Skip to main content

Fermi Estimation

TL;DR

Fermi Estimation: Break an unknown quantity into a product of estimable components, estimate each from common knowledge and reasoning, then multiply. A piano tuner estimate: city population Γ— fraction who own pianos Γ— tuning frequency Γ— time per tuning. Each estimate is rough; the product is often within an order of magnitude of the truth. Essential for quick sanity checks, market sizing, and decisions where approximate answers beat no answers.

What Is Fermi Estimation?​

Enrico Fermi (1901–1954) was an Italian-American physicist who built the first nuclear reactor and won the 1938 Nobel Prize in Physics. He was legendary among colleagues for his ability to estimate complex quantities from basic reasoning β€” "How many piano tuners are in Chicago?" was his famous pedagogical example. The technique he epitomised has since been formalised as Fermi Estimation or "back-of-the-envelope calculation."

The core insight is that many seemingly unanswerable questions can be decomposed into components that can each be estimated from knowledge and reasoning. The estimates are individually rough β€” often within a factor of 2 or 3 β€” but the errors tend to partially cancel when multiplied together, producing final estimates that are typically within an order of magnitude (factor of 10) of the actual answer. For most practical purposes, order-of-magnitude accuracy is all that's required.

Fermi Estimation is valuable precisely where it seems unnecessary: when you think you need an exact answer, you often only need to know whether the number is in the thousands, millions, or billions. Is this market worth pursuing? Is this engineering approach feasible? Does this business model generate positive unit economics? These questions can often be answered definitively with back-of-envelope estimates, saving weeks of rigorous research.

At its heart, Fermi Estimation is calibration training β€” the practice of knowing what you know and quantifying your uncertainty, rather than retreating to "I don't know" when exact data is unavailable.

How It Works​

Step 1: State the target quantity precisely
  β€” "How many iPhones are sold globally per year?"
  β€” Vague questions produce misleading estimates

Step 2: Identify knowable anchors
  β€” What do you know for certain or near-certainty?
  β€” World population: ~8 billion; US population: ~330 million

Step 3: Decompose into a product of factors
  β€” Express the unknown as: A Γ— B Γ— C Γ— D
  β€” Each factor should be independently estimable

Step 4: Estimate each factor from first principles
  β€” Use reference points, logic, and reasoning
  β€” Bound each estimate: "between X and Y, probably around Z"

Step 5: Multiply and record
  β€” Calculate the product
  β€” Note the range (pessimistic Γ— optimistic estimates)

Step 6: Sanity check
  β€” Does the answer "feel right"?
  β€” Can you check one or two factors against known data?

Three Real-World Examples​

Market Sizing: UK Coffee Shop Revenue​

Question: What is the total annual revenue of coffee shops in the UK?

Decomposition:

  • UK population: ~67 million
  • Fraction who drink coffee shop coffee at least weekly: ~25% = ~17 million people
  • Average spend per visit: ~Β£4.50
  • Average visits per week among weekly customers: ~1.5
  • Total annual spend: 17M Γ— Β£4.50 Γ— 1.5 Γ— 52 weeks = Β£5.95 billion

Actual figure (2023): approximately Β£4.5 billion. Our estimate is within 30% β€” well within order-of-magnitude accuracy. For a market entry decision, this estimate is entirely sufficient.

Engineering Feasibility: Server Capacity​

Question: Can a single server handle 10,000 simultaneous users of our web app?

Decomposition:

  • RAM per user session: ~2 MB (estimated from application data)
  • 10,000 users Γ— 2 MB = 20 GB RAM required
  • Modern server: 128 GB RAM available
  • Active requests at any moment (10% concurrency): 1,000 simultaneous requests
  • Requests per second the server can handle: ~5,000 (from benchmark data)
  • At 1,000 concurrent users making ~1 request every 2 seconds: ~500 req/sec

Conclusion: both RAM (20 GB vs 128 GB capacity) and CPU (500 req/sec vs 5,000 capacity) suggest the server can handle load with significant headroom. No detailed capacity testing needed to make the architectural decision.

