In a prior post Fermi Problems are Quadrature Problems we worked through the classic Fermi problem "how many chocolate bars are eaten each year in the United States?" twice - once as a naive one-point estimate (~40 billion bars/year) and once as a refined 4×4 age-by-season quadrature that preserved the age/season interaction (~44 billion bars/year). Both estimates hovered in the same order of magnitude, which was satisfying but also unfalsified: we worked from first principles and back-of-the-envelope estimates. No published, macroeconomic data from the real world went into the estimates.
This post is the falsifiability check. It turned into a blog post because it turned out to be much more interesting than "look up two or three published numbers and back out the real answer" because "real" is hard to pin down, there isn't actually an "answer," and the rabbit hole keeps going deeper, because building out bounds for the original Fermi problem is itself a Fermi problem. And the numbers that go into those Fermi problems pack in many assumptions that can themselves be unpacked into Fermi problems.
It's Fermi problems all the way down.
The Aggregator Trap
The obvious first move is to do a web search for "how much chocolate do Americans eat per year" and take the top result. Do that (like a rube) and you get some variation of these two claims, repeated across dozens of sites:
- Americans eat 2.8 billion pounds of chocolate per year (~11 lb/person).
- The average American eats about 3 chocolate bars per week.
Both numbers convert cleanly into "bars per year." Claim 1, at a standard 1.55 oz Hershey's bar (~10 bars/lb), gives 28 billion bars/year. Claim 2, over 52 weeks × 340M people, gives 53 billion bars/year. That would be a tidy [28B, 53B] corridor that both of our estimates sit inside. Post over. Ship it. Except that it is total bullsh-t.
Sites using this claim are a giant Gordian knot of self-referential listicle sites, paid aggregators, and misspelled/garbage URLs grifting as "legitimate" sources.
Quality score: 1/10.
The Actual Primaries
Two sources survive scrutiny:
USDA Economic Research Service publishes cocoa bean import volumes as part of its agricultural trade tracking. Between 2000 and 2022, the US imported an average of ~425,000 metric tons of cocoa beans per year. (2023 and 2024 were unusual crop-failure years and dropped to 269kt and 198kt respectively; we'll use the long-run average since the recent dip is a supply shock, not a demand signal.) This is a real customs-derived number with a known methodology.
(Side note: reporing an average consumption per year over TWENTY YEARS reeks of Fermi sub-problems.)
National Confectioners Association's State of Treating 2025/2026 report, compiled using Circana retail-panel data and Euromonitor market modeling, reports US chocolate sales of $28.4 billion in 2025 (51.7% of the $55B total confectionery market). This is a real industry number with named data providers.
To turn either one into "bars per year," we have to introduce multiple layers of estimates (models). Which is to say, we have to do a Fermi problem.
Lower-Bound Fermi Problem: Cocoa Tonnage to Bars
Starting from 425,000 metric tons/year of imported cocoa beans, the chain of unit conversions goes something like:
- Bean → cocoa mass. After shell removal, roasting, and winnowing, yield is roughly 80% of raw bean mass. ~340,000 t of usable cocoa mass.
- Cocoa mass → finished chocolate. Finished chocolate is a blend. Milk chocolate is ~10–20% cocoa content by mass; dark is 50–85%; white is 0% solids. Weighted by US market mix (heavily milk), call the average cocoa content ~20% → 340,000 / 0.20 → 1.7 million metric tons of finished chocolate → 3.75 billion pounds.
- Non-bar uses. Some cocoa becomes baking cocoa, hot chocolate mix, cosmetic cocoa butter, ice cream inclusions - not bars. Call it 15–25% of cocoa mass. Subtract ~20%.
- Net trade in finished chocolate. The US is a net importer of finished chocolate (Toblerone, Lindt, Cadbury via Canada, etc.), which adds back roughly the same order of magnitude as the non-bar leakage. Call these a wash to first order.
- Pounds → bars. At 1.55 oz/bar → 10.3 bars/lb → 3.75 × 1e9 lb × 10.3 → ~39 billion bars/year.
Every step in that five-step decomposition carries real uncertainty. Bean yield could be 75% or 85% (±6%). Average cocoa content could be 15% or 25% (±25%). Non-bar share could be 10% or 30% (±33%). Net trade is a wash only to first order. Individually these are tens-of-percent knobs, not factor-of-two ones - but they compose multiplicatively, and five stacked ±25%-ish knobs is already enough to smear the answer by roughly ±2×. And every arrow in the chain is a Fermi sub-problem whose inputs are numbers someone else estimated the same way.
