thinking-fermi-estimation

Fermi Estimation

Overview

Fermi estimation, named after physicist Enrico Fermi, is the art of making reasonable estimates for quantities that seem impossible to know without direct measurement. By decomposing a question into factors you can estimate, then multiplying, you often get surprisingly accurate order-of-magnitude results.

Core Principle: Break the unknown into known (or estimable) pieces. Even rough estimates combine to reasonable accuracy due to errors canceling out.

Trigger Card

When you need a number you can't measure or look up, and order-of-magnitude is enough:

1. Decompose: Quantity = Factor₁ × Factor₂ × ... × Factorₙ (by component, by rate×time, or by population×fraction).
2. Estimate each factor with a range, not a point; use geometric mean for order-of-magnitude; round to one significant figure.
3. Multiply and sanity-check: Does the order of magnitude make sense? Would a 10× error change the decision?
4. Report: "~X, within 3-5×" — never false precision.

If the number is lookup-able (COUNT(*), pricing page, profile it), do that instead — a made-up estimate where a real value exists is strictly worse.

When to Use

• Capacity planning ("How much storage will we need?")
• Cost estimation ("What will this infrastructure cost?")
• Market sizing ("How many potential users exist?")
• Feasibility assessment ("Is this even plausible?")
• Sanity checking ("Does this number make sense?")
• Interview questions ("How many piano tuners in Chicago?")
• Quick prioritization ("Is this worth pursuing?")

Decision flow:

Need a number you don't have? → Can you measure OR look it up cheaply? → yes → DO THAT (don't Fermi it)
                                                                       ↘ no → Will an order-of-magnitude answer suffice? → yes → FERMI ESTIMATE
                                                                                                                          ↘ no → you need real data, go get it

When NOT to Use

• The number is cheaply lookup-able or measurable. Don't Fermi the size of a table you can COUNT(*), the latency you can profile, the file size you can ls, the price on a pricing page, or any fact one query/search away. A made-up estimate where a real value is available is strictly worse — it adds error and false confidence. Look it up.
• You need precision, not order of magnitude. Fermi gives you "~4 TB, within 3-5x." If the decision turns on 3.8 vs 4.2, estimation can't carry it — get the real number.
• The factors are correlated or you'd be inventing the base rates. If the decomposition is just stacked guesses with no anchor, the result is theater. Either find one real anchor or flag the whole thing as a rough guess.
• The decision is identical across the plausible range. If "somewhere between 1 GB and 1 TB" doesn't change what you do, skip the estimate and act.

The Fermi Process

Step 1: Clarify What You're Estimating

Be precise about the quantity:

Vague: "How big is the market?"
Precise: "How many SaaS companies with 50-500 employees in the US
         would pay $1000/month for our product?"

Step 2: Decompose into Estimable Factors

Break into pieces you can estimate:

Storage needs for user data:
= (Number of users)
  × (Data per user per day)
  × (Days of retention)
  × (Overhead factor)

Decomposition strategies:

Strategy	Example
By component	Total = Sum of parts
By rate × time	Total = Rate × Duration
By population × fraction	Target = Base × Percentage
By analogy × adjustment	New ≈ Similar × Ratio

Step 3: Estimate Each Factor

For each factor, estimate based on:

Source	Example
Known data	"We have 10,000 DAU"
Industry benchmarks	"Average SaaS churn is 5%"
Physical constraints	"A human can make ~50 decisions/day"
Logical bounds	"At least 1, at most 1 million"
Personal experience	"I've seen systems handle 1000 req/s"

When estimating:

• Use ranges, not point estimates: "10,000 to 50,000"
• Prefer geometric mean for order-of-magnitude: √(10,000 × 50,000) = 22,360
• Round to one significant figure: ~20,000

Step 4: Combine Factors

Multiply (or add) factors together:

Storage = 50,000 users × 10 KB/user/day × 365 days × 1.5 overhead
        = 50,000 × 10,000 × 365 × 1.5 bytes
        = 274 billion bytes
        ≈ 270 GB/year

Step 5: Sanity Check

Verify reasonableness:

• Does the order of magnitude make sense?
• Is it physically possible?
• Does it match any known data points?
• Would a 10x error change the decision?
• Is any factor actually lookup-able? If so, replace the estimate with the real value — every measured factor you substitute shrinks the error bars.

