Six ways to build a qubit,
and nobody has won yet.
Classical computing settled on silicon transistors decades ago and never looked back. Quantum computing has not settled on anything. Six physical substrates are being pursued seriously, by companies with real money and real machines, and they disagree about everything: what a qubit is made of, how cold it has to be, and what the hard problem actually is. This page is the map, with the trade-offs stated plainly.
The whole field on one screen
Numbers here are order-of-magnitude, not benchmark results, and they describe the best published devices rather than what you can rent. Scroll sideways on a narrow screen.
How each one actually works
So who is winning?
Nobody, and the question is usually asked with the wrong metric attached. Qubit count is the number that reaches the press and the least informative one available. Two thousand noisy qubits compute less than fifty good ones, because what you can run is set by how many operations you get before the errors swamp the signal.
Ask these instead
- What is the two-qubit gate error, measured across the whole device rather than on the best pair?
- How many physical qubits does one logical qubit cost at that error rate?
- How long is a single round of error correction? Clock speed compounds over billions of cycles.
- Did the logical qubit actually go below threshold, meaning more physical qubits made it better rather than worse?
- What is the strongest classical baseline for the task that was demonstrated?
The scoreboard, honestly
Superconducting leads on error correction demonstrated end to end, and on raw clock speed. Trapped ions lead on fidelity per gate by a clear margin. Neutral atoms lead on how fast they are improving, and have moved the field's expectations more than anyone in the last three years. Photonics leads on manufacturability and has the least built. Silicon spins lead on theoretical density with the fewest working qubits. Topological leads on nothing yet, and would leapfrog the entire field if the physics holds up.
If you want one sentence: superconducting and trapped ions are ahead today, neutral atoms are closing fastest, and the two manufacturing bets are playing a longer game.
The likelier outcome is not one winner
Error-corrected machines are heading toward being modular: many small processors linked together rather than one enormous chip. The links are photonic whatever the modules are made of, which means the photonics companies may end up supplying the interconnect for everyone else even if they never win the processor race outright. It is entirely plausible that two platforms coexist the way CPUs and GPUs do, one fast and one accurate.
On timing: every serious roadmap in the industry points at the end of this decade for a first genuinely useful fault-tolerant machine. Those roadmaps have slipped before, and the ones that have not slipped yet are mostly the ones that are newest. Treat any specific year, including the ones the companies publish themselves, as a plan rather than a forecast.
Quantum annealers
D-Wave builds superconducting machines with thousands of qubits, which sounds like it should top every table here. It is a different category: an annealer minimises an Ising energy landscape and cannot run arbitrary circuits, so it is not comparable to a gate-based processor and its qubit count is not the same quantity. The approach is covered in the algorithms section instead.
Where the claims come from
Almost every hardware headline traces back to a paper, and the paper is usually milder than the headline. When a record is announced, find the preprint, check what was held fixed, and check whether the figure quoted is a median or a best case. The papers page lists the results that set the baselines these announcements are measured against.