In 2021, economists Arvind Subramanian, Justin Sandefur, and Dev Patel announced that poor countries had finally started catching up to rich ones, vindicating the Solow growth model's prediction of "convergence." Now they say the rumors of Solow convergence's vindication were premature; the convergence trend reversed and poor countries are falling behind again.

David Oks, writing about their retraction, argues that the whole convergence episode was a mirage produced by the Chinese commodities boom. China's industrial buildout created massive demand for raw materials. Poor countries that exported copper, soybeans, iron, and oil experienced a surge of income. When Chinese demand slowed in the mid-2010s, their growth collapsed.

Oks is probably right about the proximate cause, but the convergence debate is asking a malformed question. The way economists measure convergence doesn't correspond to the abstractions Solow uses to justify the intuitions behind his model. And in the rare case where economists have bothered to measure a quantity relevant to the Solow growth model, they find convergence after all.

What Solow actually predicts

The Solow growth model says that capital investment has diminishing returns. Your first production line adds a lot, because its outputs are badly needed inputs to other productive processes. Your thousandth adds little; what are you going to do with your hundredth wrench, use it as a paperweight? So economies closer to their production-possibilities frontier grow slower, and economies further from it grow faster. Poor countries should catch up to rich ones.

This prediction is specifically about produced means of production: factories, machines, infrastructure, tools. This includes intangible productive infrastructure like software, logistics systems, and organizational know-how, though intangibles are easier to compromise with job-creation grift than a crane or a power grid. The convergence prediction comes from diminishing returns to the producible stuff, not from anything about land or natural resources.

The convergence test doesn't measure what Solow predicts

Solow predicts convergence in capital per worker. But the standard empirical test, which Subramanian, Sandefur, and Patel use, doesn't measure capital at all. It regresses GDP growth on initial GDP per capita. If poorer countries grow faster, that's "convergence."

This substitution involves several steps, each of which introduces distortion. Solow's production function maps capital and effective labor to output, the result of combining productive inputs. This matters because in the Solow model, output is where capital comes from: output that can be reinvested as input to further production is the thing that converges. GDP mixes this with rents, imputations, and institutional overhead that don't correspond to output from any production function.¹ Even if GDP tracked output, the test substitutes income levels for capital stocks, relying on a stable production function to make income a reliable proxy for capital. And the test assumes technology levels are similar across countries, absorbing the differences into an error term. Mankiw, Romer, and Weil's foundational 1992 paper, which set the template for the whole literature (it has over 20,000 citations), regresses income per capita on saving rates and population growth without ever directly measuring capital stocks.

GDP doesn't measure productive capacity

GDP in rich countries is substantially composed of activity that doesn't represent productive capacity in any straightforward sense.

Housing appreciation shows up as imputed (or actual) rent in GDP. Credential inflation shows up as education spending. The expansion of administrative and compliance labor in healthcare, finance, and government shows up as service-sector output.

In The Domestic Product, I trace the historical process by which this happened. The conditions under which something like GDP corresponds to an improvement in human capacities are those in which the economic activity it represents is a complex and progressive solution to a set of problems (or, equivalently, exploitation of a set of opportunities). When this economic activity remains unironic, it tends to obviate itself by solving the problems that made it necessary, so if we see secular, smooth GDP growth like we've seen in the past century, we need to either assume some consistent supply of new problems and opportunities of comparable value to the ones being solved, or that GDP has become an optimization target decoupled from the process of problem-solving it originally measured.²

A system that requires people to work longer hours to afford housing, creates jobs to absorb those hours, and counts both the housing appreciation and the job output towards GDP growth, will have higher per-capita GDP than a system with the same underlying capacities, in which people work less and live cheaply. But it is not richer in any sense relevant to the Solow model's production-possibilities frontier.

Does measuring capital directly help?

The convergence test uses GDP as a proxy for productive capacity, and as we've seen, it's a bad proxy. But the underlying Solow prediction is about capital per worker, so you might think the fix is to test convergence using capital measures instead. This runs into its own problems at every level of measurement.

Start with what people ordinarily mean by "capital" or "wealth." Total US household net worth is north of $150 trillion, with housing as the dominant component for most people. Davis and Palumbo (2006) estimated that land accounts for about 50% of US residential real estate value, up from 32% in 1984, and the trend has continued since. Gianni La Cava's research found that the increase in capital's share of national income is largely attributable to housing appreciation. So a substantial part of what makes rich countries look "capital-rich," in the ordinary sense, is expensive land, not an abundance of productive tools approaching diminishing returns. This is obviously not what Solow meant by "capital."