Step 6: State Confidence and Implications

Estimate: ~270 GB/year
Confidence: Within 3-5x (80-1,500 GB)
Implication: Standard database tier sufficient; no special infrastructure needed

Fermi Estimation Template

# Fermi Estimate: [Question]

## Question (Precise)
[Exactly what we're estimating]

## Decomposition
[Quantity] = [Factor 1] × [Factor 2] × ... × [Factor N]

## Factor Estimates

### Factor 1: [Name]
- Estimate: [Value]
- Source/Reasoning: [Why this number]
- Confidence: High / Medium / Low

### Factor 2: [Name]
- Estimate: [Value]
- Source/Reasoning: [Why this number]
- Confidence: High / Medium / Low

[Continue for all factors...]

## Calculation
[Show the math]

## Result
- Point estimate: [Value]
- Range: [Low] to [High] (representing Xx uncertainty)

## Sanity Check
- Physical plausibility: [Check]
- Comparison to known data: [Check]
- Order of magnitude reasonable: [Check]

## Implications
[What does this estimate mean for the decision?]

Example 1: Data Storage Needs

Question: How much storage will our new feature need in Year 1?

## Decomposition
Storage = Users × Events/User/Day × Event Size × Days × Replication

## Factor Estimates

### Users (DAU)
- Estimate: 100,000 (current) growing to 200,000 (end of year)
- Average over year: ~150,000
- Confidence: High (we have current data)

### Events per User per Day
- Estimate: 50 events (based on current feature usage patterns)
- Confidence: Medium (new feature might differ)

### Event Size
- Estimate: 500 bytes (JSON with typical payload)
- Confidence: High (we can measure similar events)

### Days in Year
- Estimate: 365
- Confidence: Certain

### Replication Factor
- Estimate: 3x (standard for durability)
- Confidence: High (architectural requirement)

## Calculation
Storage = 150,000 × 50 × 500 × 365 × 3
        = 150,000 × 50 × 500 × 365 × 3
        = 4.1 × 10^12 bytes
        = 4.1 TB

## Result
- Point estimate: ~4 TB
- Range: 1 TB (pessimistic assumptions) to 15 TB (growth beats expectations)

## Sanity Check
- 4 TB for 150K users = ~27 MB/user/year = reasonable
- Similar feature at other company uses "several TB" = consistent
- Standard database can handle 4 TB = feasible

## Implications
- Standard managed database tier sufficient
- No need for sharding or special storage architecture in Year 1
- Budget ~$500/month for storage costs

Example 2: API Rate Capacity

Question: Can our API handle Black Friday traffic?

## Decomposition
Required RPS = Peak Daily Users × Requests/User/Session × Sessions/Day × Peak Multiplier / Seconds in Peak Hour

## Factor Estimates

### Peak Daily Users
- Estimate: 500,000 (3x normal 170K)
- Source: Last year's Black Friday
- Confidence: Medium

### Requests per Session
- Estimate: 30 API calls (measured)
- Confidence: High

### Sessions per Day
- Estimate: 2 (mobile + desktop)
- Confidence: Medium

### Peak Multiplier
- Estimate: 5x (traffic concentrated in 4-hour window, spiky within that)
- Confidence: Medium

### Seconds in Peak Hour
- Estimate: 3,600
- Confidence: Certain

## Calculation
Required RPS = (500,000 × 30 × 2 × 5) / 3,600
             = 150,000,000 / 3,600
             = 41,667 RPS
             ≈ 40,000 RPS peak

## Result
- Point estimate: 40,000 RPS
- Range: 15,000 to 100,000 RPS

## Sanity Check
- Current capacity: 10,000 RPS
- Gap: 4x capacity needed
- Similar scale companies report 20-50K RPS on peak days = consistent

## Implications
- Need 4x capacity increase
- Auto-scaling must handle 40K+ RPS
- Load test to 60K RPS (1.5x safety margin)

Example 3: Market Size

Question: How many potential customers for our developer tool?

