In short: A qubit is described by continuous complex amplitudes, but a measurement returns only one classical bit and destroys the rest, so n qubits hold 2^n amplitudes yet give you at most n bits of readable information.
Two ways of holding information
A classical bit in a CMOS flip-flop is a node held near 0 V or near VDD. Noise margins make it robust: a node sitting at 0.1 V and one at 0.05 V are both read as logic 0, and the next clock edge restores the full level. Copying is trivial too. Fan-out a net to ten gates and every receiver sees the same value. Your STA signoff in Synopsys PrimeTime or Cadence Tempus relies on exactly this property: a value is a value, it can be observed without being changed, and it can be duplicated.
A qubit is a two-level quantum system (a superconducting transmon, a trapped ion, a spin) whose state is written as |ψ⟩ = α|0⟩ + β|1⟩, where α and β are complex numbers with |α|² + |β|² = 1. Ignoring global phase, two real numbers (the polar angle θ and the phase φ on the Bloch sphere) fix the state. If you read our primer What is a Qubit? this will be familiar. The tempting conclusion, that a qubit therefore stores "infinite information" because θ and φ are continuous, is where most interview answers go wrong.
Measurement collapse: you get one bit, not two numbers
Measuring in the computational basis returns 0 with probability |α|² and 1 with probability |β|² (the Born rule), and afterwards the qubit is left in |0⟩ or |1⟩ respectively. The numbers α and β are not retrievable from one shot. To estimate them you must prepare the identical state many times and tally outcomes; resolving a probability to ±1% needs roughly 10,000 shots by the standard 1/√N error scaling. This is why quantum computers always report histograms from thousands of shots.
Key fact: Holevo's theorem (1973) proves that n qubits can convey at most n classical bits to a receiver, no matter how cleverly the receiver measures.
So the continuous angles are real physics, and they steer interference inside the computation, but they are not an output channel. The information capacity "as seen from outside" is one bit per qubit. Entanglement does let two parties exploit a shared pair to send two classical bits using one transmitted qubit (superdense coding), but that pair already consumed an entangled partner; the Holevo bound is not violated.
Why n qubits do not store 2^n usable bits
An n-qubit register is described by 2^n complex amplitudes. For 50 qubits that is about 1.1 × 10^15 amplitudes; at 16 bytes each (complex128) a classical simulator would need about 16 PiB just to hold the state vector. This exponential bookkeeping is why simulating around 50 qubits exactly is a supercomputer problem.
But describing a state costs 2^n numbers, while extracting from it costs n bits per shot. Consider a quick example: put 3 qubits through Hadamards and you hold an equal superposition of 8 basis states, each amplitude 1/√8 with probability 1/8. Measure and you obtain one 3-bit string, uniformly random. You did not "read all 8 values in parallel". Quantum algorithms are useful only when they choreograph amplitudes so that wrong answers cancel and the right one is reinforced before measurement.
The no-cloning theorem
Classical logic copies by fan-out. Quantum mechanics forbids the equivalent: there is no unitary operation that maps |ψ⟩|0⟩ to |ψ⟩|ψ⟩ for every unknown |ψ⟩ (Wootters, Zurek and Dieks, 1982). A short proof: unitary maps preserve inner products. If U copied both |0⟩ and |+⟩, the inner product before is ⟨0|+⟩ = 1/√2, while after copying it would be (1/√2)² = 1/2. These cannot match, so no such U exists.
Key fact: You can copy a known basis state (a CNOT does that), but you cannot copy an unknown arbitrary state. CNOT on (|0⟩+|1⟩)/√2 produces an entangled Bell pair, not two copies.
The consequences are practical. Classical error correction repeats a bit three times and takes a majority vote. A qubit cannot be repeated, so quantum error correction instead spreads one logical qubit across many physical ones by entanglement and extracts only error syndromes (parity checks) without learning the data. The surface code needs on the order of hundreds to thousands of physical qubits per logical qubit at today's error rates. It also means a quantum "fan-out buffer" does not exist, and wiring a quantum chip is a very different routing problem from the place-and-route you run in Innovus or IC Compiler II. On the other hand, no-cloning is the foundation of quantum key distribution: an eavesdropper cannot copy the photon without disturbing it.
Comparison at a glance
| Property | Classical bit | Qubit |
|---|---|---|
| State description | 1 binary value | 2 complex amplitudes (2 real parameters after normalisation and phase) |
| Readout per element | 1 bit, non-destructive | 1 bit, destructive (collapse) |
| Copying | Fan-out, free | Forbidden for unknown states |
| n elements: description size | n bits | 2^n amplitudes |
| n elements: extractable data | n bits | at most n bits (Holevo) |
| Noise | Restored by gate gain | Decoherence; needs error correction |
Seeing the collapse
Interview traps and the hardware angle
- "A qubit is 0 and 1 at the same time." Better: it is a coherent amplitude pair, and probabilities arise only at measurement.
- "Quantum computers try all answers at once." They do not; interference must concentrate amplitude on correct answers. Grover search yields a quadratic speedup (about √N queries), not an exponential one.
- "Superposition means you can store a million bits in 20 qubits." Holevo: 20 bits out, at most.
For silicon engineers the story continues at the control layer. Superconducting qubits sit at roughly 10-20 mK, and research groups design cryo-CMOS controllers in standard EDA flows (Cadence Virtuoso and Spectre for analog, Synopsys tools for digital) because bias lines per qubit do not scale. Teams at Avecas VLSI design services see growing demand for exactly this mixed-signal skill set. Follow-up reading: Quantum Entanglement Basics.
Want to go from concepts to hardware-level fluency? Explore the quantum and VLSI tracks in our CourseTron courses, or pair them with hands-on lab training at ChipXpert.
FAQ
Can a qubit store more information than a bit?
Not in extractable terms. Its state needs continuous parameters to describe, but a measurement yields one bit, and Holevo's bound caps n qubits at n retrievable classical bits.
Why does measurement destroy a qubit's state?
Measurement projects the state onto a basis outcome. Afterwards the qubit sits in |0⟩ or |1⟩, and the original amplitudes and relative phase are lost.
Why can't we just copy a qubit before measuring it?
The no-cloning theorem shows that no unitary can duplicate an unknown state, because copying would not preserve inner products.
Does 2^n amplitudes mean exponential speedup?
No. Only specific algorithms, such as Shor's factoring, exploit interference to turn that structure into a speedup; many problems get at most a quadratic gain.