If we swing the knobs pessimistically, we land at ~25B bars. Optimistically, ~55B. The "lower bound from real primary data" is a range of ~25–55B, which is not really a bound at all - it's a Fermi estimate wearing a lab coat.
Upper-Bound Fermi Problem: Dollar Sales → Bars
Starting from $28.4B/year in US chocolate sales, the chain is shorter but each step is fuzzier:
- Dollar sales → bar-equivalents. We need an average retail price per "bar." Full-size checkout bar: ~$1.50–$2.00. Premium bar (Lindt, Ghirardelli): $3–$5. Fun-size piece from a Halloween bag: ~$0.15–$0.30. Bulk seasonal chocolate (Easter eggs, Valentine's boxes, advent calendars): wildly varying $/oz, but each piece often counts as an eating event.
- Weighted average. If bar-form full-size chocolate dominates sales dollars, the average is close to $1.50/bar. If fun-size and bulk dominate the count, the average is closer to $0.50/piece.
That's not a factor of two; it's a factor of three, and the whole answer sits on it:
| Average $/bar | Implied bars/year |
|---|---|
| $0.50 | ~57 billion |
| $1.00 | ~28 billion |
| $1.50 | ~19 billion |
The "upper bound from real primary data" is anywhere from 19B to 57B, depending entirely on how we resolve one knob - which is itself the same "what counts as a bar" linguistic knob we called out in the very first section of the prior post. Primary data didn't retire the knob. It just relocated it into the unit-conversion step.
Fermi Problems All the Way Down
Here's the pattern. We started with a Fermi problem (bars/year). To bound it, we went hunting for real data. The real data we found isn't denominated in the units we want, so converting it into bars/year required building another Fermi problem - a decomposition into estimated factors each guessed to within a factor of two. And when we look at where those factors come from, they are the outputs of other Fermi problems that somebody else already solved:
- USDA's 425,000 metric tons of cocoa imports is itself the output of a chain: customs declarations aggregated across ports, unit conversions, imputations for missing data, seasonal adjustments. USDA analysts did their own Fermi decomposition to produce that number; we just don't see the pieces.
- NCA's $28.4B in chocolate sales is Circana's retail-panel extrapolation (some measured stores × a scaling factor for coverage) blended with Euromonitor's channel-mix model. Both of those are big decomposition chains ending in a single reported number.
There is no bedrock number that lives outside a Fermi decomposition. Every "primary" figure is the leaf node where someone earlier in the chain decided to stop decomposing and report a total. Push on any leaf hard enough and you find yet another decomposition underneath.
And the decomposition doesn't just go deeper; it goes sideways into whole other disciplines. Take the "$/bar" knob from the upper-bound problem. We handwaved it as $0.50 to $1.50 depending on whether fun-size or full-size dominates. But the honest way to attack that knob is to estimate the joint distribution of chocolate bar types across the US market - Hershey's checkout bars vs. Costco Toblerones vs. Halloween fun-size vs. seasonal boxed chocolates vs. premium Lindt vs. bulk baking chocolate - and that is an entirely different subclass of Fermi problem. It lives in economics and finance, not in demography. It asks:
- Which manufacturers dominate US chocolate volume? (Hershey, Mars, Ferrero, Mondelez, Lindt, Nestlé, Ghirardelli, and a long tail.)
- What is each manufacturer's product mix, and what fraction of their US revenue is bar-form vs. boxed vs. seasonal vs. inclusions?
- What are their per-unit margins, wholesale prices, and retail markups?
- What's the channel distribution - grocery vs. drug vs. mass vs. club vs. convenience vs. specialty - and how does the price-per-bar-equivalent vary across those channels?
Each of those questions is itself a Fermi problem, and now the primary data isn't customs tonnage or retail-panel dollar totals - it's public companies' 10-K filings, segment reporting, investor decks, and gross margin disclosures, and you're reverse-engineering unit economics out of consolidated financial statements that were never designed to answer "how many bars." Hershey's 10-K tells you global net sales for the "North America Confectionery" segment, but not how many kilograms it shipped, and definitely not how many bars; you have to estimate the mapping using industry-average margin data, competitor benchmarks, and public price-per-unit surveys - each of which is another Fermi sub-problem several layers down.
