Governments logged over 2,500 industrial policy measures since 2023, and 71% distort trade. What the evidence says about when state money builds capacity.
The Return, Measured
How a discredited idea became the default setting of economic policy
Since 1 January 2023 the New Industrial Policy Observatory has logged more than 2,500 new industrial policy measures worldwide, 71% of them trade-distorting✓ Established [4]. The United States, the European Union and China account for 48% of that total between them✓ Established [4]. Industrial policy is no longer a heterodox position defended in development seminars; it is the operating assumption of the three largest economies on earth. The argument has moved from whether governments should do it to whether any of it works.
The instrument of choice is not the tariff. Across the whole period from 2009 to 2023, subsidies were consistently the main vehicle for industrial policy in both rich and developing countries: their share of measures fell from 84% to 75% in advanced economies and rose from 56% to 71% in emerging and developing ones✓ Established [3]. The same IMF work identifies a structural break around 2020, after which activity accelerates sharply and the stated motive shifts from competitiveness and climate toward supply-chain resilience and national security✓ Established [3]. What began as a climate programme has become a security programme wearing the same fiscal clothes.
That shift is measurable in the language of the measures themselves. By 2024 and 2025, more than half of American industrial policy interventions explicitly cited national security or geopolitical rationales, and the instrument mix moved away from domestic financial support toward tariffs, local-content requirements and security-justified trade defence◈ Strong Evidence [5]. The proclamation of 14 January 2026 imposing a 25% Section 232 tariff on a narrow band of advanced semiconductors and the products containing them belongs to that second phase rather than the first✓ Established [37]. Subsidy and protection are converging into a single policy posture.
The fiscal weight is easier to underestimate than to exaggerate. The World Bank's 2026 assessment finds that upper-middle-income countries now provide business subsidies averaging 4.2% of GDP, the highest figure on record and above what high-income economies spend✓ Established [6]. China spent 1.73% of GDP on industrial policy in 2019 on a deliberately conservative accounting, against 0.67% in South Korea and 0.39% in the United States; on broader assumptions that include government procurement, the Chinese number approaches 4.9%✓ Established [7]. The developing world is not following the rich world into industrial policy. It was already there.
In the United States the programme has a paper trail. Nineteen companies received $30.7 billion in direct awards and $5.5 billion in loans across 40 commercial fabrication projects under the CHIPS and Science Act, with roughly 59% of the award value directed at leading-edge logic and about 20% at leading-edge memory✓ Established [15]. Alongside it, the Inflation Reduction Act's manufacturing credits triggered an announcement wave that private analysts have been counting ever since, and unwinding since 2025✓ Established [19]. Two statutes, one country, and a combined commitment larger than the annual output of most European economies.
Everyone else responded in kind. The European Union assembled a Chips Act carrying about €86 billion in triggered funding✓ Established [16]. Japan has committed roughly ¥2.35 trillion of cumulative research support to a single start-up, Rapidus, with a further ¥631.5 billion scheduled for fiscal 2026-27✓ Established [24]. South Korea raised its semiconductor package to 33 trillion won, about $36 billion, from 26 trillion won a year earlier✓ Established [25]. India runs fourteen production-linked incentive schemes with an outlay of ₹1.91 lakh crore✓ Established [26]. None of these governments is experimenting. All of them are matching.
Subsidies accounted for 75% of advanced-economy industrial policy measures and 71% of emerging-market measures by 2023, having been the dominant instrument for the entire period since the financial crisis✓ Established [3]. The tariff revival is real but recent, concentrated in the United States after 2024, and layered on top of the subsidy architecture rather than replacing it◈ Strong Evidence [5]. Any assessment that treats the new industrial policy as a protectionist turn is measuring the smaller half.
This report asks the only question that matters about a policy this expensive: under what conditions does the money produce industrial capacity, and under what conditions does it produce announcements. The answer is not ideological. Over the past fifteen years, econometricians have built a body of properly identified evidence on industrial policy that did not exist when the Washington Consensus declared the question settled✓ Established [1]. That evidence is more favourable to industrial policy than the previous consensus allowed — and far more specific about why most of it fails.
What the Evidence Revolution Found
Causal identification changed the answer, but not to a simple yes
The older empirical case against industrial policy rested on cross-country correlations that could not separate the policy from the conditions that produced it✓ Established [1]. Since roughly 2015 a new literature using natural experiments, administrative microdata and structural models has revisited the canonical episodes. Its verdict is more favourable than the previous orthodoxy, and considerably more demanding: effects depend on whether the instrument matches the market failure, and on whether the state can withdraw support◈ Strong Evidence [1].
