Ask a classical computer to add two numbers and it gives you one answer. Ask a quantum computer to run a circuit and it gives you a single bitstring drawn from a probability distribution. Run it again and you may get something different. This is not a bug. It is the nature of measurement in quantum mechanics, and it forces a habit that surprises newcomers: a quantum program is almost never run once. It is run hundreds, thousands, sometimes millions of times. Each run is called a shot, and the number of shots quietly governs how long a quantum job takes and whether it is even worth attempting.
Why one run is not enough
Inside the machine, qubits live in superpositions of many states at once. The whole point of a good quantum algorithm is to arrange interference so that the right answers become more likely and the wrong ones cancel out. But you never see the superposition directly. The instant you measure, each qubit collapses to a zero or a one, and you get one sample from the underlying distribution.
If the answer you want is encoded as the probability of measuring a particular pattern, a single sample tells you almost nothing. It is like trying to learn whether a coin is fair from one flip. You need a histogram, and histograms need data. So you reset the qubits, run the identical circuit again, measure again, and repeat. After enough shots you can estimate the probabilities well enough to read off the result.
The arithmetic of repetition
The catch is that statistical accuracy improves slowly. To halve your error bar, you generally need to quadruple the number of shots. That square-root scaling is forgiving for rough answers and brutal for precise ones. An expectation value that you want accurate to three decimal places can demand on the order of a million repetitions of the same circuit.
This matters because each shot has real cost. On superconducting hardware a single shot might take tens or hundreds of microseconds, including the time to prepare the qubits, apply the gates, read out the result, and reset everything for the next round. Trapped-ion systems, which are accurate but slower, can take far longer per shot. Multiply a modest per-shot time by a million and a job that sounds instant becomes minutes or hours of dedicated machine time. On a cloud service where you are billed by the second, the shot count is also the bill.
Where the tax bites hardest
The repetition problem becomes severe in exactly the algorithms people are most excited about. Variational methods, the workhorses of near-term chemistry and optimization experiments, estimate the energy of a molecule by measuring many circuits and averaging. Each step of the optimization needs fresh measurements, and a single chemistry problem can require splitting the measurement across hundreds of separate circuit configurations, each demanding its own pile of shots. The shot budget can balloon into the billions, which is one of the quieter reasons these demonstrations have stayed small.
Noise makes it worse. Every gate error and readout mistake adds spread to the distribution you are trying to measure, so noisy hardware needs even more shots to see the signal through the scatter. Error mitigation techniques, which try to scrub away noise after the fact, often work by running extra variants of the circuit, multiplying the shot count further. You pay the sampling tax twice.
Ways to lower the bill
Researchers attack the problem from several directions. Better measurement strategies group commuting observables so that a single batch of shots extracts more information at once. Techniques like classical shadows squeeze many properties out of a smaller random set of measurements. Faster qubit reset and parallel readout shave the per-shot time. And some algorithms, including phase estimation built on top of error-corrected qubits, are designed to need far fewer repetitions because they read out answers more directly rather than estimating fuzzy probabilities.
That last point hints at the long game. Fault-tolerant machines change the calculus. Once logical qubits are reliable, certain algorithms return crisp answers that need only modest sampling, and the wasteful repetition of today's noisy experiments fades. Until then, the shot count remains one of the most honest measures of what a quantum computer is actually doing. When a company reports a flashy result, it is worth asking how many times the machine had to run to get there. The answer often reveals more about the state of the field than the qubit count on the spec sheet.
