In 2024, China produced roughly 93% of the world’s polysilicon, 97% of its silicon wafers, 92% of its solar cells, and 86% of its finished modules.1 Those aren’t leads. Those are shares that would trigger an antitrust investigation in any other industry on earth.
And in the first quarter of 2026, the twenty-two largest listed Chinese solar manufacturers lost about $1.5 billion between them. For Tongwei, LONGi, and TCL Zhonghuan, it was the tenth consecutive quarter in the red.2 Forty-plus smaller firms have gone bankrupt or simply left. The top five cut their workforces by more than 30% in a single year.
Total dominance. Negative profit. For years.
Every Western analyst reaches for the same word: irrational. Overcapacity. Subsidy waste. Involution, Beijing’s own euphemism for the more embarrassing term.
They’re missing the point. Those losses aren’t the failure of the strategy. They are the price of it. They’re what the first rotations of a flywheel cost, and the reason almost nobody else can pay them.
A flywheel is not a good industry
A policy solves a problem. A subsidy buys an outcome. A moat defends a position.
A flywheel is a loop in which solving one problem makes the next problem cheaper to solve.
That last clause is the whole thing. It’s the only test that matters…
Does the next rotation cost less than the last one?
If yes, you have a flywheel. If no, you have an expense you’ll be paying forever, until a budget committee notices and stops.
Run China’s solar loop through the test. State capital funded scale. Scale pulled suppliers into physical proximity. Proximity created density — a module maker in Jiangsu sources polysilicon, wafers, silver paste, glass, and frames from firms clustered within a few hundred kilometers, at lead times measured in hours. Density crushed cost. Cost took share. Share funded research: China’s slice of global solar patent applications went from 13% in 2004, when Japan led with 43%, to roughly 65% by 2024, three-quarters of them filed by private firms. Better technology feeds back into cheaper manufacturing.
Six links. Each makes the next one cheaper.
The wheel now spins fast enough that Beijing is removing the supports. Since June 2025, new solar projects sell power into the market instead of collecting a fixed tariff. On April 1, 2026, the export VAT rebate, already cut from 13% to 9% in late 2024, was eliminated outright.3
You don’t take the training wheels off a bicycle you’re still carrying.
The first three turns always look like waste
This isn’t a Chinese-genius story or an American-decline story. It’s an accounting story.
A flywheel’s early rotations have terrible economics. That’s definitional. The advantage hasn’t compounded yet, so you pay full price for everything and get nothing back. Losses, redundancy, duplicated capacity, firms that will die… the first turns look exactly like a failing industry. On the spreadsheet, they’re indistinguishable.
So ask who’s holding the spreadsheet.
Public markets price the next four quarters. A flywheel is negative NPV for years under any discount rate a CFO is permitted to use.
📎 The flywheel is an American idea
Here’s the part that should sting.
The word is ours. Jim Collins coined it in Good to Great. In 2001, Amazon in a tailspin, the dot-com bust gutting the NASDAQ, the company bleeding cash and cutting staff, Bezos brought Collins to Seattle to coach his executive team. Collins told them the answer to a crisis isn’t to react to bad news. It’s to build a flywheel.4
They sketched theirs. Legend says it was on a napkin; the napkin is probably apocryphal and the loop certainly isn’t:
Lower prices → more customers → more third-party sellers → wider selection → better experience → more traffic → economies of scale → lower costs → lower prices.
Amazon went public in 1997. It didn’t post a full-year profit until 2003, $35 million, immediately after a $149 million loss. Six years of a public company being ridiculed for not making money while it built the loop. AWS didn’t launch until 2006.5
So the United States didn’t fail to understand flywheels. It invented the vocabulary and produced the canonical example.
What it lacked was the thing Bezos had: a principal who couldn’t be fired for the first six rotations. He held control, and in his 1997 shareholder letter he told the market flatly that he was optimizing for the long term — then reattached that letter to every annual report for the rest of his tenure, as if to say I warned you, and I haven’t changed my mind.
That’s the whole trick. Not genius. Not planning. Immunity from the people who price rotation one.
A founder with voting control can buy that immunity. A CEO on a four-year tenure can’t. A democracy on a two-year electoral cycle can’t. Hold that thought… it comes back with a vengeance.
Electoral politics prices the next election, and the United States has now run that experiment on itself twice, in public.
The Inflation Reduction Act was signed in August 2022 with a decade-long horizon. In July 2025, the One Big Beautiful Bill Act accelerated most of it into an early grave: consumer credits dead by year-end, EV credits dead by September, wind and solar credits compressed into a closing window. Three years, start to finish. No flywheel has ever ignited in three years.6
The CHIPS Act is the same story in a different key. The manufacturing subsidies bought fabs… that part largely worked. But the part meant to build an ecosystem rather than a building has been dismantled: the Natcast/NSTC research consortium discontinued and its funding clawed back, the planned design and advanced-packaging facilities cancelled, and the Industrial Advisory Committee, a body Congress mandated, now described on NIST’s own website in three words: “currently inactive.”7
Cost-benefit analysis prices the project, not the system. It cannot see the thing it should be measuring, because the value of rotation one lives entirely inside rotation nine.