The methodological problem was never subtle. Governments target industries that are already growing, already politically organised, or already strategically important, which means the correlation between support and success carries no causal content in either direction. The review by Réka Juhász, Nathan Lane and Dani Rodrik is explicit that the earlier generation of work was largely correlational and marred by interpretational problems, and that the newer papers — attentive to measurement, identification and economic structure — offer a more positive and more contextual account✓ Established [1]. This is a change in the evidence, not a change in ideology.
The best-identified episode is South Korea's Heavy and Chemical Industry drive of 1973 to 1979, launched for defence reasons after the partial withdrawal of American troops. Nathan Lane's study, published in the Quarterly Journal of Economics in 2025, finds that the policy promoted the expansion and dynamic comparative advantage of the industries it targeted, that it indirectly benefited downstream users of the targeted intermediates through the input-output network, and that the benefits persisted after the drive itself ended✓ Established [8]. The last of those three findings is the one that matters: the effect outlived the money.
Jaedo Choi and Andrei Levchenko put numbers on the persistence using the universe of firm-level subsidy records. Firms receiving the average subsidy recorded 919% larger sales growth between 1982 and 2009 than firms that received none — an 8.6% higher annual growth rate, measured entirely after the subsidies had stopped✓ Established [10]. Their structural model attributes the persistence to learning-by-doing and relaxed financial constraints, and estimates that Korean welfare would have been 22% to 31% lower had the drive never happened◈ Strong Evidence [10]. Between half and two-thirds of that gain comes from the long-run productivity channel rather than the direct transfer◈ Strong Evidence [10].
Subsidised firms outgrew never-subsidised firms for thirty years after the Heavy and Chemical Industry drive was wound up, with the advantage transmitted through learning-by-doing and relieved financial frictions rather than through continuing transfers◈ Strong Evidence [9]. Upstream suppliers of the subsidised firms benefited as well◈ Strong Evidence [9]. This is the strongest single piece of causal evidence that industrial policy can change an economy's trajectory rather than merely its accounting.
The rich-country evidence is more sobering. Chiara Criscuolo, Ralf Martin, Henry Overman and John Van Reenen exploited seven-yearly changes in European state aid eligibility to identify the effects of Britain's Regional Selective Assistance programme across two decades of firm data. A ten-percentage-point increase in the maximum investment subsidy raised manufacturing employment by 10%, with positive effects on investment and net entry✓ Established [12]. But there was no effect on total factor productivity, and the entire treatment effect came from small firms — large companies took the money without changing behaviour✓ Established [12].
Allocation turns out to matter more than amount. Philippe Aghion and co-authors examined every medium and large Chinese enterprise between 1998 and 2007 and found that subsidies, tax holidays, loans and tariffs raised productivity growth when they were directed at competitive sectors or spread across many firms, and failed when they were concentrated on incumbents✓ Established [13]. The mechanism is competition: support that preserves rivalry between recipients performs measurably better than support that creates a national champion◈ Strong Evidence [13]. That result cuts directly against the instinct most governments follow.
The emerging picture, generally, paints a more positive view of industrial policy — but also highlights important nuances.
— Réka Juhász, Nathan Lane and Dani Rodrik, CEPR VoxEU, 2024Instrument design is the third axis. Panle Jia Barwick, Myrto Kalouptsidi and Nahim Bin Zahur reconstructed China's shipbuilding push through a dynamic model of entry, exit, investment and production. The policy worked on its own terms — the Chinese share of world orders climbed from 14% in 2003 to 53% by 2009, while Japan's fell from 32% to 10% and South Korea's from 42% to 32%✓ Established [14]. But entry and investment subsidies produced fragmentation and idle yards, while production subsidies delivered far more capacity per dollar◈ Strong Evidence [14]. The same budget, spent differently, would have bought substantially more.
What the new literature does not establish is equally important. It does not show that industrial policy raises aggregate growth on average, because the studies are of episodes rather than of countries. It does not show that any particular government can identify a winning sector in advance. And it does not resolve the selection problem in the literature itself: the well-identified episodes are disproportionately the famous ones◈ Strong Evidence [1]. The honest summary is that industrial policy can work, that its success is a function of design rather than of ambition, and that the design features which predict success are precisely the ones most political systems find hardest to supply.
