Endogenous Growth Theory

Why Solow needed an alternative

The Solow model has a famous limitation. Per-capita growth is pinned to an exogenous rate of technical progress. Whatever fraction of GDP a country saves, raising it produces a one-time level effect — output per worker rises, but the long-run growth rate doesn't change. The only thing that drives long-run growth is the parameter g, and g is handed in from outside the model.

That was unsatisfying for two reasons. First, the cross-country data of the 1960s through 1980s showed striking divergence: South Korea pulled away from Ghana even though both started poor. The conditional convergence Solow predicted only emerged after Mankiw, Romer, and Weil (1992) added human capital. Second, Solow's accounting exercise had attributed roughly seven-eighths of U.S. per-worker growth to the residual A — and labelling something as "exogenous technical progress" wasn't an explanation.

Endogenous growth theory tries to explain the residual. Where does technology come from? Why do some countries produce more of it than others? What policies make innovation more productive? Paul Romer's 1990 paper "Endogenous Technological Change" is the foundational answer.

The trick — knowledge as a non-rival good

Goods are rival if my using them prevents you from using them: a steel beam in my building is not in yours. Goods are non-rival if any number of people can use them at once without diminishing each other's use: a mathematical theorem, a chemical formula, a software algorithm. Romer's key observation was that ideas — the inputs to technological progress — are non-rival.

This breaks Solow's diminishing returns. In the Solow setup, the marginal product of capital falls as you accumulate more, because each new factory produces less than the last. But knowledge is different. If a country accumulates a stock of useful ideas A, every worker in the economy can use every idea simultaneously. Doubling A doesn't halve the value of each idea — they all keep their full productivity. With non-rival inputs, returns to scale at the economy level can exceed one, even when no individual firm gets increasing returns.

That single observation rebuilds growth theory. Output is now a function of capital, labor, and a knowledge stock A that grows because firms and people invest in producing it.

Romer 1990 — the canonical model

Romer's 1990 paper writes the economy with three sectors. The final-goods sector produces consumer output Y using labor and a continuum of intermediate inputs:

Y = LY1−α · ∫0A x(i)α di

Here A is the variety of intermediate goods (the stock of designs), x(i) is the quantity of input i, and L_Y is labor in production. The intermediate-goods sector is monopolistically competitive: each firm holds a patent on one design and charges a markup above marginal cost.

The R&D sector produces new designs at a rate proportional to research labor and the existing stock of knowledge:

Ȧ = δ · LA · A

The crucial term is the multiplication by A. New ideas come from researchers standing on the shoulders of existing ideas. This is what generates the perpetual growth: the more we know, the more efficiently we produce new knowledge.

In equilibrium, the long-run growth rate of output per worker is g = δ · L_A. Doubling research labor doubles long-run growth. R&D subsidies, immigration of researchers, and trade openness (which expands the effective L_A) all permanently raise the growth rate. Solow's exogenous g has been replaced by an expression in policy variables.

Lucas 1988 — human-capital externalities

Robert Lucas's 1988 paper "On the Mechanics of Economic Development" took a parallel route. Instead of explicit R&D, growth comes from human-capital accumulation with external effects: when I become a better engineer, my productivity rises, but so does the productivity of the engineers I work with, because skilled colleagues raise everyone's output. The externality breaks diminishing returns to human capital and sustains perpetual growth.

Lucas used this to argue why poor countries don't simply borrow rich-country technology and catch up. Workers in countries with thin human-capital stocks can't absorb advanced technologies as productively, even when the blueprint is freely available. The model gives a formal version of the everyday observation that you can't run a chip fab in a country without engineers.

The AK shortcut

The simplest endogenous growth model strips Romer down to one equation:

Y = A·K

where K is interpreted broadly to include physical capital, human capital, and knowledge. Marginal product of capital is constant at A — no diminishing returns. With a constant savings rate s and depreciation δ:

K̇/K = s·A − δ
gY = s·A − δ

Growth is permanent and depends linearly on the savings rate. That's a clean prediction. Empirically it's too clean — countries with similar savings rates show different growth, and post-1990 China's savings rate didn't predict its growth path the way the AK model would. But it's the cleanest pedagogical example of how breaking diminishing returns generates endogenous growth.

