Solow's model explains growth in output per worker with g and then declines to explain g, which arrives from outside the model. That is a large hole to leave in a theory of long-run growth, and closing it is what the endogenous-growth literature of the 1980s and 1990s did.
What does non-rival mean for growth?
The pivot is a property of ideas that Paul Romer made central in his 1990 paper: ideas are non-rival. A lathe used by one worker cannot simultaneously be used by another; a recipe, a chemical process, a compiler or a management method can be used by everyone at once without being consumed. Rival goods face diminishing returns because you have to keep making more of them. Non-rival goods do not.
Once ideas are non-rival and are produced deliberately by people who respond to incentives, three things change. Growth becomes a function of choices about research, education and openness rather than an exogenous constant. Scale can matter, because a larger market supports more research on a fixed idea. And intellectual property becomes a genuine trade-off rather than a nuisance, because the same monopoly rent that funds the research restricts the use of the non-rival output.
Advertisement
What is the AK model?
The limiting case is worth solving because it shows exactly what breaks. Set α = 1 so that capital's exponent is one and there are no diminishing returns at all. The production function collapses to Y = A·K, and the capital accumulation equation becomes
gY = s · A − δwith A now interpreted as broad productivity of a broad capital stock that includes knowledge and skills. There is no k* here, because the marginal product of capital never falls, so there is nothing to converge to. And the saving rate is back inside the growth rate.
With A = 0.30 and δ = 5%, raising s from 0.28 to 0.35 lifts growth permanently from 3.40% to 5.50%, a gain of 2.10 points that compounds to a 2.73-times larger economy over fifty years. In Solow the same policy bought a 12.77% level gain and no growth at all. Two models, one policy, opposite answers, and the difference is entirely whether α is 0.35 or 1.
Solow vs endogenous growth: which is right?
Which is right is an empirical question, and the honest answer is neither in pure form. Charles Jones showed in 1995 that if research effort set the growth rate, growth should have risen sharply as researcher numbers exploded, and it did not. Nicholas Bloom, Charles Jones, John Van Reenen and Michael Webb quantified the same point in 2020, finding that research productivity in the United States falls at roughly 5% a year across a wide range of activities.
Ideas are produced rather than given, which is Romer's contribution and it stands. They are also getting harder to find, which caps what any country can buy with research spending alone.
For an analyst the practical residue is a rule about forecast horizons. Over a decade, the capital term and the labour term dominate and are forecastable from investment rates and demography. Over half a century, the residual dominates and is not forecastable at all, which is why a terminal value should assume mediocrity in it rather than continuation of a good run.
{
"skill": "Building a growth-accounting decomposition and its capital stock from primary data",
"use": "Have an assistant write the perpetual-inventory loop and the decomposition from a downloaded national-accounts table, then have it produce the full sensitivity grid over the capital share, the initial capital stock and the depreciation rate, so that the point estimate never travels without its range.",
"tools": ["Chat assistants with code execution - Claude, ChatGPT, Gemini", "Python: pandas for the national-accounts table, NumPy for the recursion", "A spreadsheet for the side-by-side reconciliation of the two decomposition routes"],
"prompt": "Here is a table with columns year, real GDP, real gross fixed capital formation, and total hours worked. Write Python that (a) builds a real capital stock by the perpetual inventory method K[t+1] = (1-delta)K[t] + I[t] from a stated initial stock and depreciation rate, (b) computes the compound annual growth rate of output, capital and labour over the full window, (c) decomposes output growth into alpha*gK, (1-alpha)*gL and the residual, (d) asserts that the per-worker identity gY - gL = alpha*(gK - gL) + residual closes to within 1e-12, and (e) prints a grid of the residual over alpha in {0.30, 0.35, 0.40, 0.45, 0.50}, initial capital stock at -10%/base/+10%, and delta in {0.04, 0.05, 0.06}. Do not round intermediate values, and state which vintage of the source table the numbers came from.",
"verify": "Recompute the two end-point growth rates by hand from the levels before trusting any of it, because a single mis-parsed year silently changes every downstream number. Confirm the per-worker assertion actually fires by feeding it a deliberately wrong alpha. Trace every input series to the primary publisher and its release date, not to a secondary aggregator: real output and real gross fixed capital formation to the national statistics office, hours or employment to the labour-force survey, and note any series break inside your window because the residual will absorb it invisibly. Recompute the totals independently; Date-check the vintage against the window; Cite the table and release for each input (Guardrail 1-5). A decomposition without its sensitivity grid attached is not a result.",
"diy": "You must be able to run the decomposition and the per-worker cross-check on a four-input table with a calculator, and you must be able to say from the capital-output ratio alone whether the capital term can continue. Worked examples 1 and 2 are the drill. An assistant will happily hand you a residual computed on a headcount series when hours were available, and on a capital share it chose for you, and the output will look identical to a correct one.",
"market": "IN"
}