The Korean Template and the Latin American Ruin
Two continents ran the same experiment with opposite discipline
Between 1950 and 1980 Latin America and East Asia both protected infant industries, both directed credit, and both built state-sponsored heavy industry. Latin American GDP per head went from 27% of the American level in 1950 to about 29% in 1980✓ Established [32]. Korea's went from among the poorest in the world to OECD membership. The instruments were similar; the accountability was not.
Import-substituting industrialisation was not a fringe doctrine. It was the mainstream development prescription of the post-war decades, implemented across Brazil, Argentina, Mexico and Chile with tariffs, quotas, directed credit and state enterprises. Its measurable failure was specific: the region combined strong capital deepening with weak total factor productivity growth, meaning that investment went in and output did not come out at the expected rate◈ Strong Evidence [32]. Three decades of protection moved relative income by two percentage points✓ Established [32].
The diagnosis has held up. Protected firms served domestic markets, never reached the scale at which learning economies operate, and never faced the discipline of foreign competition◈ Strong Evidence [32]. Because the state held discretionary power over who received protection, the profitable activity shifted from raising productivity to lobbying for continued protection◈ Strong Evidence [32]. The subsidy became an entitlement, and the sectors that received it acquired precisely the political weight needed to prevent its withdrawal.
East Asia used the same instruments with an added condition: export performance. Rents were granted against measurable results in foreign markets, which meant the recipient faced a price signal the domestic government could not fake and could not politically override. Mazzucato and Rodrik argue that rents provided through industrial policy must carry either performance requirements or close monitoring of use, because without a stick to discipline the recipient, support degenerates into a handout secured by lobbying◈ Strong Evidence [36]. Export discipline is the cheapest such stick ever invented, because a foreign buyer cannot be lobbied.
The distinction the evidence supports is not between picking winners and staying out. It is between regimes that can stop paying and regimes that cannot. The test of an industrial policy is not whether the state identifies the right sector in advance, which no state reliably does, but whether it can withdraw support from the sectors it got wrong◈ Strong Evidence [36]. Selection is a forecasting problem and governments are bad at forecasting. Exit is an institutional problem, and institutions can be designed.
Even Korea paid for weak exit. Minho Kim, Munseob Lee and Yongseok Shin examined the drive at plant level and found that while plant-level productivity rose in the targeted sectors, industry-region productivity did not, because allocative efficiency deteriorated as resources concentrated✓ Established [11]. Had misallocation within the targeted industries not worsened relative to the untargeted ones, their average total factor productivity would have been 40% higher by 1980◈ Strong Evidence [11]. The canonical success story contains a 40% efficiency loss inside it.
The comparison between East Asia and Latin America is usually read as evidence that some governments are cleverer than others. The evidence supports a narrower and more useful claim. Both regions chose sectors badly at times; only one attached a measurable, externally verified performance test to the money and enforced it◈ Strong Evidence [36]. Export discipline works not because bureaucrats in Seoul understood steel, but because a Japanese shipbuyer could not be persuaded by a Korean lobbyist.
This is where transferability becomes the real question. Korea's discipline rested on an authoritarian state with an insulated economic bureaucracy, a small number of family conglomerates that could be monitored individually, and an external security threat that made industrial failure politically intolerable. The World Bank's 2026 review is blunt that past efforts failed principally on implementation capacity and on fiscal and institutional constraints rather than on the choice of sector✓ Established [6]. A democracy with a lobbying industry and a four-year electoral cycle is attempting the same trick with none of the same instruments.
The Semiconductor Test Case
Four jurisdictions, one technology, four different answers
No other sector has absorbed as much industrial policy money in as short a time as advanced semiconductors, and none offers as clean a natural experiment. The United States, the European Union, Japan and South Korea launched broadly similar programmes within three years of one another, with different instruments, different conditionality and different institutional capacity. Four years in, the outcomes have already diverged✓ Established [15].
The American programme is the largest and the best documented. Under the CHIPS and Science Act, nineteen companies received $30.7 billion in direct awards and $5.5 billion in loans across 40 commercial fabrication projects, with about 59% of award value going to leading-edge logic, 20% to leading-edge memory and 10% to mature nodes✓ Established [15]. A further twelve companies signed preliminary agreements without reaching a final award✓ Established [15]. Announced private commitments across TSMC, Intel, Micron and Samsung run into the hundreds of billions of dollars, and American fabrication capacity is expanding at about 5% a year to 3.2 million wafers per month◈ Strong Evidence [31].