Endogenous growth vs Solow vs other theories

	Solow (1956)	AK model	Romer (1990)	Lucas (1988)	Schumpeterian (Aghion-Howitt 1992)	Semi-endogenous (Jones 1995)
Source of long-run growth	Exogenous A	Constant marginal product of K	R&D producing new designs	Human-capital externalities	Quality-improving innovations	Population growth
Diminishing returns	Yes, in K	No	Avoided via non-rival ideas	Avoided via spillovers	Replaced by Schumpeterian creative destruction	Yes; growth from population
Convergence	Conditional	None — divergence possible	None automatically	None	Conditional	Yes
Policy raises long-run growth?	No (level only)	Yes	Yes — R&D, IP, trade	Yes — education spending	Yes — innovation incentives	Limited
Empirical fit	Decent for OECD	Poor cross-country	Good for divergence patterns	Mixed	Mixed	Best for "ideas getting harder"
Scale effect	None	None	Yes — more researchers ⇒ faster growth	Yes	Yes	None — addressed Romer's bug

Evidence in favor

The empirical case for endogenous growth rests on three patterns the Solow model can't explain.

• R&D pays off. Coe and Helpman (1995) found that a 1% increase in domestic R&D capital raises productivity by 0.23% in large countries — a return well above the discount rate. Trade-weighted foreign R&D spillovers added another 0.06%, supporting the non-rival channel.
• Patenting and growth co-move. Cross-country panels show patent counts per capita track productivity growth tightly, even after controlling for capital and labor inputs.
• Education matters. The marginal social return to schooling — once you account for spillovers between workers — is consistently estimated above the marginal private return, consistent with Lucas's externality story.

Counterarguments

The scale-effect problem. Romer's first-generation model implies that if you double the number of researchers worldwide, you double the long-run growth rate. The number of full-time-equivalent researchers in OECD economies has grown roughly 20-fold since 1950, but TFP growth has not accelerated 20-fold — it has slowed. Charles Jones (1995) made this critique cleanly and proposed semi-endogenous models where ideas get harder to find as the stock grows, so growth ultimately depends on population growth, not research effort.

The replication argument. Robert Solow himself argued that it's enough to imagine running the U.S. economy twice — at double the size — to get double the output. That's constant returns to scale. If knowledge is truly non-rival and increasing returns are pervasive, doubling the economy should more than double output, which the data don't show clearly.

Augmented Solow as competitor. Mankiw, Romer, and Weil (1992) — note the same Romer — showed that adding human capital to the standard Solow model fits cross-country income variation roughly as well as endogenous growth without abandoning the convergence prediction. Many macroeconomists treat MRW as a sufficient extension of Solow without endorsing endogenous growth's stronger claims.

The "ideas getting harder" finding. Bloom, Jones, Van Reenen, and Webb (2020) document that across many sectors, the number of researchers needed to sustain a given rate of productivity growth has been rising for decades — Moore's Law now requires 18× as many semiconductor researchers as in the 1970s. Their conclusion: ideas are getting harder to find, eroding the optimistic side of endogenous growth.

Variants

• R&D-driven (Romer 1990). Knowledge produced in a dedicated research sector; growth proportional to research labor.
• Human-capital (Lucas 1988). Skills and externalities drive growth; no formal R&D sector.
• Schumpeterian (Aghion–Howitt 1992). Each new innovation displaces the previous one (creative destruction); growth from quality ladders rather than variety expansion.
• Semi-endogenous (Jones 1995). Removes the scale effect; long-run growth driven by population growth times research productivity.
• Directed technical change (Acemoglu). Innovation is biased toward factors that are abundant or whose markets are large — explains skill-biased technical change.
• AK (Rebelo 1991). Pedagogical limit case where capital is broadly defined and faces no diminishing returns.

Common pitfalls

• Treating "endogenous" as a synonym for "complicated." The defining feature is that growth is determined inside the model. The simplest endogenous growth model — AK — is one equation.
• Confusing endogenous growth with R&D growth. Lucas's model is endogenous growth without R&D. The unifying feature is the absence of diminishing returns to a producible factor, not the presence of a research sector.
• Inferring policy from first-generation models. Romer 1990 implies infinite returns to research effort, which is unrealistic. Use Jones-style semi-endogenous models for policy comparisons.
• Ignoring the scale-effect critique. Any policy claim resting on "more researchers" without a Jones-style correction is fragile.
• Reading divergence as evidence for endogenous growth. Augmented Solow generates divergence too if savings rates and human-capital investments differ persistently. The presence of divergence isn't decisive.