Luca Barberis
Unit economics & systems thinking

The Numbers With a Currency Sign

Sometimes a business is the same system in two markets and only the currency changes. When that happens the P&L can flip sign with nothing else moving. System theory explains the shape. It doesn’t explain the outcome.

00

The argument in brief

The argument in brief
ClaimWhat it rests onHow sure I am
1. Structure sets the shape, not the outcomeWhen two markets run the same system, they produce the same cohort curves and the same plateaus, and profit is a threshold the curve either clears or doesn’t. Often the system changes across markets too; the cases where it doesn’t are the instructive ones, because they isolate the currency. Seen twice in my own P&Ls, and again in Global Fashion Group’s regional reporting and in Meta’s and Netflix’s regional ARPU.Pattern: established. My two cases: one operator’s sample.
2. When the system holds still, the difference is the numbers with a currency signRevenue per unit follows the local price level (ICP 2021, Hong Kong = 100: Singapore 80, Indonesia 44, India 37). A share of cost per unit does not. In a clean comparison that gap, not execution, decides which side of zero the P&L lands on; in a messy one it is one of several things moving.Price levels: established. The cost-floor mechanism: hypothesis, illustrated with a model.
3. Competition, monopoly scale, stack depth and zero marginal cost can override the Y axisLamoda was GFG’s most profitable region at a 6% margin, in Russia. Shopee’s Asia EBITDA swung from +$320m to -$193m in a year with the same price levels. Management quality explains about 30% of productivity gaps across countries.The caveat is as sourced as the claim. See Where this breaks.
The three takeaways and the honest caveat. Claim 1 and the price-level data are established; the mechanism in claim 2 is my hypothesis, tested with a model rather than measured.
01

Two businesses I watched twice

I have always been a big fan of system theory but I have also always been amazed at how different it is from the financials.

At Eatigo I ran Singapore and Indonesia, about 70 people across the two markets. The app was identical. So was the discount grid, and so was the way we signed restaurants, which I know because I launched Indonesia with it: 600 restaurants in under six months, Ritz-Carlton and Marriott among them, a thousand bookings a day inside three.

Same system, same nuances, one was 10M in revenue and 30% EBITDA, the other was 1-2M in revenue and cash burn. The cohort charts looked like photocopies.

It was amazing, the curves of revenue identical, the Y axis completely different, and one was making profits, another one losing money. Same execution, same people, same meetings, even same numbers when the number doesn’t have a currency sign.

ZALORA was the same film a few years earlier. I ran Malaysia and then Hong Kong, and Indonesia and THE ICONIC sat in the same group reviews as me, month after month.

Again, identical business, similar strategy, similar goals, similar meetings. I would even say the ID was better run, and bam, opposite outcomes. Again the difference was only on the Y axis, the difference was only for numbers with a currency attached. Everything else, fundamentally the same.

I want to be careful with the word same, because most of the time it isn’t. Markets differ in ways that have nothing to do with the currency: cash on delivery, who the competitor is, how many people don’t show up for a booking, whether the roads work. Sometimes the whole system is different and the comparison is useless.

The two cases above are the ones where, as far as I could see from inside, it wasn’t. And those are the interesting ones. When the structure holds still you get to watch what the currency does on its own, and it turns out it does a lot.

Two comparisons is a thin sample and it’s my own, so I went looking for the same pattern in other people’s filings. If it’s there, you read cohort charts differently and you stop grading country managers on a number that was decided before they were hired.

I’ll go through what a systems view sees and why it can’t see currency, then what the P&L sees instead. After that, Meta and Netflix as the clean test, the Rocket Internet fashion family and Uber as the messy one, a small model of the mechanism, the cases that break it, and what I’d do differently now.

