What is industrial policy's actual track record — CHIPS Act, IRA, and EU equivalents: what's working and what isn't
When Governments Try to Build Factories: What's Actually Working?
Based on analysis of a 139-node, 474-edge knowledge graph mapping industrial policy outcomes across the US CHIPS Act, Inflation Reduction Act, EU industrial programs, and their global equivalents.
The Basic Question
Governments around the world are spending hundreds of billions of dollars trying to build industries they think are strategically important — computer chips, electric vehicle batteries, solar panels. The United States passed two big laws (the CHIPS Act and the Inflation Reduction Act). The European Union launched its own versions. China has been doing this for decades.
Is it working? The honest answer is: some of it is working, some of it is backfiring, and some of the things that are working right now are setting up problems for later.
Here is what the structure of the evidence actually shows.
The IRA’s Clever Trick — and Why It Creates Its Own Problems
The Inflation Reduction Act (the IRA, passed in 2022) used an unusual design. Instead of giving companies grants — here is a check, now go build something — it offered tax credits. If you build a solar panel factory, you get money back on your taxes for every panel you make. No cap on how much you could claim.
Think of it like a promise: “For every lemonade you sell, we’ll pay you back 30 cents.” That’s very different from “Here’s $1,000 to start a lemonade stand.”
The tax credit design was chosen partly because of a painful memory. In 2011, a solar company called Solyndra received a government loan and then went bankrupt. The political backlash was so severe that grant-based approaches became nearly impossible to defend. So designers chose tax credits instead — which private companies love because they can collect them at scale without government approval for each project.
This worked. Private investment in clean energy manufacturing flooded in. But the design had a hidden consequence: because there was no cap, the projected cost kept rising. The original estimate was around $383 billion over ten years. Updated estimates reached $1.1 trillion. That fiscal exposure gave political opponents the ammunition they needed. In 2025, a budget bill (called OBBBA) began rolling back parts of the IRA.
The graph shows this as a loop: the IRA’s success at attracting investment generated the cost overruns that enabled its partial dismantling. The very thing that made it work — unlimited, fast, private-capital-friendly — made it politically vulnerable. The mechanism is self-undermining when it succeeds too well.
The Thing You Can’t Buy With Money
Across the entire map of industrial policy evidence, one concept keeps appearing as the explanation for why things succeed or fail: tacit knowledge.
Tacit knowledge is the kind of knowledge that lives in people’s hands and minds — not in textbooks or manuals. It is what a skilled baker knows about when bread dough is ready. It is what an experienced semiconductor engineer knows about why a batch of chips failed that morning. You cannot learn it by reading about it. You accumulate it by doing it, for years, alongside people who have already done it.
This turns out to be the hidden variable in almost every industrial policy story.
Why does Taiwan make the world’s most advanced computer chips? Not primarily because of subsidies. Taiwan built tacit knowledge in semiconductor manufacturing over 30 years, starting in the 1970s through a government research institute called ITRI that trained engineers who later spread through the industry. The knowledge accumulated in the workforce.
Why is China so dominant in battery manufacturing? Again, not just subsidies. Chinese battery companies have been building cells at scale for longer than Western competitors, accumulating process knowledge about yield, defect rates, and cost reduction that cannot be purchased directly.
The CHIPS Act is spending roughly $52 billion to build semiconductor fabs in the United States. The graph shows this running into the tacit knowledge wall. TSMC, the Taiwanese company building factories in Arizona, has partially addressed this by flying experienced Taiwanese engineers to Arizona to train American workers — transferring knowledge through people, not documents. Other programs that are trying to build capability from scratch face a harder problem: you cannot shortcut the accumulation timeline with money.
The clearest statement of this in the evidence: the knowledge gap between what TSMC can do in Taiwan versus what it can do in Arizona is closing, but it is closing at the speed that experience accumulates in workers, not at the speed that capital is deployed.
Who Is Actually Getting What They Paid For?
There is a design problem called the “additionality problem.” It sounds technical but the idea is simple: if a company was going to build a factory anyway, and the government pays it to build the factory, did the public get anything for its money?
Industrial policy designers worry about this a lot. The evidence in the graph suggests no one has fully solved it.
India tried an interesting approach with its Production Linked Incentive programs: the government only pays subsidies on production that exceeds a baseline. If you were already making 100,000 phones a month, you only receive support for the phones you make above that number. This structurally eliminates the “you would have done it anyway” problem.
The CHIPS Act tried to address it with milestone requirements — companies only receive funds when they hit specific production targets. The IRA tax credits tie payments to actual units produced, which provides some link between subsidy and output.