The schedules tell a different story from the headlines. Negotiated completion dates for CHIPS-funded projects run from November 2024 all the way to October 2033✓ Established [15]. Intel's Ohio site slipped from 2026 to 2030; Samsung's Texas production moved from 2024 to 2025; TSMC pushed initial Arizona production back by at least a year, citing uncertainty over the award itself; Micron's New York timeline extended into the third quarter of 2030 on labour shortages✓ Established [15]. A programme justified by the urgency of supply-chain risk is delivering on a schedule measured in decades.
Then the instrument changed character entirely. In August 2025 the United States government converted Intel's outstanding CHIPS awards into equity, purchasing 433.3 million shares at $20.47 for a 9.9% stake, funded by $5.7 billion of unpaid grant money and $3.2 billion from the Secure Enclave programme✓ Established [17]. The holding was structured as passive, with no board representation or governance rights✓ Established [17]. A conditional manufacturing subsidy became an unconditional ownership position in a single firm, without the performance requirements the evidence identifies as the operative variable◈ Strong Evidence [36].
The August 2025 agreement exchanged $5.7 billion of unpaid CHIPS awards and $3.2 billion of Secure Enclave funding for 433.3 million Intel shares at $20.47, a 9.9% stake held without board representation or information rights✓ Established [17]. Whatever its merits as a financial trade, it removes the mechanism — milestone-linked disbursement — that the causal literature identifies as the difference between Korean and Latin American outcomes◈ Strong Evidence [36].
Europe's programme failed on arithmetic before it failed on execution. The EU Chips Act triggered roughly €86 billion in funding against a target of 20% of global semiconductor manufacturing by 2030 — a target that would require quadrupling European production capacity✓ Established [16]. The European Commission's own July 2024 forecast puts the EU share of the global value chain at 11.7% in 2030, up from 9.8% in 2022✓ Established [16]. Over the same window, the leading global manufacturers budgeted €405 billion, nearly five times the European commitment✓ Established [16]. The European Court of Auditors concluded that the target was not reachable at the current rate of progress✓ Established [16].
The 20 percent target was essentially aspirational — meeting it would require us to approximately quadruple our production capacity by 2030, but we're nowhere close to that with our current rate of progress.
— Annemie Turtelboom, Member of the European Court of Auditors, April 2025Japan chose concentration over breadth. Cumulative government research support for Rapidus, a start-up founded in 2022 with no production history, has reached about ¥2.35 trillion, with ¥631.5 billion scheduled for fiscal 2026-27 and roughly ¥300 billion for the year after✓ Established [24]. The company is targeting mass production of 2nm logic at its Chitose plant in Hokkaido from fiscal 2027✓ Established [23]. It still needs to raise around ¥1 trillion in private equity and more than ¥2 trillion in private financing◈ Strong Evidence [24]. This is the highest-variance bet any advanced economy has placed on a single industrial firm in a generation.
South Korea, which has the manufacturing base the others are trying to build, spends to defend rather than to acquire: 33 trillion won in the latest package, up from 26 trillion won the year before✓ Established [25]. The underlying distribution has barely moved. Chinese capacity grew 14% to 10.1 million wafers per month, close to a third of the world total; Taiwan holds 5.8 million and South Korea 5.4 million, against 3.2 million in the Americas✓ Established [31]. Five economies — China, Taiwan, Korea, Japan and the United States — hold nearly 90% of global capacity✓ Established [31].
American manufacturing construction ran at a $172.7 billion annual rate in June 2026, down 21.4% in a year, and computer, electronic and electrical construction is down 44% from its July 2024 peak◈ Strong Evidence [18]. The electronics category had accounted for over half of all manufacturing construction at the peak◈ Strong Evidence [18]. Strip it out and the rest of manufacturing construction grew 5.6% — real, but not the reindustrialisation the programme was sold as◈ Strong Evidence [18].
The final turn is the most revealing. Having spent four years subsidising domestic fabrication, the United States began in January 2026 to tariff the imports that competed with it, imposing a 25% Section 232 duty on a narrow set of advanced semiconductors and the goods containing them, while exempting most legacy chips and those destined for large domestic data centres✓ Established [37]. That sequence — subsidise, then protect the subsidised asset — is the precise pattern that turned Latin American infant industries into permanent dependants◈ Strong Evidence [32].