Same curve, different Y axis
Same revenue curve, different Y axisTwo panels. Left panel: monthly revenue indexed to each market’s own plateau; the two curves lie on top of each other. Right panel: the same curves in US dollars; one plateaus around 800 thousand a month, the other around 130 thousand. The shape is illustrative and rounded from my own experience at Eatigo; it is not company data.Indexed to own plateau (both markets)0%25%50%75%100%06121824Months from launchIn US dollars, monthly revenue$0k$250k$500k$750k$1000kSingapore + Hong KongDeveloping markets06121824Months from launch
Left, both markets indexed to their own plateau: the curves sit on top of each other. Right, the same curves in dollars. Illustrative shape, rounded from one operator’s memory of Eatigo, not company data; the ratio between the two plateaus is the point, not the absolute numbers.
What was the same, what was different
DimensionEatigo: Singapore and Hong Kong vs Indonesia and other developing marketsZALORA Indonesia vs THE ICONIC
SameProduct, discount mechanics, restaurant acquisition playbook, weekly meeting, cohort shape, time to plateau, share of repeat bookings.Model (Zappos-derived retail plus marketplace), strategy, goals, group reviews, KPIs without units. Indonesia arguably better run.
DifferentRestaurant bill size, fee per seated diner, revenue: ~10M vs 1-2M. EBITDA: ~30% vs cash burn.Basket size, spend per customer (GFG reports €271 per active customer in ANZ vs €130 in SEA), contribution per order, EBITDA sign.
What I concluded at the timeOps was fine, marketing was fine, the team was fine. The market paid a third as much per booking for the same work.Better management in Jakarta didn’t buy a better P&L. It bought a better-run loss.
My reading of the two comparisons. “Same” means I couldn’t tell the markets apart in the operating reviews; it doesn’t mean measured equality. Everything in the right column has a currency sign or is a ratio of two things with a currency sign.
02

What system theory sees

Forrester built system dynamics on one idea: the structure of a system, its stocks, flows, loops and delays, produces its behaviour, and the people at the desks matter less than the loops they sit in.[8]

Meadows later ranked the places where you can push on a system, twelve of them, and put constants, parameters and numbers at the bottom. Her line is that almost all the attention goes there and it rarely changes behaviour.[9]

She’s right, and that is what fooled me. The Jakarta cohorts behaved like the Singapore cohorts. A systems-minded operator looks at that, recognises the loops, and files the two markets under the same business. I did.

The systems view can’t see currency and isn’t meant to; that is what lets one model describe a restaurant marketplace in six countries.

Physicists have a formal version of this, and I’m a tourist here. Buckingham’s π theorem says the behaviour of a system depends on dimensionless groups; anything with a unit attached can be scaled away.[11]

Retention is dimensionless. Orders per customer, growth rate, months to plateau, all dimensionless. Revenue has a unit.

West and his co-authors looked at 31,553 US and 3,160 Chinese public companies and found sales scale with assets at nearly the same exponent in both countries.[12] The exponent is the shape and the constant in front of it is the level. Two markets can share one and differ tenfold on the other, and no dashboard I have ever used would notice.

Currency-blind metrics and currency-bound metrics
Currency-blind (same in both markets)Currency-bound (different in both markets)
Retention curve, month-12 retention %Revenue, revenue per active customer
Orders or bookings per customerBasket size, bill size, fee per booking
Growth rate, time to plateau, share of repeatContribution per order after delivery, returns, payment, service
Conversion, no-show rate, review scoreFixed cost floor per market: tech, rent, group overhead, capital
Meetings, dashboards, playbook, moraleEBITDA sign
The left column is what a systems model and a cohort dashboard show; the right column is what a P&L shows. My split, informed by the dimensional-analysis argument above; the last row is the one that decides survival.
03

What the P&L sees

Profit isn’t a behaviour that a loop produces. It’s a subtraction: revenue per unit minus cost per unit, times units, minus whatever fixed cost the market carries. Above zero you’re in Singapore. Below, Jakarta.

The two halves of the subtraction don’t cross borders together, and that’s the point, at least in the cases where the structure holds still.

Revenue per unit follows local prices, almost by definition. Eatigo charged restaurants a fixed fee per seated diner, in local currency, against a local bill.[28] A marketplace commission is a cut of a local basket.

The reason local prices differ so much is old economics. Balassa and Samuelson, separately in 1964, argued that tradable goods obey the law of one price and services don’t, so the general price level climbs with income.[1][2] Bhagwati wrote the plain version in 1984 under the title “Why are services cheaper in the poor countries?”, which does most of my work for me.[3]

The International Comparison Program measures the result, and its 2021 Asia-Pacific round happens to cover the exact Eatigo footprint.

