Fermi Estimation

How many gas stations are there in the United States? You almost certainly don't know this number, and neither does your interviewer expect you to. What they're testing is whether, given no data at all, you can produce an answer within a factor of two or three of the truth by breaking the question into pieces you can reason about.

Try it yourself before reading on: give yourself two minutes and a blank sheet of paper, no internet.

The trap: refusing to guess, or guessing wildly

Two failure modes show up constantly. The first is freezing — "I don't know how many gas stations there are, how could I possibly figure that out" — which fails the test outright; the interviewer isn't grading the number, they're grading whether you'll even attempt structured reasoning under uncertainty. The second is guessing a single top-down number with no visible reasoning ("um, maybe 50,000?") — which might accidentally be close, but demonstrates nothing, and falls apart the moment they ask "why?"

Worked example: gas stations in the United States

Break the total into a chain: (population) × (cars per person) × (gas stations needed to serve those cars).

• US population: roughly 330 million people.
• Cars per person: call it 0.8 — most adults have one, not every person is a driver (kids, some households share). That's about 260 million cars.
• Fill-ups: a typical car fills up about once a week, so roughly 260 million fill-ups a week.
• Capacity per station: a gas station has, say, 8 pumps, each serving a car in about 10 minutes, so one pump handles roughly 6 cars an hour. Over a busy 12-hour operating window that's about 70 cars a day per pump, or 560 cars a day for an 8-pump station — call it 500 for round numbers, so 3,500 cars a week per station.

Divide: $260{,}000{,}000 / 3{,}500 \approx 74{,}000$ stations.

$$N_{\text{stations}} \approx \frac{P \times c \times f}{s},$$

where $P$ is population, $c$ is cars per person, $f$ is fill-ups per car per week, and $s$ is cars served per station per week. In words: total weekly demand for fill-ups divided by weekly capacity per station gives the number of stations needed to clear that demand. The real US figure is around 145,000 — the estimate is low by roughly a factor of two, which for a cold Fermi problem with zero data is a solid result; the exercise is graded on structure, not on nailing the digit.

Worked example: piano tuners in Chicago — the canonical version

Chicago's metro population: about 9 million. Households: assume 2.5 people per household, so 3.6 million households. Fraction owning a piano: maybe 1 in 20, so 180,000 pianos — plus a smaller number in schools, churches and concert halls; ignore that as a rounding error. Tuning frequency: once a year, so 180,000 tunings a year. Time per tuning including travel: call it 2 hours. A tuner works, say, 1,500 hours a year (accounting for non-tuning time). One tuner can do $1500/2 = 750$ tunings a year. Divide: $180{,}000 / 750 = 240$ piano tuners.

$$N_{\text{tuners}} = \frac{(\text{population}) \times (\text{piano fraction}) \times (\text{tunings/year})}{(\text{tunings a tuner can do per year})}.$$

In words: total tuning-demand-hours divided by the hours a single tuner can supply gives the headcount. This is the exact classic Fermi problem, and estimates like this one consistently land within a factor of two or three of directory counts of actual tuners — which is remarkable given the input was pure structured guessing.

→ × 1 tuning/year 180,000 tunings/yr

÷ 750 tunings/tuner/yr ≈ 240 tuners

The transferable technique

Every Fermi problem is the same three moves: decompose the unknown quantity into a chain of ratios or counts you can estimate individually; anchor each link to something you actually know (population figures, typical household sizes, plausible work hours); and state your assumptions out loud as you go, because a wrong number with visible reasoning beats a right number with none. This is the identical skeleton behind market-sizing questions and behind rough-and-ready trading estimates — "how big is this market" is just a Fermi problem wearing a business suit.