The Chinese Counterfactual
What a patient state buys, and what it destroys
Made in China 2025 is the largest industrial policy programme ever attempted and the only one with a published scorecard. Rhodium Group's assessment finds China achieved 86% of its goals, far exceeding its benchmarks in new energy vehicles and electrical equipment while meeting only 75% of new-materials targets, the weakest of the ten sectors✓ Established [20]. The programme is simultaneously the strongest evidence that industrial policy works and the clearest demonstration of what it costs.
The pattern of hits and misses is diagnostic. China closed the gap fastest in sectors where the binding constraint was capital, scale and process engineering — batteries, electric vehicles, electrical equipment, solar, shipbuilding — and slowest in sectors where the binding constraint was accumulated basic research: high-end semiconductors, advanced aerospace, biomedicine and new materials◈ Strong Evidence [20]. New materials came last precisely because it depends on breakthrough discovery rather than deployment◈ Strong Evidence [20]. Money buys capacity reliably; it buys frontier science slowly.
Shipbuilding is the cleanest quantitative case. Between 2006 and 2013 Beijing directed free coastal land, financing assistance for buyers, and entry and production support into the sector, and the Chinese share of world orders rose from 14% in 2003 to 53% by 2009 while Japan collapsed from 32% to 10% and South Korea slipped from 42% to 32%✓ Established [14]. The academic reconstruction is unambiguous that the policy worked and equally unambiguous that it was expensive: entry and investment subsidies created fragmentation and idle yards that production subsidies would have avoided◈ Strong Evidence [14].
The vehicle sector shows the scale of firm-level support. The Kiel Institute calculated that BYD received at least €3.4 billion in direct state aid, with annual support rising from €220 million in 2020 to €2.1 billion in 2022✓ Established [21]. That excludes subsidised input pricing on steel and batteries, purchase rebates paid to its customers, and discriminatory public procurement — all of which are real and none of which is easily measured◈ Strong Evidence [21]. Chinese industrial subsidies run at three to four times OECD levels on conservative estimates and up to nine times on broader ones◈ Strong Evidence [21].
Aggregate estimates converge on the same conclusion from a different direction. On CSIS's deliberately conservative accounting China spent 1.73% of GDP on industrial policy in 2019, more than double South Korea's 0.67% and more than four times the American 0.39%✓ Established [7]. Widening the definition to include government procurement pushes the Chinese figure toward 4.9% of GDP◈ Strong Evidence [7]. Whatever the precise number, the order of magnitude is not in dispute, and neither is the fact that the rest of the world is now trying to close a gap that took two decades to open.
Rhodium's scorecard shows benchmarks far exceeded in new energy vehicles and electrical equipment and a 75% completion rate in new materials, the weakest of the ten sectors, with high-end semiconductors, advanced aerospace and biomedicine remaining the persistent gaps✓ Established [20]. The failures cluster in exactly the domains where the constraint is basic research rather than deployment capital◈ Strong Evidence [20]. State money is a capacity instrument, not a discovery instrument.
The cost is now visible on Chinese balance sheets. Listed solar companies lost an estimated $7.3 billion in 2025, the top six lost a further $2.8 billion in a single quarter, and manufacturers announced another $1.5 billion of losses in the first quarter of 2026 — roughly three continuous years of unprofitability in an industry that supplies over 80% of the world's panel components✓ Established [22]. Three years of official anti-involution policy have not fixed it, because local governments keep loss-making plants alive to protect employment◈ Strong Evidence [22]. The subsidy that built the industry now prevents it from consolidating.
What China's record establishes is narrow and important. A state with sustained fiscal capacity, a twenty-year horizon and tolerance for enormous waste can relocate global manufacturing capacity in specific sectors✓ Established [20]. What it does not establish is that this was cheap, that the returns exceeded the outlay, or that it can be replicated by governments that must show results within an electoral cycle. The programme's greatest advantage was not the money. It was the absence of any requirement to stop.
The Anatomy of a Boondoggle
Where the money actually goes when it fails
Failures in industrial policy do not usually look like a bad sector choice. They look like a well-chosen sector, a headline investment figure, a subsidy paid against promises rather than performance, and a quiet renegotiation three years later. Wisconsin, Sweden and Saxony-Anhalt each produced that sequence within a decade, at a combined public cost in the tens of billions◈ Strong Evidence [28].