The Y axis, measured by statisticians
Price level index, 2021 ICP Asia-Pacific round, Hong Kong = 100Bars: Hong Kong 100, Singapore 80, Philippines 52, Malaysia 48, Indonesia 44, India 37. Ratio of purchasing power parity to exchange rate at the GDP level.0285582110General price level, Hong Kong = 100100Hong Kong80Singapore52Philippines48Malaysia44Indonesia37India
General price level at the GDP level, Hong Kong = 100, 2021 ICP round coordinated by the Asian Development Bank. Philippines computed as nominal over real GDP from the same tables. A country-wide average: the Jakarta or Manila diner Eatigo served is richer than the national figure, which narrows but doesn’t close the gap. Thailand, Eatigo’s home market, is in the round but I didn’t extract it.

Cost per unit is the problem, because only part of it follows. Wages and rent do. Cloud, software licences, fuel, the motorbike, the parcel bag, the phone in the courier’s hand, the engineering team in the hub, the group overhead and the cost of the money mostly don’t. GFG’s central costs alone were about €25m in 2024, which is 3.4 points of margin across the whole group.[18]

So if a basket is worth 44% of a Hong Kong basket and the work to pick, pack, ship, take back and support it costs, say, 70% of the Hong Kong work, the margin has moved a lot further than the basket did.

Which line items follow the local Y axis
Line itemFollows local price level?Why
Commission, fee per booking, basket-linked revenueYes, fullyPriced against a local bill or basket in local currency.
Couriers, customer service, sales reps, warehouse labourYes, mostlyNon-tradable services; wages track local productivity and prices.
Rent, local marketing in local auctionsPartlyLocal, but capital-city rents and ad auctions carry global bidders.
Fuel, vehicles, packaging, devicesNoTradable goods, close to world prices.
Cloud, SaaS, payment gateway fixed feesNoPriced in dollars from a global list.
Central tech, group management, cost of capitalNoSits in a hub; funded in hard currency.
My classification, a hypothesis, using the tradable versus non-tradable logic of Balassa and Samuelson. GFG states that the majority of its operating costs are a natural hedge because they are local; I address that in Where this breaks.
04

The same code, four Y axes

The cleanest test is a product that doesn’t change at all when it crosses a border. Meta runs one ad system, one feed, one auction, everywhere. In Q4 2023 it made $68.44 per user in the US and Canada, $23.14 in Europe, $5.52 in Asia-Pacific and $4.50 in the rest of the world.[23]

Netflix runs one app and one catalogue. From its 2024 10-K, monthly revenue per paid membership works out at about $17 in the US and Canada, $11 in Europe, Middle East and Africa, $8 in Latin America and $7 in Asia-Pacific.[24] Same code, same recommendation engine and, I assume, the same meetings.

One ad system, four prices
Meta average revenue per user by region, Q4 2023Bars: United States and Canada 68.44 dollars, Europe 23.14, Asia-Pacific 5.52, Rest of World 4.50. Worldwide 13.12. One quarter, one company, one product.$0$20$40$60$80Quarterly ARPU, US$$68.44US & Canada$23.14Europe$5.52Asia-Pacific$4.50Rest of World$13.12Worldwide
Meta’s own regional ARPU for Q4 2023, from its earnings presentation. Revenue is apportioned to the user’s estimated geography; the regions are coarse, and Asia-Pacific mixes Japan and Australia with Indonesia and India, which flatters the low end.
One catalogue, four prices
Netflix monthly revenue per paid membership by region, 2024Bars computed from the 2024 10-K: United States and Canada about 17.0 dollars a month, Europe Middle East and Africa about 10.9, Latin America about 8.1, Asia-Pacific about 7.2. Computed as regional streaming revenue divided by the average of opening and closing paid memberships, divided by twelve.$0$5$10$15$20Monthly revenue per paid membership, US$ (computed)$17.04UCAN$10.87EMEA$8.12LATAM$7.15APAC
Computed by me from Netflix’s 2024 10-K: regional streaming revenue divided by the average of opening and closing paid memberships, divided by twelve. Netflix’s own reported figures differ slightly (it uses average memberships across the year). Includes hedging gains in 2024 revenue.

India shows who sets the price. In December 2021 Netflix cut its Indian basic plan from ₹499 to ₹199 a month, a 60% cut, and the CFO said afterwards they had been overpriced relative to the market.[26] Netflix got to choose where on India’s price line to sit, and that was all it got to choose.