This is why the West’s real flywheels are almost all accidental — they arose wherever someone happened to be immune to quarterly accounting. Silicon Valley’s loop was never designed; it emerged because venture funds have ten-year horizons and because California wouldn’t enforce non-competes, so engineers carried knowledge across firms and every failure became someone else’s input. Defense R&D compounds because procurement runs on decade-long programs. The great American flywheel of the twentieth century, universities → immigration → founders → wealth → universities, ran for eighty years largely because nobody was measuring it, so nobody thought to cut it.
We didn’t build those. We inherited them. You can’t inherit your way through the next thirty years.
Buying a rotation is not building a wheel
The obvious counterexample is TSMC in Arizona, and it deserves a fair hearing.
Arizona works. Fab 21 reached mass production with yields comparable to Taiwan. The cost premium is far smaller than the doomsayers claimed, TechInsights puts the wafer-processing gap under 10%, AMD’s CEO has said 5–20% on finished chips, and the fourth Arizona fab is sold out through 2027 at a 25–30% price premium, because customers will pay for supply-chain security.8
The fab isn’t a failure. It’s a success. It’s also not America’s flywheel. It’s TSMC’s flywheel, rotating on American soil.
The $52 billion bought capacity. It didn’t buy the thing that makes Taiwan’s capacity cheap, the tooling engineers, the supplier density, the twenty-year institutional memory of running a fab at the limit. Those live in Hsinchu. And the programs designed to grow an American version were the ones that got cancelled.
You can rent a rotation. You cannot rent a wheel.
The law is the thing that stops flywheels
Now the part nobody wants to say out loud.
A mature flywheel and an illegal monopoly are, structurally, the same object.
Read the test again, each rotation makes the next one cheaper, and notice that it’s also a fair description of what antitrust law exists to interrupt. Standard Oil was a flywheel. Bell was a flywheel. Microsoft in the nineties was a flywheel. In every case the state’s considered judgment was that a wheel nobody can stop is a wheel nobody can compete with, and it broke the wheel on purpose.
So the West’s failure to compound is not an oversight. It is a policy, with a statute behind it. We built an elaborate legal apparatus specifically to interrupt self-reinforcing advantage, on the entirely defensible theory that the alternative is living inside somebody’s monopoly. Then we spent forty years wondering why we don’t compound.
China’s flywheels face no equivalent brake. That is the actual asymmetry — not work ethic, not planning, not five-year plans. One system prosecutes compounding; the other subsidizes it.
Which brings us to the live American case, and it is not a comfortable one.
SpaceX is the most complete flywheel in the country. Reusability drives launch cost down; cheap launch makes your own constellation viable; Starlink throws off recurring revenue; revenue funds Starship, which drives launch cost down again. It is self-consuming, SpaceX is its own biggest customer, which is the purest form the loop can take.
And it just got bigger. In February 2026, SpaceX absorbed xAI, which had already absorbed X. In June it went public, raising $75 billion and clearing $2 trillion in market capitalization. The loop now extends from rockets through satellites through data centers into frontier AI, with compute rented back out, reportedly including to Anthropic and Google. Musk holds the property from the Amazon callout: he cannot be fired during rotation one.9
It has also generated the friction that a flywheel at scale always generates. Residents near the Memphis-area data centers have filed a class action over noise and vibration. The NAACP has sued over gas turbines allegedly running without proper Clean Air Act permits. A whistleblower suit landed days before the IPO. A shareholder group has asked the SEC to examine what the S-1 left out.10
None of that is adjudicated, and I’m not going to pretend it is. But here’s the thing… the outcome of those cases is not the point.
In June 2026, the Department of Justice moved to dismiss the NAACP’s suit, arguing that shutting off the turbines would threaten “American national, economic, and energy security.”
Read that again. The state did not weigh the flywheel against the law. The state intervened on the side of the flywheel, against a civil-rights organization enforcing an environmental statute, on the grounds that the wheel had become too important to stop.
That’s the Chinese model. Not the industrial policy… the indemnity. The willingness to let a compounding system externalize its costs onto a community because the compounding matters more than the community does.
So here is the honest ledger, and I’d rather state it plainly than pretend the argument comes free…
At sufficient scale, a flywheel stops being subject to the law and becomes an input to it.
That is not a slander against one man. It’s a structural property, and it is the same property that made Standard Oil and Bell and every other flywheel eventually intolerable. The compounding works. That’s exactly the problem. A wheel spinning fast enough acquires the political mass to protect itself, and at that point the accountability mechanism and the growth mechanism are pointed at each other.
