A man draws a smile, 1992.
In 1992, the founder of Acer sketched a curve to explain a problem his own company was living. Stan Shih had built one of the largest PC makers on Earth, and the better Acer got at making computers, the less money making computers seemed to earn. His sketch plotted value added against the stages of the hardware chain, and it came out shaped like a smile: high on the left, where the intellectual property and components live; high on the right, where the brand, the services, and the customer live; and lowest in the exact middle — at manufacturing and assembly, the part everyone could see and touch.1
Shih drew it as a diagnosis. It turned out to be a law. Every hardware market that has commoditized since has arranged itself along his smile, without exception and without mercy. The middle of the curve works hardest, employs the most, and keeps the least — because the middle is the only part of the chain where the output of one firm can be swapped for the output of another. Value does not reward effort. Value rewards what cannot be substituted.
The economics under the smile are worth stating precisely, because they will run this industry too. Three kinds of money are made in any hardware era. At the upstream end, innovation rent — the temporary monopoly of knowing what others do not, defended by patents, silicon, and tacit craft. At the downstream end, scarcity rent — the durable income of owning what others need: the customer, the data, the installed base. And in the middle, commodity margin — the wage of doing well what many can do at all, competed down toward cost by every entrant. The classical economists drew the distinction two centuries ago: rent flows to what is scarce by nature; wages to what is reproducible.2 The middle earns wages. The ends earn rents. Everything else in this dispatch is that sentence, with dates.
The data behind Figure I
| Position | Physical-AI equivalent | Substitutability |
|---|---|---|
| Upstream end — high value | Embodied models, IP, silicon | Low — scarce minds, scarce talent |
| Middle — low value | Body manufacturing and assembly | High — merchant parts, many makers |
| Downstream end — high value | Fleet ownership, data, service | Low — experience is rivalrous and private |
The experiment has been run twice. Same result.
Autopsy one: the personal computer. IBM invented it, and in an act of speed over control, built it from open, merchant parts. Every element of the body could be bought by anyone — so within a decade, hundreds of firms made PC bodies and none of them made money doing it. The two components that could not be substituted — the processor and the operating system, the machine's mind — were owned by exactly two companies, and Intel and Microsoft divided the industry's entire profit pool for twenty years. In 2005, IBM sold the business it had invented to Lenovo for $1.75 billion3 — a company then worth roughly a week of Microsoft's market value.
Autopsy two: the handset. Nokia was, by units, the greatest hardware company of its era — at peak it built roughly four of every ten phones on Earth, with supply chains and manufacturing discipline nobody could match. None of it mattered. When the phone's value migrated into its mind — the OS, the silicon, the ecosystem — Nokia's mastery of the body became a mastery of the middle of the curve. It sold its devices business to Microsoft in 2013 for about $7.2 billion.4 Meanwhile Apple, holding a minority of unit share, captured more than eighty percent of the industry's profits — in some quarters over ninety.5 The company that assembled those iPhones, the largest manufacturer in human history, operates on margins in the low single digits.6
Two eras, one verdict. The body is where the work is. The mind is where the money is. And note who the losers were: not fools — the best manufacturers alive. The curve does not punish incompetence. It punishes position.
Clayton Christensen gave the mechanism its sharpest name: the law of conservation of attractive profits. When a layer of the stack becomes modular and good enough — interfaces standardized, performance beyond what the customer can absorb — the profits in that layer do not vanish. They migrate to the adjacent layers where hard, integrated problems remain.7 The PC body modularized; profit migrated inward, to the processor and the operating system. The handset body modularized; profit migrated upward, to the silicon-OS-ecosystem stack. The robot body is modularizing now, in public, on a published price sheet — and the adjacent layers still hard to do are exactly two: the mind and the fleet. The profits are already in the mail.
The data behind Figure II
| Player | Unit share | Profit share |
|---|---|---|
| Apple (owner of OS + silicon) | ~15–20% | 80%+, above 90% in some quarters |
| All other handset makers | ~80–85% | Under 20%, several loss-making |
| Contract assembler of the iPhone | Largest manufacturer on Earth | Low-single-digit operating margin |
The robot body is already racing to the middle.
Now watch the same film start again, at higher speed. The robot body is being built from parts other industries already commoditized — Dispatch No. 001 called it the child of a car and a phone — which means it inherits their cost curves on day one, with no decade of grace. A capable humanoid has fallen roughly 70 percent in two years; a commercial platform now lists near $16,000, and dozens of manufacturers — a Chinese industrial base above all — are racing each other down the price sheet exactly as the handset makers once did.8
This is not a failure of the body business. It is the body business working as designed. Cheap bodies are what makes the era possible — someone has to play the middle of the curve, and the middle will be played magnificently, at scale, on margins that would make an airline flinch. The question the macro asks is not whether bodies will be built. It is who ends up holding Nokia's position — and who ends up holding Apple's.