Both companies can live with a $5 region because their marginal cost is close to zero. Shapiro and Varian described information goods 25 years ago: big fixed cost, almost nothing per extra copy, so a fourth region is worth having at any price above the cost of moving bits.[17]

A restaurant marketplace isn’t that. Every booking needs a rep who signed the restaurant and someone to chase the no-show; every parcel needs a person on a motorbike. Low prices hurt everyone’s revenue line; they only kill the businesses that have to touch every unit.

05

One playbook, five continents

The Rocket Internet and Kinnevik fashion family is the closest thing to a controlled experiment I know, partly because I was in it. Zalando in Europe, THE ICONIC in Australia, ZALORA in South-East Asia, Dafiti in Latin America, Jumia in Africa, and before the exits Jabong in India, Namshi in the Gulf and Lamoda in Russia.

One Zappos-derived playbook, similar org charts, the same review format, launched within a couple of years of each other on five continents. Here is what the public filings say about them now.

One playbook, five continents: spend per customer and profitability, most recent full year
Platform (region)GMV or NMV per active customerProfitabilitySource
Zalando (Europe)€295 (GMV, LTM 2024); basket €60.9+4.8% adj. EBIT margin, €511m, FY2024[21]
THE ICONIC (Australia, New Zealand)€271 (NMV, FY24R, Edison estimate)about +€16m adj. EBITDA, FY24R; €26m in FY2025[18][19]
ZALORA (Indonesia, Philippines, Malaysia, Singapore, Hong Kong)€130 (NMV, FY24R, Edison estimate)about -€3m adj. EBITDA, FY24R; the only GFG region never profitable in FY19-21[18]
Dafiti (Brazil, Colombia)€89 (NMV, FY24R, Edison estimate)about -€7m adj. EBITDA, FY24R; €3m positive in FY2025[18][19]
Jumia (Nigeria, Egypt and others)Average order value $39.6 (Q1 2024); GMV $721m on $167.5m revenue-$51.3m adj. EBITDA, FY2024, about -31% of revenue[22]
ExitedThailand and Vietnam (2016), Jabong India (sold 2016), Namshi Gulf (from 2017), Lamoda Russia (sold 2022, for the war), Argentina (2023), Chile (2025), Taiwan (2025)[18]
Zalando figures are group adjusted EBIT on revenue; GFG regional figures are adjusted EBITDA on estimated regional revenue, before €25m of central costs; Jumia is adjusted EBITDA on revenue. The margin definitions differ, so read the signs and the spend-per-customer column, not the decimals. GFG’s SEA segment includes Singapore and Hong Kong, which pulls its average up.
The sign of the margin follows the spend per customer
Spend per active customer against operating margin, four fashion platformsFour points. Zalando in Europe, 295 euro of GMV per active customer and a 4.8 percent adjusted EBIT margin. THE ICONIC in Australia and New Zealand, 271 euro per active customer and about 4.4 percent adjusted EBITDA margin. ZALORA in South-East Asia, 130 euro and about minus 1.8 percent. Dafiti in Latin America, 89 euro and about minus 3.6 percent. The two profitable points sit on the right, the two loss-making on the left.Margin on revenue, %-8%-4%0%4%8%€0€100€200€300GMV or NMV per active customer, last twelve monthsZalando (Europe)€295, +4.8% (adj. EBIT, FY24)THE ICONIC (ANZ)€271, +4.4% (adj. EBITDA, FY24R)ZALORA (SEA)€130, -1.8% (adj. EBITDA, FY24R)Dafiti (LatAm)€89, -3.6% (adj. EBITDA, FY24R)
Four platforms, one axis each. Margin definitions differ (Zalando adj. EBIT, GFG regional adj. EBITDA before central costs), so this is a picture of the sign and the ordering, not a regression. Four points prove nothing on their own; they are consistent with the two cases I lived.

Uber is the ride-hailing version. It burned about $2bn in China before selling to Didi in 2016, merged with Yandex in Russia in 2017, and sold South-East Asia to Grab in 2018 after putting in $700m.

Reuters reported at the time that India was more than 10% of Uber’s trips and not making money, while the markets Uber called core, the US, Australia, New Zealand and Latin America, were profitable or close to it.[25] Competition drove those exits at least as much as price levels did. I’ll come back to that, because it’s the objection I find hardest to answer.