Economists know this, which is why the formal capital stock estimates used in growth accounting try to exclude land. The widely used Penn World Tables and IMF Investment and Capital Stock Dataset rely on the perpetual inventory method: they cumulate monetary investment flows (gross fixed capital formation) and apply depreciation rates. Land purchases are excluded from these flows by construction; national accounts have treated land as a non-produced asset since Kuznets. But the convergence literature mostly doesn't use these measures either.

McQuinn and Whelan (2006) are a rare exception, and their results look very different from the mainstream. Working from the Central Bank of Ireland, they constructed capital stock series from PWT investment data using the perpetual inventory method, then examined how fast the capital-output ratio (still using GDP for output) converges to its steady-state value. This is closer to what Solow actually predicts: the model's endogenous dynamics are entirely in the capital-output ratio, not in output alone. They found convergence at about 7% per year, roughly what the Solow model predicts, and well above the 1-2% the GDP-based regressions report. Their explanation for the discrepancy is that stochastic technology shocks bias the output-per-worker regressions downward.

And even the formal measures inherit distortions. If investment spending is driven by land-value-inflated construction costs rather than productive need, the "capital stock" goes up without productive capacity increasing. The perpetual inventory method cumulates monetary flows and therefore inherits all the distortions of the monetary system they're denominated in.

John S. Wentworth's 2017 survey of nonfinancial capital assets across 102 major US public companies gives a useful picture of what real produced capital looks like when you strip away financial claims. It's overwhelmingly infrastructure and productive equipment: oil wells, power grids, telecom networks, railroads, manufacturing equipment, buildings. The grand total across these 102 companies was about $6.3 trillion, against that $150+ trillion in household net worth.

Wentworth's data also reveals that public corporations systematically minimize their holdings of non-productive assets like land. As I noted in the comments on his post, public corporations make ownership decisions close to the finance-theoretical ideal, minimizing assets that aren't part of their production function and increasing return on capital. People who want to hold claims on rents buy them separately. This is consistent with the model I advance in The Domestic Product: rentiers hold real estate, as they did when the term was coined, and everyone else operates within a system that conflates land rents with productive returns to inflate our long-run estimate of the latter.

In the published academic literature, the closest analogue to Wentworth's bottom-up approach is Calderón, Moral-Benito, and Servén (2015), who estimated infrastructure's output contribution across 88 countries using physical measures (road kilometers, power generation capacity, telephone lines) rather than monetary values. This approach sidesteps the contamination problem in monetary capital measures; I'm not aware of anyone who has applied it to the convergence debate.

What was the "Great Convergence"?

If rich-country GDP is inflated by accounting theater, then the measured gap between rich and poor countries is not tracking real productive capacity. The convergence test is regressing one distorted number against another.

On the rich-country side, a growing share of GDP reflects land rents, credentialism, and administrative bloat rather than productive output. On the poor-country side, the "convergence" period saw GDP temporarily inflated by commodity extraction that often left little durable capacity behind. In some cases the real gap may be smaller than measured (where rich-country GDP is mostly theater), in others larger (where poor-country GDP was inflated by resource liquidation). But since "rich" is defined by the distorted measure, the gap is likely smaller on average than it appears, and in some cases even negative.

As I understand it, here's what concretely constituted the "Great Convergence" of 1995-2015: China mobilized Chinese people to build out massive industrial capacity, which created transient demand for other countries, with weaker mobilization capacities, to mobilize their people to extract and sell raw inputs to China. This showed up as GDP growth in poor countries. But converting natural endowments into cash flows is not the sort of capital accumulation we should expect to lead to Solow-model capacity growth.

The divergence that preceded convergence also looks different under this lens. From roughly 1970-1995, a growing fraction of the apparent rich-poor gap was rich countries getting better at the nominal-value game, not rich countries accumulating more real productive capacity while poor countries fell behind.