History suggests most of today's most photographed robot companies are competing, with total sincerity, for the Nokia seat. The bodies will be marvels. The margins will be memories.
And the deflation is not a forecast — it is a schedule, and the schedule has a name. In 1936, studying aircraft factories, the engineer T. P. Wright observed that every doubling of cumulative production cuts unit cost by a roughly fixed fraction — a learning rate the industry then never escapes.9 Solar photovoltaics have obeyed it for half a century at about twenty percent per doubling; lithium-ion cells at roughly eighteen. Now see what Wright's Law implies for any young product: when cumulative output is small, doublings arrive absurdly fast — the ten-thousandth humanoid doubles the fleet the five-thousandth completed. A product early on its learning curve does not descend gently. It falls down the stairs.
The data behind Figure III
| Product | Learning rate | Record |
|---|---|---|
| Aircraft (Wright's original study) | ~20% per doubling | Observed 1936, direct labor hours |
| Solar photovoltaic modules | ~20% per doubling | Sustained since 1976 — Swanson's law |
| Lithium-ion cells | ~18% per doubling | Sustained since 1991 |
| Humanoid platforms | Early-cycle — faster than mature rates | −70% in two years from a tiny base |
What deflates, and what appreciates.
So strip the era to its balance sheet. On one side, everything that deflates: actuators, sensors, batteries, compute-per-dollar, the body entire — all of it on merchant supply chains, all of it substitutable, all of it sliding down the middle of the smile. On the other side, the four things that cannot be bought off a price sheet.
Why do minds resist the commoditization devouring the bodies? The deepest answer was written in 1966 by the philosopher Michael Polanyi: we know more than we can tell. Most human competence — how to see, how to grip, how to balance — is tacit: real, reliable, and impossible to write down.10 Polanyi's paradox is why the physical world resisted automation for seventy years — you cannot program what no one can articulate — and it is also, note the symmetry, why the people who finally cracked it cannot be mass-produced. The craft of training embodied minds is itself tacit: it lives in a few hundred practitioners and moves by apprenticeship, not curriculum. The paradox that protected the work now protects the workers.
The minds. Frontier embodied models are made by a population of engineers small enough to fit in a lecture hall, and the market has begun pricing them like franchises: compensation packages for individual researchers have been reported in nine figures.11 The compute. Frontier training runs now sit on the order of 10²⁶ FLOPs, concentrated in a handful of institutions.12 The data. As No. 001 argued, there is no internet of touch — fleet experience is manufactured, rivalrous, and private. And the position. As No. 002 argued, a mind with a body is an asset that can, for the first time, be held.
Deflation on one side of the ledger, appreciation on the other — and the deflation funds the appreciation: every dollar the body sheds widens the market the mind can address. That is the whole macro, in one sentence. Cheap bodies are not the threat to this industry. Cheap bodies are the subsidy.
One last piece of theory, because it decides the endgame. The economist W. Brian Arthur showed that markets built on increasing returns — where each unit of success makes the next cheaper to win — do not settle into the polite equilibria of commodity markets. They tip: small early leads compound into locked-in dominance.13 The fleet flywheel of No. 001 is increasing returns in its purest form — more bodies, more experience, better minds, more bodies — with a property no software market ever had: the compounding asset is titled. The screen era's winners tipped their markets and still leaked their data. This era's winner tips the market and keeps the corpus. That is not a bigger prize. It is a different kind of prize.
Own the ends. Rent the middle.
Every institution in this industry is, knowingly or not, choosing a seat on Shih's curve. Ours is chosen and stated. ARBX builds minds — the upstream end, where substitution fails because the asset is a trained intelligence and the people who made it. And ARBX holds fleets — the downstream end, where substitution fails because the asset is accumulated experience, titled to one ledger. The middle we buy, gratefully, at whatever price the magnificent body-makers race each other down to.
We do not make bodies. We make what makes bodies valuable — and we keep what the bodies learn. The PC century and the handset decade were tuition. This time, the position is taken before the music starts. Made here, held here.