06

A floor under a shrinking basket

Here is the mechanism as a toy model, three assumptions and one chart. Revenue per unit equals the local price level. Cost per unit splits into a part that follows local prices and a part that doesn’t, which I’ll call the floor. Calibrate to a 30% margin at price level 100, roughly what Singapore and Hong Kong gave me, and let the price level fall.

With no floor, the margin stays at 30% wherever you go. That’s the McDonald’s case, in principle: the burger, the wage and the rent all move together, so a franchise in Jakarta should earn a similar margin on a much cheaper burger. I haven’t checked franchise margins by country, so treat that as a guess.

With a 25% floor, the margin at Indonesia’s price level of 44 drops to about 8%. With a 50% floor it’s about minus 14%. Nothing else in the model moved.

In real life the structure and the currency usually move together, and this only isolates the second.

A floor under a shrinking basket
Margin against local price level for three cost-floor sharesA model, not data. Per-unit revenue equals the local price level. Per-unit cost is split into a share that follows local prices and a share that does not, the floor. Margin is calibrated to thirty percent at price level one hundred. With no floor, margin stays at thirty percent everywhere. With a twenty-five percent floor, margin at price level forty-four is about eight percent. With a fifty percent floor it is about minus fourteen percent. Vertical markers show the 2021 ICP price levels of Hong Kong, Singapore, the Philippines, Malaysia, Indonesia and India with Hong Kong at one hundred.EBITDA margin, calibrated to 30% at price level 100 (model)-40%-20%0%20%40%India 37Indonesia 44Malaysia 48Philippines 52Singapore 80Hong Kong 100406080100Local price level (Hong Kong = 100)0% cost floor25% cost floor50% cost floor
A model, not data. Per-unit revenue equals the local price level; per-unit cost is split into a local-tracking share and a floor; margin calibrated to 30% at 100. The markers are the ICP 2021 price levels from the earlier chart. The floor shares are assumptions; a real business needs its own split, and the split changes as the business localises.
Worked example: one order in two price levels (25% cost floor)
Per orderPrice level 100 (Hong Kong)Price level 44 (Indonesia)What moved
Revenue captured10044Follows the basket.
Local-tracking cost (labour, rent, local marketing)52.523.1Follows local wages and prices.
Floor cost (fuel, devices, cloud, central tech, capital)17.517.5Doesn’t move.
Contribution30 (30%)3.4 (8%)The floor now eats 40% of revenue instead of 17.5%.
Illustrative per-order economics at a 25% floor, the middle line of the chart. Numbers are invented to match the model’s calibration, not taken from any company. The point is the last row: the same work, the same number of touches, half the basket.

GFG would disagree with me, or at least its disclosure does. It says most of its operating costs are a natural hedge because they are local, and that only about €24m of a €416m cost base sits centrally.[18] That’s true about currency translation and, I think, misleading about unit economics.

The courier in Jakarta is paid in rupiah. The motorbike isn’t 56% cheaper, and neither is the fuel, the phone or the bag. And a customer who spends €130 a year needs the same number of picks, packs, deliveries and returns as one who spends €271. That’s my reading, not GFG’s, and the regional cost lines I would need to prove it aren’t public.

07

Where this breaks

Everything above holds management constant, which it rarely is. Bloom, Sadun and Van Reenen scored management practices at more than 11,000 firms in 34 countries and estimate that those differences explain about 30% of the productivity gap between countries.[15] If you’re comparing a well-run market with a badly run one, that is the story, and the currency is noise.

Competition is the objection I can’t fully answer, because it is currency-blind too and it moves faster. Uber didn’t lose China to the price of a ride, it lost it to Didi. Shopee’s Asia business posted +$320m of adjusted EBITDA in Q4 2022 and -$193m in Q4 2023, same price levels, after competition intensified.[27]

So the price level is a ceiling, not a forecast. In Singapore the ceiling was high enough that we could lose share and still make money. In Jakarta it wasn’t, and losing share there meant losing money faster.

Scale can beat it. Lamoda was GFG’s most profitable region, at a 6% adjusted EBITDA margin and 36% of group NMV in 2021, and it was in Russia, not Australia.[18]

Sea’s first annual profit, in 2023, came with Shopee still losing money in Q4 (adjusted EBITDA minus $225m) while SeaMoney’s credit business made $550m for the year; owning more of the stack changes what you have on the revenue side of the floor.[27]

And Rachleff’s law, the closest thing VCs have to my claim, says that when a great team meets a lousy market the market wins.[16] That’s about whether the dogs eat the dog food. Mine is narrower. The dogs can love the food and still not cover the delivery.