Wisconsin's Foxconn package is the canonical American case. The state offered roughly $4.5 billion, mostly in direct cash payments plus land acquired through eminent domain, against a promise of 13,000 jobs — implying $200,000 to $346,000 of public money per job, seven to twelve times the average American state incentive✓ Established [28]. Foxconn hired 113 of the 260 employees required in 2018 and 281 of the 520 required in 2019, then stopped trying◈ Strong Evidence [28]. The state's non-partisan Legislative Fiscal Bureau put break-even at 2043 in the best case✓ Established [28].
Europe's flagship battery champion failed in a different register. Northvolt raised more than $14 billion, positioned itself as the EU's answer to Asian cell manufacturing, and filed for bankruptcy in Sweden on 12 March 2025✓ Established [27]. Its Skellefteå plant reached about 1 GWh of annual output against a 16 GWh plan, a shortfall that led BMW to cancel a $2 billion supply contract✓ Established [27]. The failure was not sectoral — European battery demand is real — but operational, in the specific discipline of scaling cell production, which no amount of capital substitutes for.
Germany's Intel commitment failed before a single wafer moved. Berlin had committed €10 billion to the Magdeburg plant, the largest single state subsidy in German history, of which €3.96 billion was scheduled for 2024✓ Established [29]. Intel suspended the project in September 2024 and abandoned it in July 2025, leaving the federal government to reclaim funds and reopen the argument about where the money should go✓ Established [29]. The state had underwritten a firm's capital plan without any mechanism to compel the plan.
| Failure mode | Severity | Assessment |
|---|---|---|
| Political capture and inability to exit | The recipients of industrial policy acquire the political weight needed to prevent withdrawal, which converts a temporary subsidy into a permanent entitlement and shifts firm effort from productivity to lobbying◈ Strong Evidence [32]. This is the single failure mode common to every unsuccessful episode in the record◈ Strong Evidence [36]. | |
| Policy reversal and time inconsistency | Since the start of 2025, $39.6 billion of planned American clean-energy manufacturing investment and 53,400 announced jobs have been cancelled✓ Established [19], and the 2025 budget law is projected to cut new clean power build-out by 53% to 59% over the following decade◈ Strong Evidence [30]. Capital will not commit to fifteen-year assets on four-year policy. | |
| Overcapacity and the subsidy race | Chinese solar manufacturers have run at a loss for roughly three years, with $7.3 billion of combined losses in 2025 alone✓ Established [22], while the IMF warns that managing cross-border spillovers requires cooperation precisely when geopolitics makes it least available◈ Strong Evidence [35]. | |
| Cost per job and measurement failure | Subsidies are routinely priced against announced rather than delivered employment: Wisconsin's package implied $200,000 to $346,000 per promised Foxconn job against a delivered headcount a fraction of that✓ Established [28]. Announcement-based accounting systematically overstates returns. | |
| Implementation and institutional capacity | The World Bank attributes past failures principally to weak implementation capacity and fiscal and institutional constraints rather than to poor sector selection✓ Established [6], a diagnosis the IMF repeats for the current wave◈ Strong Evidence [35]. |
Reversal is now the dominant risk in the United States, and it is self-inflicted. Companies have cancelled $39.6 billion of planned clean-energy manufacturing investment and 53,400 announced jobs since the start of 2025, with the cancellations concentrated in electric vehicles, hydrogen electrolysers and solar manufacturing✓ Established [19]. In the second quarter of 2026 alone, 5,900 new manufacturing jobs were announced against 2,800 cancelled✓ Established [19]. Rhodium estimates the 2025 budget law will cut new clean power capacity build-out by 53% to 59% across 2025 to 2035◈ Strong Evidence [30].
The construction data show what that does to physical investment. American manufacturing construction ran at a $172.7 billion annual rate in June 2026, down 21.4% year on year, with computer, electronic and electrical construction down 44% from its July 2024 peak◈ Strong Evidence [18]. That category had represented more than half of all manufacturing construction spending at the top◈ Strong Evidence [18]. Excluding electronics, construction spending rose 5.6% since the tariffs began — a real effect, and a much smaller one than the political framing◈ Strong Evidence [18].
Industrial policy asks private capital to commit to assets with fifteen- to thirty-year payback periods on the strength of statutes that survive at most one change of government. When those statutes are amended — as they were in 2025, cancelling $39.6 billion of planned investment✓ Established [19] — the state does not merely lose the projects. It raises the risk premium on every future programme it announces, which means the next round has to be larger to buy the same behaviour.