The Original: Piano Tuners in Chicago​

Fermi's classic:

  • Chicago population: ~3 million
  • People per household: ~2.5 β†’ ~1.2 million households
  • Fraction with pianos: ~5% β†’ ~60,000 pianos
  • Tuning frequency: ~1 per year β†’ 60,000 tunings/year
  • Time per tuning (travel + work): ~2 hours
  • Working hours per year per tuner: ~2,000 hours
  • Tuners needed: 60,000 Γ— 2 / 2,000 = 60 tuners

Actual number circa Fermi's time: approximately 50–80. Spot on.

When to Use It​

βœ… Fermi Estimation is essential for:

  • Market sizing and TAM calculations in business planning
  • Quick feasibility checks before committing engineering resources
  • Due diligence sanity checks ("does this business model make sense?")
  • Consulting interviews and analytical problem-solving
  • Any decision where approximate answers are sufficient and exact data is unavailable

❌ Not appropriate for:

  • Decisions where precise numbers are legally or contractually required
  • Physical engineering where tolerances matter (safety margins, structural calculations)
  • Cases where known data exists and is accessible β€” look it up instead
Pairs well withWhy
First PrinciplesFermi Estimation applies first-principles reasoning to quantity problems
Expected ValueExpected value calculations require probability estimates that Fermi methods can provide
Proximate-Root CauseFermi Estimation helps quantify the magnitude of causes in root cause analysis
Bayesian ThinkingBoth involve updating beliefs from limited evidence; Fermi provides initial estimates

Common Misuses and Limitations​

False precision. The output of a Fermi estimate is an order-of-magnitude figure, not a precise number. Reporting "I estimate 47,500 piano tuners" from a Fermi analysis is misleading β€” "roughly 50,000, probably between 30,000 and 100,000" is appropriate.

Anchoring bias in factor selection. The factors you choose and how you estimate them can be heavily influenced by the first number that comes to mind. Counter this by working the problem from multiple decomposition approaches and comparing results.

Compounding systematic errors. Fermi estimates are tolerant of random errors (they partially cancel). They are not tolerant of systematic errors β€” if every factor is biased in the same direction, the product is badly wrong. Check whether your assumptions are consistently optimistic or pessimistic.

Using it to avoid looking things up. Fermi Estimation is for when data is unavailable. If relevant data exists and is accessible, retrieve it β€” don't estimate when facts are available.

ModelRelationship
First Principles ThinkingFermi Estimation is first principles applied to quantity problems
Bayesian ThinkingBoth techniques develop calibrated reasoning under uncertainty
Expected ValueEV calculations depend on probability estimates that Fermi methods can generate
Scientific MethodFermi estimates generate testable hypotheses about quantities

Frequently Asked Questions​

How accurate does a Fermi estimate need to be to be useful?

It depends on the decision. For market sizing (should we enter this market?), within an order of magnitude is usually sufficient. For unit economics (does this business model work?), within a factor of 2–3 is needed. For engineering feasibility (can this server handle the load?), within 50% may be needed. Calibrate the required accuracy to the sensitivity of the decision β€” some decisions are robust to wide uncertainty ranges; others are not.

How do you get better at Fermi Estimation?

Three practices: (1) make explicit estimates before looking up answers β€” estimate how many airports exist in the US, then check; calibration comes from confronting the gap; (2) build a library of reference quantities β€” memorise key anchors (world population, US GDP, average salaries, common speeds and distances); (3) practice decomposition β€” any time you see an impressive aggregate number, practice decomposing it from first principles to see if you can reconstruct it.

What are the most useful reference anchors for Fermi Estimation?

A short list of high-utility anchors: world population (8B), US population (330M), US GDP ($27T), median US household income ($75K), average working hours per year (~2,000), average human lifespan (~75 years), driving speed (~60mph), light speed (~300,000 km/s), number of seconds in a year (~31.5M), and the density of water (1 g/cmΒ³). These anchors appear in a surprisingly large fraction of useful Fermi estimates.

Further Reading​

  • Weinstein, L. & Adam, J. (2008). Guesstimation: Solving the World's Problems on the Back of a Cocktail Napkin
  • Mahajan, S. (2010). Street-Fighting Mathematics β€” rapid estimation techniques from MIT
  • Tetlock, P. & Gardner, D. (2015). Superforecasting β€” calibration and structured estimation under uncertainty

Apply with AI​

πŸš€ Build a Fermi estimate with MindMax β†’

This page is part of the MindMax Mental Models Knowledge Base.