## Decomposition
TAM = Software Companies × Avg Developers × Adoption Rate × Price Tolerance

## Factor Estimates

### Software Companies (US)
- Estimate: ~500,000 (SBA data: tech companies)
- Confidence: Medium

### With 10+ Developers (our target)
- Estimate: 10% = 50,000 companies
- Confidence: Low (rough estimate)

### Developers per Target Company
- Estimate: 30 average
- Confidence: Medium

### Adoption Rate (would consider)
- Estimate: 20% (dev tools are crowded)
- Confidence: Low

### Price Point
- Estimate: $50/developer/month
- Confidence: Medium (based on similar tools)

## Calculation
Addressable Users = 50,000 × 30 × 20% = 300,000 developers
Revenue = 300,000 × $50 × 12 = $180M/year TAM

## Result
- TAM: ~$180M/year
- Realistic serviceable market: 5-10% = $10-20M/year

## Sanity Check
- Similar dev tools (Datadog, etc.) have $100M+ revenue = plausible ceiling
- 300K potential users in a niche = reasonable

## Implications
- Market size justifies investment if we can capture 5%+
- Need differentiation in crowded space

Common Decomposition Patterns

Capacity Planning

Needed = Users × Usage/User × Factor/Usage × Growth × Safety

Cost Estimation

Cost = Resources × Unit Cost × Duration × Overhead

Time Estimation

Time = Tasks × Time/Task × (1 + Risk Factor)

Market Sizing

Market = Population × Segment% × Adoption% × Price × Frequency

Tips for Better Estimates

Use Multiple Approaches

Estimate the same thing different ways:

Website traffic estimate:
Method 1: Bottom-up from user base
Method 2: Top-down from market share
Method 3: Analogy to similar company

If methods agree within 3x, confidence increases
If they diverge wildly, investigate assumptions

Bound First

Start with upper and lower bounds:

"Definitely more than 1,000, definitely less than 10 million"
"So somewhere in 10,000-1,000,000 range"
"Let me narrow from there..."

Watch for Correlated Errors

If factors are correlated, errors don't cancel:

BAD: Users × Revenue/User (both depend on same growth assumption)
BETTER: Estimate revenue directly, or use independent factors

One Significant Figure

Don't false precision:

Calculation: 47,832,519 bytes
Report: ~50 MB (not "47.8 MB")

Verification Checklist

• Question stated precisely
• Decomposed into 3-6 estimable factors
• Each factor has reasoning/source
• Factors are relatively independent
• Calculation shown and checked
• Result sanity-checked against reality
• Uncertainty range stated
• Implications for decision clarified

Key Questions

• "Can I break this into smaller, estimable pieces?"
• "What do I already know that constrains this?"
• "What's the upper bound? Lower bound?"
• "Does this number pass the smell test?"
• "Would being off by 10x change my decision?"
• "Can I estimate this a different way to cross-check?"

Fermi's Wisdom

When asked how many piano tuners were in Chicago, Fermi didn't look it up—he estimated from population, households, pianos, tuning frequency, and tuner capacity. His estimate was reportedly within 20% of the actual number.

The lesson: You know more than you think. Decompose, estimate, combine. The errors often cancel, and you get surprisingly close to truth.

"Never make a calculation until you know the answer." — John Wheeler (Fermi's colleague)

Meaning: Estimate first to know what answer to expect, then calculate to verify.