The failure modes have a common shape. In each case the subsidy was paid, or committed, against an announcement rather than a delivered milestone; there was no mechanism to compel performance; and the political cost of cancellation fell on the government rather than the recipient. This is the mirror image of the export discipline that separated East Asian from Latin American outcomes◈ Strong Evidence [36]. The states that failed were not worse at choosing sectors. They were worse at attaching conditions and enforcing them.
The Argument That Will Not Resolve
Four genuine disagreements, and where each one actually sits
Some of the dispute over industrial policy is ideological noise. Four parts of it are real, in the sense that competent researchers looking at the same data reach opposing conclusions⚖ Contested [1]. Separating the genuine disagreements from the performative ones is the difference between a policy debate and a culture war.
The first genuine dispute is over the Korean canon. Choi and Levchenko estimate that Korean welfare would have been 22% to 31% lower without the Heavy and Chemical Industry drive◈ Strong Evidence [10], while Kim, Lee and Shin find that average total factor productivity in the targeted industries would have been 40% higher in 1980 had misallocation not worsened⚖ Contested [11]. Both are careful, both use Korean microdata, and both are probably right: the drive raised the level of industrial capability and simultaneously degraded the efficiency with which resources were allocated within it.
The second is over attribution in China. Made in China 2025 hit 86% of its targets✓ Established [20], but China also had the world's largest domestic market, the deepest manufacturing supply chains, a decade of rising engineering graduates and an exchange-rate position that would have supported export growth without any programme at all. The Aghion results suggest the composition of support mattered — dispersed, competition-preserving subsidies raised productivity where concentrated ones did not✓ Established [13] — which is evidence that the design did work, without settling how much of the outcome the design explains⚖ Contested [13].
The third is over whether the CHIPS Act has succeeded. Capacity is genuinely being built and American wafer output is rising◈ Strong Evidence [31]. Completion dates run to 2033, the flagship Ohio project slipped four years✓ Established [15], and manufacturing construction in the relevant category has fallen 44% from its peak◈ Strong Evidence [18]. Whether this reads as a success with delays or a delay with successes depends almost entirely on the counterfactual chosen, and no honest analyst can observe that counterfactual⚖ Contested [15].
The case that it works
Properly identified studies of infant-industry promotion find support for increased activity in targeted sectors in every episode reviewed, a marked change from the correlational literature✓ Established [2].
Korean firms receiving the average subsidy grew sales 919% more between 1982 and 2009 than firms that received none, three decades after the programme ended✓ Established [10].
China went from 14% to 53% of world shipbuilding in six years, displacing two incumbent industrial powers✓ Established [14].
Made in China 2025 achieved 86% of its stated goals across ten sectors, exceeding its benchmarks in new energy vehicles and electrical equipment✓ Established [20].
A ten-percentage-point rise in the maximum British investment subsidy raised manufacturing employment by 10%, with gains in investment and net entry✓ Established [12].
The case against
The same British programme produced no measurable effect on total factor productivity, and the entire employment effect came from small firms while large ones took the money and did not change✓ Established [12].
Targeted Korean industries would have had 40% higher average productivity in 1980 without the deterioration in allocative efficiency the drive produced◈ Strong Evidence [11].
Entry and investment subsidies in Chinese shipbuilding produced fragmentation and idle capacity that production subsidies would have avoided at the same cost◈ Strong Evidence [14].
Chinese solar manufacturers lost $7.3 billion in 2025 after three years of price war, and local governments keep loss-making plants alive to protect jobs✓ Established [22].
$39.6 billion of American clean-energy manufacturing investment and 53,400 jobs were cancelled within three years of the statute that induced them✓ Established [19].
The fourth dispute is the one with the largest stakes. If every major economy subsidises the same technologies simultaneously, the result may be faster decarbonisation at lower global cost, or an expensive arms race in which each country pays to relocate capacity that already existed elsewhere. The IMF's position is that managing these spillovers requires international cooperation at precisely the moment when geopolitical tension makes cooperation hardest◈ Strong Evidence [35]. Both readings are consistent with the data available in 2026⚖ Contested [35].
What is not contested is narrower than the volume of argument suggests. Nobody serious now claims that industrial policy never works; the Korean and Chinese microdata foreclosed that position✓ Established [1]. Nobody serious claims that governments can reliably identify winning sectors in advance either. And nobody disputes the two design findings that survive every study: support allocated in ways that preserve competition outperforms support that creates champions✓ Established [13], and support without an enforced exit condition degenerates into rent◈ Strong Evidence [36].