The main difference between the (partial) hollowing out of the USA and the hollowing out of China's raw goods suppliers was that the US specialized in producing and exporting illusions of value (brand rents, financial claims, asset-price appreciation),³ and as the world's money-creator (and thus export market) of last resort, the US retains too-big-to-fail negotiating leverage. Like they say, "if you owe the bank a million dollars, you have a problem; if you owe the bank a hundred billion dollars, the bank has a problem."⁴⁵

Nonspecific Industrial Policy Is Zero-Sum

Oks and Dani Rodrik note that the commodities boom led to "premature deindustrialization" in poor countries. The standard framing invokes Dutch disease: the commodities boom strengthened local currencies, making manufacturing exports less competitive, so manufacturing capacity withered. This framing makes it sound like something that just happened to people, but it's actually the result of someone's decisions.

Chinese state-backed manufacturers, operating with subsidized credit, cheap energy, and export-contingent tax breaks, flooded African and Latin American markets with finished goods that directly displaced local producers. South Africa's manufacturing-to-GDP ratio fell from 24% in 1990 to 13% today. Sub-Saharan Africa's share of manufacturing value added in GDP dropped from over 16% to about 10% over the same period.

Comparative advantage implies that in a common market, when one participant uses policy to become relatively better at producing manufactured goods, everyone else would profit by adjusting away from manufacturing. China's choice to industrialize, through tariffs, subsidies, and state-directed credit, was structurally a choice to deindustrialize its trading partners.⁶ In principle, trading partners can integrate upward in the supply chain rather than being pushed out of manufacturing entirely; Vietnam has done something like this, using Chinese demand and inputs as a platform for its own export-oriented industrialization, expanding manufacturing capacity by over 100% in the last decade. But this requires exactly the kind of directed industrial policy that Western-imposed trade liberalization under the Washington Consensus stripped from most developing countries. The policy tools (tariffs, import quotas, industrial policy) that every successful industrializer, including China itself, had used to protect infant manufacturing were taken away. So most of these countries left the "convergence" period with less manufacturing depth, and less capacity to build on, than the GDP numbers suggested.

Resource extraction and the question of feedback

Resource extraction apparatus (e.g. mines, wells, processing facilities) are not categorically excluded from Solow-type capital accumulation. In principle, the surplus purchasing power from resource extraction could be reinvested into durable productive infrastructure that outlasts the boom: power grids, transport networks, manufacturing equipment, education. If a country uses its resource windfall to build an integrated productive system, the extraction apparatus feeds into genuine capital accumulation even though the mine itself depreciates and the ore runs out.

Saudi Arabia's Vision 2030 is an attempt to do exactly this, using oil wealth to build non-oil productive capacity before the resource runs out or loses value. The results so far are instructive: ten development plans over several decades have targeted diversification, and oil still accounts for 43% of GDP and 75% of government revenue. Non-oil GDP growth has been strong recently (4.5% in 2024), but non-oil exports remain at 25% of GDP against a 35% target. The Kingdom has the foresight, the sovereign wealth fund ($941 billion in assets), and the political will.

What it doesn't have is a straightforward way to convert money into productive capacity. In practice, much of the Public Investment Fund's (PIF) portfolio consists not of factories but of financial bets on foreign companies (Uber, Electronic Arts) and prestige acquisitions (Newcastle United Football Club, Heathrow Airport). The $45 billion SoftBank Vision Fund, the PIF's flagship technology bet, sank roughly $16 billion, about 16% of the fund's total capital, into WeWork, a company whose business was subleasing office space (often owned by and leased from the founder personally) at a loss: a real estate intermediary dressed up as a technology platform. The fund posted a record $32 billion loss in a single fiscal year. The number of active factories in Saudi Arabia has grown (from about 7,200 in 2016 to 12,000 in 2024), but the gap between non-oil GDP targets and outcomes suggests that buying equity in foreign companies is not the same thing as building domestic productive capacity.

