Thomas Simon · September 2026

The Efficiency Flywheel

Making a thing more efficient makes it cheaper. Cheaper means more of it gets used, which drives the cost down again. The same turn that grows the market drains the margin out of whoever sits in the middle.

Cumulative volume Unit cost falls Price falls Quantity demanded rises Is demand elastic? Wright's law unit cost falls ~20% per doubling of volume price follows cost unless a held part keeps the gap as margin efficiency reaches the buyer yes, elastic quantity rises more than price fell Jevons more units built no, inelastic the saving sticks; rebound below 100%; the loop stalls R REINFORCING LOOP Parts standardize Suppliers multiply and compete Margin drains from the middle volume concentrates on the same parts Stigler the market is big enough to specialize one spec, many makers Baldwin and Clark margin pools at the ends design, integration, brand (the smile curve) and in what the curve can't touch (Baumol)

Red marks the three assumptions the loop rests on: a real learning curve, elastic demand, and a standard part.

  • The loop only closes through the elasticity check, the red diamond at the bottom. Volume drives unit cost down by Wright's law, and price follows.
  • Elastic demand turns the wheel. The cheaper price pulls in more than proportional demand, which adds volume and drives cost down again.
  • Inelastic demand stops it. The saving is kept and the wheel stalls: rebound without backfire.
  • The chain on the right runs off the same volume. Standardized parts, many competing makers, and the margin leaving the middle for the ends.
  • Cheaper parts feed back into unit cost. That is why the two sides are one curve.

The three conditions

A learning curve that is real

Cost has to fall with cumulative volume, not just with time. Airframes, chips, solar modules, radios: yes. A string quartet: no.

Elastic demand

A price cut has to bring a more than proportional rise in quantity demanded. There has to be a darker room, a longer drive, a bigger model to run.

A standard part

For the supply side to commoditize, volume has to concentrate on one interchangeable component with a fixed interface, so that many makers can build it.

Each assumption sits on one arrow of the loop:

  • The learning curve decides whether more volume lowers cost.
  • The standard part decides whether lower cost becomes a lower price, or stays as margin with whoever controls the part.
  • Elastic demand decides whether a lower price brings back more volume.

Where each assumption holds, and where it fails

A real learning curve

Cost falls with volume

  • Solar modules, about 20% cheaper per doubling
  • Lithium batteries, close to that
  • Memory chips
  • Airframes, where Wright measured it
  • The ESP32 radio at a dollar

Cost does not fall with volume

  • Nuclear plants: French and US costs rose with each build, the documented case of negative learning
  • Live performance and care work, where an hour is an hour
  • Site-built construction: every building a prototype, productivity flat for decades

Elastic demand

Use grows faster than price falls

  • Light for two centuries: spending flat, lumens up a hundred-thousand-fold
  • Computation: every hardware gain spent
  • Road capacity: driving fills it
  • Radios cheap enough to put in a lightbulb

Use barely grows as price falls

  • Refrigerators: energy per unit fell about three-quarters since the 1970s, and nobody bought ten
  • Home insulation and car fuel economy: rebound of 10 to 30%, most of the saving kept
  • Cheaper is not enough; there has to be a darker room

A standard part

Many makers can build it

  • Open-source hardware, where price falls fastest: the design given away on purpose
  • The shipping container
  • USB
  • The ESP32
  • Generic drugs: the day a patent expires the same molecule may be made by anyone, and the price falls 80 to 90%

Few makers can build it

  • Patented drugs: one maker by law
  • Certified aerospace parts: traceability limits who may build them
  • Wedgwood's Etruria: no potter knew the whole process, so the learning could not leave
  • Fixtures built to a specification: nothing interchangeable, so the middle keeps its margin. Machinery versus Furniture in the industry return tables

The last failure case is the one that matters for a specification-driven manufacturer: the same curve that drains margin from a commodity part leaves it with whoever owns the specification.

Where the moat sits: the standard part

The assumption a firm controls

Of the three assumptions, the standard part is the one a maker controls. It sits on the arrow from cost to price. Open-source hardware gives that control away on purpose, which is why price falls fastest there and why the originators get cloned.

Five ways to hold it

  • Law: a patent
  • Regulation: certification
  • Organization: secrecy by splitting the process
  • Interface: a proprietary socket
  • Ecosystem: the customer relationship

Wedgwood, 1770

Split the work so no potter knew the whole process: learning sped up at each step and could not walk out the door. Then cut prices from his cost books and opened the mass market. Held the part, passed the saving on.

Tesla, 2014

Gave away its patents to pry open supply and demand at once: more makers, more volume, a faster curve it rode too. Then opened its charging connector and it became the standard. Opened the part, kept the margin on the socket.

Control of the part is not the moat by itself. It is what makes the next decision yours: whether the cost decline goes to price, to grow volume, or stays as the gap between cost and price.

  • Held and kept: the patented drug.
  • Held and passed: Wedgwood.
  • Open and passed: open-source hardware, where the middle earns nothing.
  • Open and kept: Tesla, Arduino, Red Hat. Give away the module, own the socket.

An example: light, 1800 to today

NumberWhat it is
0.7%of GDP spent on light, roughly constant for two centuries (Tsao and colleagues, 2010)
100,000×more light consumed over the same period, as the cost of a lumen collapsed (Nordhaus, 1996)
~20%cost decline per doubling of cumulative production, the experience curve (Wright, 1936; Henderson, 1968)
10 to 30%rebound for most household efficiency gains; full backfire is rare. Light is the exception (Sorrell, 2007; Gillingham and others, 2016)

Reading the loop

The R marks a reinforcing loop, the systems-dynamics term for what a business calls a flywheel: each turn makes the next turn easier. Jevons's paradox is the name for what this loop does to a resource when demand is elastic. The elasticity check is the whole argument: efficiency saves less than it promises, and in the cases where demand is truly elastic it saves nothing, because the saving is spent on more of the thing.

The right-hand chain is what the same volume does to the people who make the parts. It is Apple's supply chain becoming Shenzhen's, and it is why a small hardware designer's advantage is recognizing a cost curve someone else paid for and moving the commodity to a place where it is not one.

Sources