America is not incapable of building flywheels. It is building one right now. It’s just doing it the only way our system permits, by exempting a man instead of building an institution. One founder, unfireable, absorbing the externalities of an entire industrial transition, with the Justice Department clearing his path.
If you want compounding without that, you have to say what you’re willing to trade. I don’t think anyone has.
Wheels run backwards
One more thing, missing from nearly every paper on this subject.
Compounding is not optional. It only has a direction.
Lose a factory, and you lose the suppliers who served it. Lose the suppliers, and the tooling engineers retire without apprentices. Lose the engineers, and the community college cancels the program for lack of enrollment. Cancel the program, and the next factory can’t hire, so it doesn’t get built, so two more suppliers close.
Each rotation makes the next loss cheaper to incur and the eventual recovery more expensive. This isn’t decline as a level. It’s decline as a slope — a flywheel with all the same self-reinforcing mathematics, pointed the other way.
The choice was never “build a flywheel or stay where you are.” There is no staying where you are. A system that isn’t compounding upward is compounding downward. The steady state is a fiction invented by people who check the balance sheet once a year.
Why this is now the only way to compete
Every copyable advantage now has a half-life measured in months. A feature, a model, a process, a price — anything that can be described can be replicated, increasingly by a machine, by a competitor you’ve never heard of. If your advantage is a thing, it’s already depreciating. Only advantages that manufacture the next advantage survive an environment that copies this fast.
Cost advantage is now a system property, not a firm property. No American solar company can out-manufacture a Chinese one, because it isn’t competing with a company. It’s competing with an industrial park, a provincial bank, a technical university, a port, and a hundred suppliers within a two-hour drive of each other. Non-Chinese modules run 50–100% more expensive despite subsidy. Firm-versus-ecosystem is a category error. You lose before you start.
Efficiency and compounding are actively opposed. Efficiency means eliminating slack, redundancy, and duplicated effort. But redundancy is how flywheels ignite — a dozen firms racing each other in one city is “wasteful” right up until the churn produces the density and talent depth that make the cluster unkillable. Every efficiency program is a flywheel-suppression program wearing a green eyeshade. We optimized ourselves into brittleness and called it discipline.
The deeper issue is that efficiency is optimization against a known objective.
Compounding requires discovering the objective. You cannot optimize a loop you haven’t found yet, and finding it costs failures, because the failures are the search.
Which means “waste” is a word that only works in hindsight. Ex post, forty bankrupt solar firms look like squandered capital. Ex ante, they were forty draws from a distribution nobody had sampled, the entropy budget you spend to locate a loop that isn’t on any map. Redundancy isn’t the opposite of R&D. Redundancy is R&D, run in parallel, on somebody’s balance sheet, filed under “losses.”
Put it in the terms an engineer would recognize… efficiency is hill-climbing. It moves you steadily up the nearest slope, and it is very good at that. But the flywheel is a different hill, and reaching it means walking downhill first… accepting worse numbers, for years, on purpose. Efficiency cannot cross the valley. It is structurally incapable of it, because every step across looks like failure by its own metric. Applied before the loop is found, efficiency isn’t discipline. It’s premature convergence; locking in the local optimum and calling the global one irrational.
Every efficiency program is a flywheel-suppression program wearing a green eyeshade. We optimized ourselves into brittleness and called it discipline.
None of this requires admiring the Chinese model, and I don’t. Its wheels have real costs: an industry that dominates the planet and cannot earn its cost of capital; capital poured into sectors that will never repay it; a state now spending three billion yuan on a platform whose only job is to retire the capacity it spent fifteen years building.
Grant every word of that. The argument isn’t “be China.” It’s: the compounding loop is the unit of competition now, and if you’re not building one, you’re dismantling one.
The test
It scales down cleanly, which is the tell that it’s a real principle and not a geopolitics take.
For a nation, a company, a founder, a person… one question, asked of every commitment of time or money:
Does this make the next one cheaper?
Reading an article is a task. Reading to sharpen the essay you’re already writing turns the reading into a rotation. Writing code is a task. Writing the library that eliminates a class of future work is a rotation. Hiring is a task. Building the bench that makes the next five hires teach themselves is a rotation. A subsidy that must be renewed at full price every year is not a policy… it’s an annuity for incumbents.
And the corollary, which is the uncomfortable half:
Anything that doesn’t make the next one cheaper is making it more expensive.
The ecosystem you don’t reinforce is the one somebody else’s is eating. There is no neutral spend, no neutral year, no neutral decade. Governments are good at solving isolated problems. Corporations are good at hitting quarterly targets. Individuals are good at finishing to-do lists.
All three are the same failure, and it isn’t a failure of effort. It’s a failure to notice that effort which doesn’t compound is effort you’ll spend again, at full price, forever, against someone who no longer has to.
The wheel is already turning. The only questions are which direction, who’s standing on which side of it, and what we’re prepared to let it run over.