Capacity isn't one thing you can buy. It emerges when people face real problems and develop techniques and tools to solve them. Building it requires some way of relating means to ends. There are a few known options:

• Explicit central planning: the Soviets tried to replace market prices with input-output matrices plus a specified desired consumption function. When the formal problem proved too complex to solve centrally, they supplemented it with a whole ecosystem of informal workarounds: fixers, input hoarding, grey markets, and ad hoc negotiation between enterprises. Central planning works better when it targets a specific problem. China's investment in solar manufacturing, for instance, addresses a concrete strategic vulnerability (oil import dependence) and produces something of global value, which is why it's positive-sum rather than merely redistributive.
• Existential military pressure: under Stalin, "Germany is invading" supplied the feedback signal. The penalty for failure (factories that don't produce, capacities built in the wrong proportions, logistics that can't move the right outputs to where they're needed as inputs) is death.
• Export discipline: South Korea, Taiwan, and Japan successfully industrialized in the postwar period by using a specific intermediate strategy described in Joe Studwell's How Asia Works: state-directed credit into manufacturing, with export competition as the ruthless external test. Neither pure planning nor pure markets did the job; what worked was directed investment combined with a feedback mechanism that killed firms that couldn't solve real production problems.⁷

• Domestic market competition can also work, but as I noted earlier, it tends to solve the problems that call for it, producing abundance that erodes the scarcity the financial system was built to manage. This forces a choice between letting the system wither away and cooking the books to create a perceived need for continued growth. (See, again, The Domestic Product.)

Saudi Arabia can buy factories, but factories are not capacity; capacity is the web of skills, habits, supplier relationships, and problem-solving repertoires that makes the factory produce things people want at costs they'll pay.⁸

The problem for most commodity-exporting countries in Africa and Latin America during the Great Convergence was that the resource windfall was not systematically reinvested into durable productive capacity. It flowed into consumption, real estate, government payrolls, and debt service.

This is sometimes called a "resource curse," but that framing attributes misaligned incentives to inanimate objects. Oil, for instance, was never a curse to the oil-rich countries, until highly capitalized countries' demand for it awarded extractive elites a tremendous amount of purchasing power in higher-capacity economies. It's the demand-resource match, not the natural resource itself, that constitutes the curse.

During these China-driven commodity booms, the extraction apparatus depreciated fast, it didn't generalize to other uses, and the flood of Chinese manufactured imports was simultaneously eroding the incentive to maintain whatever manufacturing base might have absorbed the surplus into broader production. The feedback loop from extraction into general capacity never closed.

Nonspecific "capacity" isn't something you can try to build. You have to want something specific enough that you can tell whether you're getting it, what Eliezer Yudkowsky called having "something to protect." A factory matters when it is for something. A number going up matters when it measures something we care about.

Footnotes

1. In fairness, this conflation is baked in from the beginning. Solow's 1956 model defines Y as output from a production function, but he immediately identifies it with national accounts data in his empirical work. MRW follow suit without comment. The theory is about productive output; the empirics have always substituted GDP. ↩

2. Technological miracles that outpace extractive growth can more than offset this, but either the benefit is transient, or the uncontained explosive growth of one aspect of the system constitutes a regime change. ↩

3. American multinationals like McDonald's, Coca-Cola, and Disney have succeeded in large part by making some of their products into money-like fetishes: part of the value of a Happy Meal to my toddler, or a glass bottle of Mexican Coca-Cola to my discerning friends, or merchandise with Mickey Mouse on it, is that it's widely depicted as intrinsically desirable. The Fallout series alludes to this nicely by imagining standardized bottle caps as a post-apocalyptic currency. ↩

4. In The Debtors' Revolt, I describe how debtor-aristocratic classes have repeatedly used correlated leverage and wartime mobilization to expropriate from creditor-bourgeois classes, from medieval Italian city-states through the World Wars, ultimately overturning the prewar creditor-dominant regime and replacing it with the debtor-dominant too-big-to-fail financial system. ↩

5. The typical exit strategy for individual commodity extractors in poor countries is to purchase some sort of financial asset backed by American buzz and live off the dividends or capital appreciation. ↩

6. This is more than just static comparative advantage. Krugman's strategic trade theory from the 1980s showed that in industries with increasing returns and learning-by-doing, protection can shift capabilities in ways that persist after the protection is removed. The effects compound. ↩

7. David Oks pointed out to me that Park Chung Hee openly planned South Korea's industrialization as a defense measure against the North, and that one could make similar arguments for China under Deng (survival of the regime) and for Taiwan (invasion from the mainland). ↩

8. SoftBank's 30-year vision deck, pitched to Saudi Crown Prince Mohammed bin Salman (MBS) among others, identified his actual problem—human disconnection—and suggested an appropriate general means to address it: information technology. But it couldn't supply the right details to help MBS get friends. The solution to loneliness isn't purchasable at scale, which is exactly the problem with trying to buy capacity. ↩