Notes & Sources
- Stan Shih, Acer, 1992 — the "smiling curve" of value added across the IT hardware chain. smiling curve ↗ ↑
- Economic rent versus wages — returns to non-reproducible factors, from Ricardo onward. economic rent ↗ ↑
- IBM's sale of its Personal Computing Division to Lenovo, announced December 2004, ~$1.75B including assumed debt; completed 2005. ibm pc ↗ ↑
- Microsoft's acquisition of Nokia's Devices & Services business, announced 2013, €5.44B (~$7.2B). Nokia's peak global handset share reached roughly 40% in 2008. nokia ↗ ↑
- Industry profit-share analyses (Canaccord Genuity and others), mid-2010s: Apple above 80% — in some quarters above 90% — of global smartphone operating profits on roughly 15–20% of units. smartphone profit share ↗ ↑
- Hon Hai Precision (Foxconn) operating margins have run in the low single digits for over a decade. foxconn ↗ ↑
- Clayton Christensen, the law of conservation of attractive profits — The Innovator's Solution, 2003. christensen ↗ ↑
- Unitree G1 class platforms from ~$16,000, 2024–, against research-grade platforms at $150K–$250K+ two years prior; see Dispatch No. 001, Figure II. unitree ↗ ↑
- T. P. Wright, "Factors Affecting the Cost of Airplanes," 1936 — the original learning-curve study. Solar PV ~20% per doubling since 1976 (Swanson's law); lithium-ion ~18% since 1991. wright's law ↗ ↑
- Michael Polanyi, The Tacit Dimension, 1966 — "we can know more than we can tell." polanyi ↗ ↑
- Individual frontier-AI researcher compensation packages reported at and above $100M, 2025–. ai talent market ↗ ↑
- Frontier model training compute on the order of 10²⁵–10²⁶ FLOPs, mid-2020s. frontier compute ↗ ↑
- W. Brian Arthur, "Increasing Returns and the New World of Business," 1996; Increasing Returns and Path Dependence in the Economy, 1994. arthur ↗ ↑
What is the smiling curve?
A model drawn by Acer founder Stan Shih in 1992: across a hardware value chain, value added is high at the two ends — upstream IP and design, downstream brand, services, and ownership — and lowest in the middle, at manufacturing and assembly. Every commoditized hardware era since has obeyed it.
Why won't robot manufacturers capture the value of physical AI?
Because assembly sits in the middle of the curve, and the experiment has been run twice: IBM invented the PC and exited for $1.75B while Wintel took the profit pool; Nokia built the most phones on Earth and sold for ~$7.2B while Apple took 80%+ of industry profits on a minority of units. Robot bodies, built from EV and phone supply chains, are already on the same cost curve.
What happened to Nokia — and who is the Apple of physical AI?
Nokia mastered the body while the phone's value migrated into its mind — the OS, the silicon, the ecosystem. The Apple position in physical AI belongs to whoever owns the embodied mind and the fleet's experience, not whoever ships the most chassis. That seat is still open — which is the premise of ARBX.
Is making humanoid robots a good business?
Making them will be a magnificent, enormous, thin-margin business — the middle of the curve always is: the world's largest contract assembler runs on ~3% operating margins. Cheap bodies are not the industry's problem; they are its subsidy — every dollar the body sheds widens the market the mind can address.
What stays scarce in physical AI?
Four things: the minds (frontier embodied models and the lecture-hall-sized population who can train them, with packages reported in nine figures), the compute (frontier runs on the order of 10²⁶ FLOPs), the data (fleet experience — rivalrous and private), and the position (a mind with a body can, for the first time, be held).
What is Wright's Law?
Observed by T. P. Wright in 1936: every doubling of cumulative production cuts unit cost by a roughly fixed fraction — ~20% for solar panels since 1976, ~18% for lithium-ion cells since 1991. For young products like humanoids, doublings arrive extremely fast, so early price collapse is a schedule, not a surprise.
What is the law of conservation of attractive profits?
Clayton Christensen's observation that when one layer of a technology stack becomes modular and good enough, profits don't disappear — they migrate to adjacent layers where hard, integrated problems remain. PC bodies modularized and profit moved to Wintel; handset bodies modularized and profit moved to Apple's silicon-OS stack; robot bodies are modularizing now, and the adjacent hard layers are the mind and the fleet.
What is Polanyi's paradox?
Michael Polanyi, 1966: "we can know more than we can tell" — most human skill is tacit and cannot be written down. It explains why the physical world resisted automation for seventy years (you cannot program what no one can articulate), and why the engineers who finally cracked embodied learning are so scarce: their craft is tacit too.
What is ARBX's position on the curve?
Own the ends, rent the middle. ARBX builds minds and holds fleets — the two ends where substitution fails — and buys bodies as commodity hardware. We do not make bodies; we make what makes bodies valuable, and we keep what the bodies learn.