and its two-qubit gates succeed 99 percent of the time. Machine B has 50 qubits and its gates succeed 99.9 percent of the time. A press release would trumpet Machine A. But try running a circuit of 1,000 gates without error correction. On Machine A, with a one in a hundred chance of error at each step, the odds that every gate works are tiny, well under one in ten thousand. On Machine B, with one chance in a thousand, the odds are better than one in three. For deep circuits, Machine B gives meaningful answers and Machine A gives noise. This is why qubit count alone is a poor measure of power: error rates and how the qubits connect also limit what a machine can run. When reading about a new record, ask about gate fidelity, coherence time and, above all, the number of error-corrected logical qubits.