Fermi Calculation — Method and Practice — Bench

What Makes a Fermi Question

A Fermi question asks for a quantity you can’t look up in the moment — or that doesn’t exist yet. The point isn’t precision. It’s the ability to reason from first principles to a number that’s useful: close enough to make a decision, check a claim, or expose a wrong assumption.

Enrico Fermi could estimate the yield of a nuclear blast by dropping scraps of paper and watching how far they scattered. The method: observe what you can, reason to what you can’t.

The Process

1. Name the unknown clearly

Before estimating, state exactly what you’re finding. Ambiguity compounds into large errors.

“How many cars are in India?” — is this registered cars, cars on the road today, cars sold per year? Pick one and commit.

2. Find a decomposition

Break the unknown into factors you can estimate independently. There’s usually more than one path; pick the one where your uncertainty is lowest.

For “cars in India”:

Supply path: cars manufactured per year × average car lifespan
Demand path: households that own a car × total households
Density path: cars per km² of urban area × urban area

Run the one you’re most confident in. If you have time, run two and compare — agreement between paths is a strong signal.

3. Estimate each factor

For each factor, ask: what do I actually know? What’s a lower bound? An upper bound? The geometric mean of the bounds is a reasonable point estimate.

Don’t reach for precision you don’t have. “About 10,000” is better than “9,400” when you’re estimating from first principles — the false precision obscures real uncertainty.

4. Combine and read the magnitude

Multiply out. Work in powers of 10. The answer is a magnitude first, a number second.

5. Sanity check

Does the answer make sense against things you know? If the answer implies every Indian household owns 3 cars, the assumption is broken. Find which factor is wrong and adjust.

Worked Examples

How many litres of tea are consumed in India per day?

Decomposition:

India population: 1.4 × 10⁹
Fraction who drink tea: ~70% ≈ 10⁹ tea drinkers
Cups per day per person: ~2
Volume per cup: ~150 mL = 1.5 × 10⁻¹ L

10⁹ × 2 × 1.5×10⁻¹ = 3 × 10⁸ litres/day

~300 million litres per day. India is the world’s largest tea consumer — that magnitude checks out.

How many commercial flights are in the air over Earth right now?

Decomposition:

Flights per day globally: ~100,000 = 10⁵
Average flight duration: ~3 hours = 3/24 of a day ≈ 1/8

10⁵ × (1/8) ≈ 1.2 × 10⁴

~10,000–15,000 flights airborne at any moment. Pre-pandemic data puts the real number around 9,000–12,000. Good.

How much does the internet weigh? (In electrons)

A playful one — but it has a real answer.

Data stored on internet: ~5 × 10²¹ bytes ≈ 4 × 10²² bits
Each bit is stored as a charge difference: ~10⁵ electrons per bit (rough estimate for flash storage)
Electron mass: 9.1 × 10⁻³¹ kg

4×10²² bits × 10⁵ electrons/bit × 9.1×10⁻³¹ kg = ~4×10⁻³ grams

A few milligrams. The internet, in terms of electron mass, weighs about as much as a grain of sand.

Common Failure Modes

Anchoring on a bad number. If your first estimate is wildly off, all subsequent steps inherit the error. State your assumptions explicitly so you can audit them.

Overconfidence in the chain. Five uncertain factors multiplied together can look precise. Remember: each factor carries its own uncertainty. If you’re off by 3× on three factors, the product could be off by 27×.

Wrong decomposition. Sometimes an estimate is hard because the chosen path requires knowing things you don’t. Switch paths — the same quantity can be approached from supply, demand, density, rate, or duration. One path is almost always easier.

Confusing the answer with reality. The estimate is a tool for reasoning, not a fact. Hold it loosely; update it when you get real data.

The Underlying Skill

Fermi estimation is calibration practice — training the gap between what you think you know and what’s actually true to stay small. The numbers matter less than the habit: state assumptions, decompose, estimate, check. Do it enough and the instinct transfers — you start noticing when claims are off by an order of magnitude, which is most of the interesting action.