Where this breaks
Counterexample or limitWhat it showsWhat survives
Lamoda, Russia: GFG’s most profitable region at 6% marginA market outside the rich world cleared the threshold, at a scale (36% of group NMV) GFG never reached elsewhere.Category scale raises the ceiling; most markets don’t get it.
Shopee Asia: +$320m to -$193m adj. EBITDA in four quartersCompetition moves the outcome with the axis unchanged.The axis is a ceiling, not a prediction.
Meta, Netflix in APAC and Rest of WorldNear-zero marginal cost makes low-ARPU regions accretive.The mechanism needs a physical cost floor per unit.
Bloom, Sadun, Van Reenen: management ~30% of TFP gapsWhen management differs, it dominates.My claim is conditional on “same execution”.
GFG’s SEA segment mixes Singapore and Hong Kong with IndonesiaThe €130 figure overstates the poor-market number and the -€3m understates the poor-market loss, probably.Direction unchanged; magnitude uncertain.
Often the system changes too: cash on delivery, return rates, no-shows, roads, competitorsThe clean comparison is the exception. In most market pairs the structure moves with the price level, and the two can’t be separated from the outside.The mechanism applies to the currency-bound share of the gap. The rest is a different business, and needs a different diagnosis.
My sample: two comparisons, one operator, memory-rounded numbersAnecdote, however vivid.The public filings point the same way; they don’t prove the mechanism.
Time: price levels rise, and so does the addressable sliceBlume’s India1 is ~120m people at ~$12k per head, Mexico-sized, inside a 1.4bn country.[29]The Y axis is a moving target; the population is the wrong one to size on.
The counterexamples and limits I found, and what I think each one does to the claim. Rows 1 to 3 limit the mechanism; rows 4 to 7 are limits of scope and of my evidence.
08

What this changes

If even half of this is right, I’d change five things. Size markets by revenue per unit against the floor, not by population.

Indonesia’s 280 million people aren’t the addressable market for anything with a per-order cost; the slice that pays a Singapore basket is, and it’s small. Blume’s India1 arithmetic, roughly 120 million people at about $12k a head inside a country of 1.4 billion, is the honest version of this for India.[29]

Ask for the currency axis on every cohort chart, because an indexed retention curve hides exactly this, which is why people index it. Put contribution per customer in dollars next to it.

Sequence rich markets before developing ones. Singapore and Hong Kong paid for the rest of Eatigo, and Netflix could cut India by 60% because three other regions had already paid for the catalogue, or at least that’s how I read it.

Decide which floor costs you will localise and price the rest in. GFG runs its central tech from Vietnam; Sea built its own logistics in Asia and took 12% off platform logistics cost per order in a year. A country manager who can’t move the floor can only move volume, and volume against a negative contribution is a faster way to lose.

And write the exit rule in terms of the floor, not growth. GFG’s exits from Argentina and Chile each helped adjusted EBITDA by under €3m, more than ten years after launch.

Claims ledger
ClaimEvidenceConfidenceWhat would change my mind
Cohort shapes are similar across markets for the same modelMy two cases; West et al. on identical scaling exponents across countriesMedium. I have no cross-market cohort dataset, only my own.A multi-market cohort study showing retention shapes diverge with income.
Revenue per unit follows the local price levelICP 2021 price levels; GFG spend per customer; Meta and Netflix regional ARPU; Netflix’s India repricingHigh.A local-consumption marketplace earning developed-market revenue per unit in a low price-level market without a premium niche.
A share of cost per unit does not follow, so margin falls faster than the basketBalassa-Samuelson logic; GFG central costs; the modelMedium-low. Hypothesis with a model, not measured cost lines.Regional cost breakdowns showing per-order cost falling in proportion to the basket. GFG’s “natural hedge” claim points this way for currency, not for units.
The two comparisons were clean, i.e. the system held stillSame product, playbook, reviews and cohort shapes, as far as I could see from insideMedium. I was inside; I couldn’t measure it, and markets usually differ in structure too.Evidence of a structural difference I missed: return rates, no-show rates, payment mix, competitor set.
This, not execution, explained my two outcomesSame teams, same reviews, same shapes; better-run Indonesia still lostMedium. One operator, no counterfactual.Evidence that Singapore’s team or restaurant mix was materially better than I remember.
Competition and scale can override the axisUber’s exits; Shopee 2022-2023; LamodaHigh.Nothing; this is the caveat, and it’s better sourced than the claim.
The claims ledger. Confidence is mine and subjective. The last column is what I’d need to see to drop each claim.
09