The debate persists because those findings are politically unwelcome from both directions. They deny the free-market position that state money is inherently wasted, and they deny the interventionist position that the main obstacle is insufficient ambition. What the evidence actually indicts is the specific institutional habit shared by almost every democratic government running an industrial policy today: paying against announcements, declining to specify a failure condition, and treating the disbursement as the outcome.
What the Evidence Tells Us
Design is the variable; ambition is not
After fifteen years of causal work on industrial policy, the findings that replicate are not about which sectors to choose. They are about how the money is attached, whether competition survives the intervention, and whether the state retains the capacity to stop◈ Strong Evidence [1]. Every large programme currently running can be scored against those criteria, and most score badly.
Five design features predict success in the literature, and they are consistent across very different settings. Support must be conditional on verifiable performance rather than on announcement◈ Strong Evidence [36]. It must be allocated in a way that preserves rivalry between recipients rather than creating a champion✓ Established [13]. The instrument must match the market failure: production subsidies where the constraint is scale, research funding where the constraint is discovery◈ Strong Evidence [14]. There must be an enforced exit condition◈ Strong Evidence [36]. And there must be an administrative apparatus capable of measuring all four✓ Established [6].
The competition finding is the most robust and the least followed. Chinese industrial policy raised productivity growth when it was dispersed across firms within a sector and failed when it was concentrated on incumbents✓ Established [13]. Yet every major current programme concentrates: American CHIPS awards went overwhelmingly to leading-edge logic at a handful of firms✓ Established [15], the Japanese state has placed ¥2.35 trillion on a single company✓ Established [24], and Germany committed €10 billion to one plant belonging to one manufacturer✓ Established [29]. Concentration is administratively convenient and empirically the worse choice.
The exit finding is the one democracies find hardest. Export discipline worked in East Asia because it delegated the verdict to a party the recipient could not lobby◈ Strong Evidence [36]. Nothing in a modern advanced-economy programme performs that function: milestone conditions exist on paper but are renegotiated when the recipient is large enough, and equity conversion — as with the American position in Intel — removes the conditionality entirely✓ Established [17]. The problem is not that governments do not know what conditionality is. It is that they cannot credibly commit to enforcing it against a firm employing 20,000 voters.
Matching the instrument to the failure is the finding with the clearest fiscal implications. Where the constraint is scale and learning, production-linked support outperforms entry and investment subsidies at the same cost, as the shipbuilding reconstruction shows directly◈ Strong Evidence [14]. Where the constraint is basic discovery, capital deployment is close to useless — which is why China met its deployment targets and missed on new materials, high-end semiconductors and biomedicine◈ Strong Evidence [20]. India's production-linked design, which pays against output rather than investment, at least has the structure right, whatever its execution: ₹28,748 crore disbursed against ₹2.16 lakh crore of investment as of 31 December 2025✓ Established [26].
Capacity is the constraint nobody wants to name. The World Bank's 2026 review attributes past failure principally to weak implementation capacity and to fiscal and institutional constraints rather than to poor sector selection✓ Established [6], and the IMF's 2026 assessment repeats the finding for the current wave: the policy is adapting to crises but remains hard to implement effectively◈ Strong Evidence [35]. A government that cannot audit a milestone cannot run a conditional subsidy, and a government that cannot run a conditional subsidy is not doing industrial policy. It is doing procurement with extra steps.
The evidence does not support the claim that industrial policy is a mistake, and it does not support the claim that more of it is better. It supports a third position that neither political camp finds convenient: industrial policy is a state-capacity test administered with public money✓ Established [6]. Governments that can verify performance, preserve competition and stop paying get the Korean result◈ Strong Evidence [10]. Governments that cannot get the Wisconsin one✓ Established [28].
The next five years will settle more of this than the previous fifty did, because for the first time several large economies are running comparable programmes simultaneously with published data. The measurable questions are already specified: whether American fabrication capacity reaches the level implied by the awards before the 2033 completion dates✓ Established [15], whether the European share of the semiconductor value chain exceeds the 11.7% the Commission itself forecasts✓ Established [16], whether Rapidus reaches volume production at 2nm◈ Strong Evidence [23], and whether any of these governments proves willing to stop paying a firm that misses its milestones. The last question is the only one that has never been answered in the affirmative by a democracy, and it is the one the evidence says matters most◈ Strong Evidence [36].