Why Quantum Computing Exists
Quantum computers are specialised accelerators, not faster general-purpose machines. A QPU pays off only when a verified quantum algorithm matches the problem's structure — and measurement returns a few classical bits per run, never the full internal state.
Why does quantum computing exist?
Developers already choose between processors. A CPU handles branching, general-purpose work. A GPU applies the same arithmetic across large arrays. An NPU accelerates a narrow set of neural-network operations.
A QPU — a quantum processing unit — is another specialist. It controls a small physical system governed by quantum mechanics, applies a programmed sequence of operations to it, and returns samples of classical bits.
These are roles, not a leaderboard. Moving work to any accelerator pays off only when the problem's structure, the algorithm, and the data-movement costs all fit. The same is true for a QPU — only more strictly.
What does a quantum computer offer that a classical one doesn't?
The state of n classical bits is a single string of n zeros and ones. The state of n qubits — the quantum counterparts of bits — is described by a list of amplitudes: one number (possibly negative or complex) for every possible bit string, and there are 2^n of those strings.
Work the numbers. For 10 qubits that is 2^10 = 1,024 amplitudes. Stored as pairs of 8-byte floating-point numbers — two numbers per amplitude, because a complex number has two parts — that is 1,024 × 16 bytes = 16,384 bytes — about 16 kB. For 50 qubits it is 2^50 ≈ 1.13 × 10^15 amplitudes — roughly 18 petabytes. No classical memory tracks that directly.
Here is the catch: that enormous description is not readable output. Reading a quantum computer means measurement, and measuring n qubits returns just n classical bits per run.
What does one run actually return?
Take 3 qubits. Their state carries 2^3 = 8 amplitudes, one for each string from 000 to 111. A single run returns exactly one of those strings.
The probability of each string is the squared size of its amplitude. If the amplitude of 101 is 0.5, then 101 appears on 0.5 × 0.5 = 0.25 — 25% — of runs. The other 75% of runs return other strings.
A useful quantum algorithm therefore has to arrange the amplitudes so that the strings you care about come out with high probability. It cannot print all 8 values, let alone all 2^50.
Run one genuine quantum operation
What just happened?
The circuit — the programmed sequence of operations you just ran, read left to right — starts one qubit in the state written |0⟩ (read "ket zero" — a state guaranteed to measure as 0). The H gate — the Hadamard gate — moves it to a state whose two amplitudes are both 1/√2 ≈ 0.7071.
Each amplitude squared gives 0.7071 × 0.7071 ≈ 0.5, so each shot — one complete run of the circuit ending in one measurement — reads 0 or 1 with 50% probability. In 100 shots you might see 47 and 53, or 55 and 45.
Notice the workflow: a classical program defined the circuit, a simulator (or a real QPU) executed it, and classical counts came back. That hybrid loop never goes away — real quantum processors always sit inside a classical system. Try variations yourself in the Lab.
Also notice what this is not: a coin flip built this way beats no classical random-number generator. Correct quantum behaviour and useful advantage are separate claims.
Doesn't it try all answers at once?
A popular story says a quantum computer "tries every answer in parallel and picks the best one". The first half is misleading and the second half is false.
A state can hold non-zero amplitudes for many strings at once, but measurement samples one string, at random, weighted by squared amplitudes. Nothing reads out the full list.
Real quantum algorithms work by interference: they choreograph the amplitudes so that wrong answers cancel and right answers reinforce, before measuring. That works only for problems with the right structure — which is why quantum speedups are specific, not universal.
How does this look on real hardware?
Today's quantum processors are limited by gate errors (each operation goes slightly wrong a small fraction of the time), restricted qubit connectivity (each qubit can interact directly with only a few neighbours), calibration drift, and small circuit sizes. An algorithm can be asymptotically faster in theory while today's devices still lose to a laptop on the full task, end to end.
Keep three levels of evidence separate: proved theory, simulated or small-scale demonstration, and practical capability on current hardware. Preskill's term NISQ (noisy intermediate-scale quantum) names the current era. Browse real devices and their measured error rates at /hardware/qpus, and compare them side by side at /hardware/compare.