Takeaways

Takeaways
#Takeaway
1System theory explains the shape of a business and is currency-blind by design. When two markets run the same system they show the same curves, and that tells you nothing about which one makes money. Often the system changes with the market too; the clean cases are the useful ones because they isolate the currency.
2Profit is a threshold on revenue per unit minus cost per unit. Revenue per unit follows the local price level (Hong Kong 100, Singapore 80, Indonesia 44, India 37). A share of cost per unit doesn’t. With a 25% floor, a 30% margin becomes 8%; with a 50% floor, minus 14%, with nothing else changed.
3The caveat: competition, monopoly scale, stack depth, zero marginal cost and management quality all override the axis, and did (Lamoda, Shopee, Meta, Netflix). The axis is a ceiling, not a forecast.
The compressed version. If you read only this table and the summary box, you have the argument and its main hole.

I’d still launch Jakarta the way I did. I’d stop expecting Singapore’s P&L from it, and I’d stop marking the people running it against a number the price level decided before they arrived.

10

Sources

Everything cited above, plus the research that shaped the argument. Where I computed a figure from a filing, the note says so.

  1. Bela Balassa. The Purchasing-Power Parity Doctrine: A Reappraisal. Journal of Political Economy, 72(6), 584-596, 1964. The productivity-differential explanation of why price levels rise with income.
  2. Paul A. Samuelson. Theoretical Notes on Trade Problems. Review of Economics and Statistics, 46(2), 145-154, 1964. The parallel statement of the same mechanism.
  3. Jagdish N. Bhagwati. Why Are Services Cheaper in the Poor Countries?. Economic Journal, 94(374), 279-286, 1984. The non-tradables argument in its plainest form.
  4. Irving B. Kravis and Robert E. Lipsey. National Price Levels and the Prices of Tradables and Nontradables. American Economic Review, 78(2), 474-478, 1988. Empirical decomposition of national price levels.
  5. Asian Development Bank, International Comparison Program. 2021 ICP economy results: Indonesia. ADB, 2024. PLI 44 with Hong Kong = 100. Companion documents for Singapore (80), Malaysia (48), India (37) and the Philippines (computed 52) at the same site.
  6. Asian Development Bank, International Comparison Program. 2021 ICP economy results: Singapore. ADB, 2024. Singapore's price level at 80% of Hong Kong's and 124% of the regional average.
  7. World Bank. Price level ratio of PPP conversion factor (GDP) to market exchange rate. World Development Indicators, accessed September 2026. Australia at 0.90 and Brazil at 0.49 of the US price level in 2023, the THE ICONIC versus Dafiti gap.
  8. Jay W. Forrester. Industrial Dynamics. MIT Press, 1961. Structure produces behaviour. Linked to the system dynamics overview; the book itself is out of print.
  9. Donella Meadows. Leverage Points: Places to Intervene in a System. The Sustainability Institute, 1999. Parameters at the bottom of the list; most attention goes there; they rarely change behaviour.
  10. Peter M. Senge. The Fifth Discipline. Doubleday, 1990. The management-book version of structure over events.
  11. Edgar Buckingham. On Physically Similar Systems: Illustrations of the Use of Dimensional Equations. Physical Review, 4(4), 345-376, 1914. Behaviour depends on dimensionless groups. I am a tourist here.
  12. Jiang Zhang, Christopher P. Kempes, Marcus J. Hamilton, Ruyi Tao and Geoffrey B. West. Scaling laws and a general theory for the growth of public companies. arXiv:2109.10379, 2022. 31,553 US and 3,160 Chinese companies; sales scale sublinearly with assets at nearly identical exponents.
  13. Geoffrey West. Scale: The Universal Laws of Growth, Innovation, Sustainability, and the Pace of Life. Penguin Press, 2017. Linked to the Physics World review. The Compustat scaling of companies; same exponent, different prefactor.
