·Business & Strategy
Section 1
The Core Idea
Most business models suffer from diminishing returns. The first million dollars in revenue is cheap to acquire. The second costs more. The tenth costs dramatically more — because easy customers are gone, competitors have responded, and the market segment is saturated. This is the default condition of economic activity, and it's been the organizing assumption of economics since David Ricardo. But a specific class of businesses operates under the opposite dynamic: the more they do something, the better they get at doing it, and the better they get, the more the system rewards them for doing it. These are improving returns — and they are the most powerful force in modern business.
The mathematical expression is
Wright's Law. In 1936, Theodore Wright — an aeronautical engineer at Curtiss-Wright Corporation — published a paper documenting a pattern in aircraft manufacturing: every time cumulative production of a given aircraft model doubled, the per-unit labor cost dropped by a consistent percentage, typically 15–20%. The pattern was not a one-time efficiency gain. It was a curve — a predictable, continuous decline in cost as a function of cumulative experience. Wright observed it in airplane manufacturing, but the law turned out to be universal. Semiconductors follow it. Solar panels follow it. Batteries follow it. DNA sequencing follows it. Any production process that involves complex coordination, iterative learning, and cumulative refinement exhibits Wright's Law — cost declines as a power function of cumulative production volume.
The distinction from economies of scale matters enormously. Economies of scale are about size at a point in time — the cost advantage of producing a million units this year versus a thousand. Improving returns are about accumulated experience over time — the cost advantage of having produced ten million units across your entire history versus a competitor who has produced one million.
Scale can be purchased with capital. A well-funded competitor can build a factory as large as yours. Improving returns cannot be purchased — they must be earned through cumulative production, cumulative learning, and cumulative refinement of processes that embed knowledge into the system itself. The competitor who builds an identical factory on day one faces a cost curve that the incumbent already descended years ago.
Tesla's battery cost trajectory is the cleanest modern demonstration. In 2010, lithium-ion battery packs cost approximately $1,100 per kilowatt-hour. By 2023, the cost had fallen below $140/kWh — an 87% decline driven primarily by Wright's Law dynamics. Each doubling of cumulative battery production reduced costs by roughly 28%. The learning was embedded in everything: cathode chemistry optimization, cell manufacturing yield improvements, pack design simplification, supply chain negotiation leverage, and factory layout efficiency gained through repetition. A new entrant building batteries in 2024 does not start at $140/kWh. They start somewhere above it — because they lack the cumulative learning that Tesla, CATL, and BYD embedded into their processes through millions of units of production experience. The cost is not in the factory. It's in the history.
TSMC demonstrates improving returns in semiconductor fabrication. Each new process node — from 28nm to 16nm to 7nm to 5nm to 3nm — requires TSMC to solve manufacturing challenges that no competitor has solved at equivalent scale. The learning from each node transfers to the next: lithography techniques, defect reduction methods, yield optimization algorithms, and equipment calibration procedures all improve with cumulative production. Intel, with decades of fabrication experience but lower cumulative production at advanced nodes, has struggled to match TSMC's yields at 7nm and below — not because Intel lacks engineering talent but because TSMC has run more wafers through more advanced processes and accumulated more learning. The improving returns advantage compounds: TSMC's superior yields attract more customers, more customers fund more production, more production generates more learning, and more learning widens the yield gap.
Amazon's logistics network is improving returns applied to distribution. Every package shipped teaches Amazon's routing algorithms something about delivery optimization. Every warehouse processed teaches the robotics and layout teams something about throughput maximization. Every holiday peak teaches the capacity planning systems something about demand forecasting. By 2024, Amazon had shipped over fifteen billion packages cumulatively. That cumulative experience — embedded in algorithms, warehouse designs, driver routing software, and packaging optimization — represents a learning advantage that no competitor can shortcut by building warehouses. The warehouses are commodities. The accumulated intelligence within them is not.
Data flywheels represent the purest form of improving returns in the digital economy. Google Search processes over 8.5 billion queries per day. Each query — and each subsequent click, scroll, and refinement — teaches the algorithm something about relevance. The cumulative dataset of trillions of search interactions produces a search quality advantage that no competitor can replicate by building better algorithms alone, because the algorithms are only as good as the data that trains them, and the data is a function of cumulative usage. Bing can match Google's engineering talent. It cannot match Google's twenty-five years of cumulative query data. The improving return is not in the code. It is in the corpus of human behavior that the code has learned from — a corpus that grows with every query and can never be purchased, only accumulated.