So the recursion isn't a single tree of decompositions; it's a forest. Different branches recruit different fields - demography for the population count, agronomy for cocoa yield, industrial chemistry for processing losses, corporate finance for the manufacturer-and-margin map. Each field has its own conventions for what counts as a "primary" number and where the field stops decomposing. Cross-field, those conventions don't line up, so when you try to reconcile a demographic estimate with a financial one you're not just multiplying uncertainties - you're translating between whole vocabularies of what "measured" means.
Which means our two bounds - the ~25–55B range from cocoa tonnage and the ~19–57B range from dollar sales - are not really bracketing the truth so much as recording where two different chains of Fermi decompositions happened to land. Union those ranges and we get roughly [19B, 57B], which comfortably contains our two original estimates (40B and 44B). But we should be honest about what that means: it doesn't mean the original estimates were validated. It means every path we can take to an answer inherits the same tens-of-percent-per-knob budget, and those knobs compose multiplicatively, so a five-step chain of ±25%-ish knobs smears the answer by roughly ±2× even if every individual knob is honest - and if any single knob is genuinely factor-of-two (like the $/bar knob in the upper-bound problem), it dominates the whole budget by itself.
The interesting thing isn't that our Fermi estimate landed in the middle of the corridor. The interesting thing is that the corridor is the same width as the original problem. We didn't tighten the answer by bringing in real data. We just relocated the uncertainty from "what's the per-capita rate?" into "what fraction of cocoa becomes bars?" and "what's the average price per bar?"
The Fractal Coastline
In a prior post Fermi Problems are Quadrature Problems took a one-sentence bar-trivia question and pulled on the thread until it turned into the population balance equation - the same formalism used to track particles in a turbulent jet, or precipitation of ice particles in the atmosphere contributing to cloud seeding and growth - but applied to Halloween candy and demographic distributions.
This post pulls on the next thread. The moment you try to bound the answer against reality, the "reality" you were reaching for turns out to be another Fermi problem, and each of its inputs is another Fermi problem, and the branches recruit different fields - demography, agronomy, industrial chemistry, corporate finance - each of which terminates its own decomposition at whatever level of aggregation happened to be convenient for its own purposes. There is no floor. The "primary data" you were hoping to stand on is a platform someone else built by decomposing their problem far enough to feel done, and if you step on it and push, it decomposes further.
This is not a bug in Fermi estimation. It's a feature of the world. Reality is a fractal coastline: the closer you look, the more structure appears, and the total "length" you measure depends entirely on the size of the ruler you brought. Every time we ask "how much chocolate?" we're really asking "at what altitude?" A satellite photo says one thing; a customs manifest says another; a Hershey 10-K says a third; a grocery store receipt says a fourth; a five-year-old's Halloween pillowcase says a fifth. None of them are wrong. They are measurements at different scales, and the scales don't compose into a single tidy total because the coastline doesn't have a single tidy length.
What the Fermi decomposition buys us is not a number. It's a map of the coastline at the altitude we chose to fly at, with every landmark labeled: here is where we assumed a population, here is where we assumed a rate, here is where we assumed a cocoa content, here is where we assumed a price per bar. Every landmark is a place we could descend and zoom in further, and the map tells us exactly what we would gain (and lose) by doing so. That's not a consolation prize for failing to find the "real" answer. That is the real answer, at the altitude the question was asked.
The aggregator's confident 2.8 billion pounds is a photograph of the coastline with no scale bar. Our 40 billion bars is a photograph with the scale bar drawn on it, plus a legend explaining what every marking means and how the picture would change if you flew lower. The picture isn't sharper - it's actually a little fuzzier, because we've been honest about the fuzz - but for the first time it's a picture you can navigate by.
And that's the whole thing. We aren't strolling along on solid ground counting Snickers wrappers. We're traversing a fractal coastline, one moment in time at a time, with a map we drew ourselves, at whatever altitude we chose.
It is a chocolate mess. But it is a beautiful chocolate mess.
References
- USDA Economic Research Service, High cocoa prices on smaller global crops lead to decreased U.S. imports in 2023 and 2024.
- USDA Economic Research Service, Food Availability (Per Capita) Data System.
- National Confectioners Association, State of Treating 2025.
- National Confectioners Association, Confectionery Sales Climb to $55 Billion in 2025.
- Prior post: Fermi Problems Are Quadrature Problems.