  14. Richard Rumelt. Good Strategy Bad Strategy. Crown Business, 2011. Diagnosis before action; here, diagnose the axis before the team.
  15. Nicholas Bloom, Raffaella Sadun and John Van Reenen. Management as a Technology?. NBER Working Paper 22327, 2016. 11,000+ firms, 34 countries; management practices explain about 30% of cross-country TFP differences. The opposition.
  16. Marc Andreessen. The only thing that matters (Rachleff's law). pmarchive, 2007. When a great team meets a lousy market, market wins.
  17. Carl Shapiro and Hal R. Varian. Information Rules: A Strategic Guide to the Network Economy. Harvard Business School Press, 1999. High fixed cost, near-zero marginal cost; why Meta and Netflix survive low-ARPU regions.
  18. Russell Pointon and Nick Hawkins, Edison Investment Research. Global Fashion Group: Refined and redefined. Edison, 29 May 2025. Commissioned by GFG. Regional adj. EBITDA (ANZ +€16m, SEA -€3m, LatAm -€7m), NMV per active customer (€271, €130, €89), central costs €25m, Lamoda's 6% margin, country exits, and the natural hedge statement.
  19. Global Fashion Group. GFG reports Q4 and FY 2025 results. Press release, 4 March 2026. ANZ €26m adj. EBITDA, LATAM €3m, SEA NMV down 15.2%.
  20. Global Fashion Group. GFG delivers first adj. EBITDA positive year. Press release, 1 March 2021. The group's first profitable year and the €10bn ambition it has since dropped.
  21. Zalando SE. Q4 2024 Fact Sheet and Annual Report 2024 key figures. Zalando, March 2025. Adj. EBIT €511m, 4.8%; GMV per active customer €295; basket €60.9.
  22. Jumia Technologies AG. Q1 2024 and Q4 2024 results. SEC Form 6-K exhibits, 2024 and February 2025. FY2024 revenue $167.5m, GMV $720.6m, adj. EBITDA -$51.3m; Q1 2024 AOV $39.6.
  23. Meta Platforms. Q4 2023 earnings presentation. Meta, February 2024. ARPU by user geography: US & Canada $68.44, Europe $23.14, Asia-Pacific $5.52, Rest of World $4.50, worldwide $13.12.
  24. Netflix, Inc.. Form 10-K for the fiscal year ended December 31, 2024. SEC, January 2025. Regional streaming revenue and paid memberships used for the computed monthly revenue per membership.
  25. Reuters (via Daily FT). Uber sells Southeast Asia business to Grab after costly battle. 27 March 2018. $700m invested in SEA, $2bn in China; India over 10% of trips and unprofitable; core markets listed as profitable.
  26. Jessica Goodfellow, Campaign Asia. Netflix slashes India subscription prices by up to 60%. 15 December 2021. Basic plan from ₹499 to ₹199. The CFO's remark that Netflix was overpriced relative to the market is reported by exchange4media, April 2022.
  27. Sea Limited. Fourth quarter and full year 2023 results. SEC Form 6-K exhibit, March 2024. First annual profit ($162.7m); Shopee Asia adj. EBITDA -$192.9m in Q4 2023 vs +$320.0m in Q4 2022; SeaMoney adj. EBITDA $550m.
  28. PhocusWire and Dealroom. TripAdvisor gets deeper into Eatigo. PhocusWire, 2018; Dealroom company profile. Eatigo's markets, the TripAdvisor rounds (over $25m total) and the fixed fee per seated diner model.
  29. Sajith Pai and Amal Vats, Blume Ventures. Indus Valley Annual Report 2024. Blume Ventures, 2024. India1 at roughly 120 million people and about $12,000 per head, the Rule of 30 households, and why population is the wrong denominator.
  30. Delivery Hero SE. FY 2025 results. Press release, February 2026. MENA as the group's profit engine, Asia (Korea-heavy) as the largest and most contested segment. A muddier case; competition and price levels are hard to separate.
  31. Alan M. Taylor and Maurice Obstfeld (FRBSF). The Penn effect and the Balassa-Samuelson story. Federal Reserve Bank of San Francisco working paper, 2004. A sceptical review of how far the productivity story explains the price-income relationship.
  32. Luca Barberis. Operator profile. lucabarberis.com, 2026. The roles behind the two cases: Eatigo Singapore and Indonesia, ZALORA Malaysia and Hong Kong.