But no program in the evidence has fully eliminated the concern. The closest thing to a theoretical escape: even if a company would have built the factory without the subsidy, the public still benefits if the factory’s existence drives down costs for everyone through accumulated learning. (Economists call this the “learning curve” — each doubling of production volume tends to reduce costs by 15-20%.) Whether this spillover benefit is large enough to justify the subsidy cost is a genuinely open empirical question.
A Legal Ruling That Cannot Be Enforced
Here is a structurally strange situation in the evidence. The World Trade Organization — the international body that is supposed to regulate when countries unfairly subsidize their own industries — issued a ruling in 2026 that part of the IRA’s domestic content requirements are illegal under international trade rules. The ruling says the tax bonus for using American-made components is WTO-inconsistent.
At the same time, the WTO’s appeals court (called the Appellate Body) has been effectively non-functional since 2019 because the United States blocked new appointments to it.
So the graph records an active legal ruling with no enforcement mechanism. Countries can violate WTO rules with limited consequences right now because the enforcement infrastructure is broken. This is not incidental — essentially every major industrial policy program in the evidence depends on this enforcement vacuum. The IRA, the EU’s state aid programs, China’s subsidies — all would face legal challenges in a fully functioning WTO system. They are operating in a period of legal-but-unenforced status.
How One Bankruptcy in 2011 Shaped a $1 Trillion Program
One of the less obvious connections in the evidence: the design of the IRA — which will cost over a trillion dollars — was significantly shaped by a solar panel company going bankrupt fifteen years earlier.
Solyndra received a government loan guarantee in 2009 and failed in 2011. The reason it failed was largely that Chinese manufacturers flooded the market with cheap solar panels, undercutting Solyndra’s cost structure. But the political memory attached to it was “government picked a loser.” Grant-based industrial policy became politically toxic.
So when designers structured the IRA, they chose tax credits over grants specifically to avoid resembling Solyndra. Tax credits are automatic — the government is not picking specific winners, it is setting a price and letting markets decide who wins.
The fiscal cost explosion that followed — the difference between $383 billion and $1.1 trillion in projected expenditure — flows partly from that design choice. Which flows from that one bankruptcy. Which was actually caused by a Chinese trade strategy, not a government failure to pick good companies.
Why Europe Keeps Falling Further Behind
The European Union’s industrial policy situation has a structural problem that policy ambition cannot fix: the EU is a union of member states that each run their own fiscal policies.
When the US passes a tax credit, it applies in all 50 states immediately. When the EU announces an industrial program, it has to be implemented by 27 different countries with different budgets, different rules, and different political priorities. Larger, richer countries (Germany, France) can afford bigger national subsidies. Smaller countries cannot compete. This fragments the internal market rather than unifying it.
The graph shows this producing a feedback loop: the global arms race (countries competing via subsidies) makes the EU’s relative position worse, which increases the diagnosed urgency for action, which the EU’s structure cannot address quickly, which makes the position worse.
Meanwhile, the largest near-term industrial policy instrument the EU actually has is a budget exception for defense spending — not a civilian manufacturing program. Defense spending has expanded significantly; the chip and clean energy programs lag.
What the Evidence Consistently Shows Works
Across all the cases, the graph identifies a recurring pattern in programs that succeeded versus those that failed. The clearest version: programs that set explicit exit conditions — requiring companies to hit targets or lose support — show better outcomes than programs that provide support regardless of performance.
South Korea’s semiconductor industry in the 1980s is the most cited example. The government supported specific companies (Samsung) in specific technologies (DRAM memory chips), with clear market performance requirements. When the companies succeeded in global markets, support continued. The companies that failed were allowed to fail. The government did not protect them from market outcomes.
This is what one framework in the graph calls “letting losers go” — the willingness to withdraw support from underperforming programs is, paradoxically, part of what makes industrial policy work. Governments that support industries regardless of performance tend to create industries that survive only because of support.
Bottom Line
The structural picture that emerges from the evidence is roughly this:
What makes industrial policy work: transferring or accumulating tacit knowledge (not just building facilities), having explicit conditions for continued support, and using mechanisms that leverage private capital quickly.
What makes it fail: protecting industries from the market feedback that drives improvement, subsidizing companies that would have invested anyway without requiring additional output, and building single facilities when what drives competitive advantage is geographic clusters of interconnected expertise.
The central tension: the policy mechanism that best achieves speed and scale (uncapped tax credits, as in the IRA) generates the fiscal exposure and political vulnerability that enables its own dismantling. The mechanisms that are more accountable (milestone grants, performance-linked payments) deploy more slowly and mobilize less private capital.
The deepest constraint: money can build factories. Money cannot, on its own, build the decades of accumulated operational knowledge that make factories competitive. Every major industrial policy program in the evidence runs into this limit — and the evidence consistently shows that the only way through it is workforce transfer or time.