Beyond the qubit count: how to read a quantum computing headline

The problem

Size is not capability

A headline that a processor has a thousand qubits tells you how big it is, not what it can do. Every operation on today's hardware has some chance of failing, and a computation is a long chain of operations. If each fails with probability \(p\), all \(N\) succeed with probability about \((1-p)^N\).

chance 100 steps run error-free at 1 error in 100

37%

chance 10,000 steps run error-free at 1 error in 1,000

1 in 22,000

the same 10,000 steps at 1 error in a million

99%

The numbers above are our calculations from that formula. Adding qubits to a machine with the first error rate does not help; it gives the errors more places to happen.

Useful algorithms such as factoring or detailed chemistry need millions to billions of operations. Microsoft's quantum team, writing with John Martinis in September 2026, puts the error rate needed at about one in ten billion for the simplest useful applications and one in a million billion for the most demanding. No physical qubit comes close. The real question behind a headline is how close the machine gets to reliable operation, and at what cost.^[1]

Question 1

What exactly is being counted?

A physical qubit is one device: an ion, an atom, a superconducting circuit. A logical qubit is information spread over many physical qubits by an error-correcting code, so that errors can be found and undone without disturbing it. The code's distance d is the number of physical errors needed to cause an undetected logical one; a surface code of distance d uses about \(2d^2\) physical qubits.

"Logical qubit" is used loosely. Some demonstrations run correction for many rounds; others prepare and measure a code once. Some correct errors; others only detect them and discard the failed runs afterwards, which looks clean in small tests but wastes a fraction of runs that grows with the computation. The comparison to look for is whether the logical qubit beats its own parts: in Google's 2024 experiment, the distance-7 logical memory outlived the best physical qubit by a factor of about 2.4.^[2]

Question 2

How good is each operation?

Two-qubit gates are usually the weakest link, because entangling operations are hardest and every algorithm needs many of them. Measurement errors, errors while qubits wait, leakage out of the two levels that form the qubit, and crosstalk between neighbours count too. A headline fidelity is often the best pair on the chip.

How the number was measured matters as much as its size. Randomized benchmarking reports an average over many inputs. Coherent errors, small systematic over-rotations, can look tiny on average yet add up in step over a long circuit, so the worst case can be far worse than the average. Long coherence times mean little on their own; what counts is how many operations fit inside them.

Question 3

Does the error fall as the code grows?

Error correction helps only below a threshold error rate. Below it, a larger code gives a more reliable logical qubit; above it, a larger code just collects more errors. The most informative single number is the suppression factor \(\Lambda\): how much the logical error falls each time the distance grows by two.

The two results below use different codes and count errors differently, per correction round or per operation, so compare their slopes, not their absolute values. The slope sets the cost. If each trend continued from distance 3, a billionfold reduction would need distance about 59 at \(\Lambda\approx2.1\) but about 21 at \(\Lambda\approx12\): roughly seven thousand physical qubits per logical qubit against nine hundred. That is our extrapolation; rare correlated errors, from cosmic rays or leakage, are what could break it.

Protecting memory is not computing. A complete demonstration also runs logical gates between logical qubits, including the costly non-Clifford gates, usually supplied as magic states, and decodes fast enough to keep up during a long run.^[2,3,4]

Question 4

Could a classical computer do the same?

Many beyond-classical demonstrations run tasks chosen because they are hard to simulate, such as sampling the outputs of random circuits. They are honest tests of hardware but say little about applications, and their classical cost is a moving target, as the revisions to the first such claim, in the table below, show.

So ask whether the comparison used the best known classical method, ideally one proposed by people outside the team, and whether the output can be checked at all. Results with a verifiable answer, or a statistical certificate of how faithfully the circuit ran, carry more weight. Composite benchmarks such as quantum volume, the largest random circuit of equal width and depth a machine runs successfully, help compare machines but do not measure usefulness.^[5,6,7,8]

Question 5

What would a useful task need?

A resource estimate turns an algorithm into hardware: logical qubits, non-Clifford gate counts, run time, and the physical qubits needed at an assumed error rate and speed. The headline number always rests on assumptions, and lower qubit counts usually rest on stronger ones.

Speed matters as much as size. On hardware whose correction cycles are a thousand times slower, a one-week computation becomes one of about two decades. John Preskill has proposed a nearer milestone: error-corrected machines running about a million operations reliably, which he calls megaquop machines, the scale at which correction starts to pay off well before factoring.^[9,10,11]

Question 6

Result, plan or product?

Programmes that fund the field increasingly pay for verified performance. DARPA's Quantum Benchmarking Initiative asks whether any approach can reach utility scale by 2033, meaning computational value above the machine's cost, and verifies hardware in its final stage. The Department of Energy's Q Competition ties payments to lab-verified demonstrations of at least 100 logical qubits. Microsoft's post argues for judging reliability, scale, capability and performance together.

Announcements mix statements of very different weight, and the wording rarely says which kind it is. Sort each claim into one of these before weighing it.^[1,12,13]

Worked example

Two real announcements, scored

Neither result below answers every question, and neither claims to. The scorecard shows which questions each one does answer. A headline that a company has unveiled a 1,000-qubit processor, by contrast, answers only the size of the machine until the same questions are put to it.^[3,4,14]

Summary

The checklist

Few announcements answer all six questions, and that is normal. The point is to know which ones a result answers and which it leaves open.