Seventeen mechanisms that decide how an economy's output is measured, who gets it, how it grows, and why it stalls. Each one is a machine you can drive. The principle comes first; China, the course's subject and the largest natural experiment of our lifetimes, is the worked example.
The mechanisms and figures come from Professor Bai’s lectures and course slides. A few comparisons from outside the course are added where they sharpen the point, and each is marked outside the course; figures taken from the slides are marked . Figure parameters the course did not give are labelled illustrative. Companion page: the full course summary.
Before any question about growth or distribution can be asked, there has to be a number for "the economy". The number is gross domestic product (GDP), and it is built from one idea that is also the right way to think about a firm, a product line, or a career: value added.
GDP is the value of final goods and services produced in a place in a period. "Final" is there to stop double counting: a phone and the chip inside it are one product, not two. But nobody can decide transaction by transaction what is final. Crude oil is final when the United Arab Emirates (UAE) exports it and intermediate when a refinery buys it; the same is true of electricity and of artificial intelligence (AI) tokens. So statisticians measure value added instead, for every producer: sales minus what was bought in. Summed over all producers it equals final output, and the classification problem disappears.
The same subtraction is the honest measure of a business. Revenue is what passes through you; value added is what you create, and it is the pool from which wages, capital income and taxes on production are paid. Move the prices in the chain below. The total is always the final price, and the sum of all sales always overstates it.
Three conventions follow, and each is a trap in practice. A machine bought for production is investment, counted as final even though it makes other things. Reselling an existing machine or apartment is a transfer, not production, and only the agent's commission enters GDP. A revaluation, an apartment worth ten million now worth eight, touches nothing. The test the lecturer kept returning to: "is somebody making a production effort that creates value?" Value added is also not profit: labour cost is not subtracted from it.
Two more matter for reading any country's numbers. Owner-occupied housing produces a service nobody pays for, so an imputed rent is added; China's imputation has been based on construction cost rather than market rents, a large under-estimate. And depletion is not deducted: a barrel of oil counts in full when sold, which is why "green GDP" keeps being proposed and never becomes operational.
One more distinction decides which country a given production belongs to. GDP counts by place: everything produced inside the borders, whoever owns the factory. Gross national product (GNP), now usually called gross national income (GNI), counts by ownership: what a country's people and firms earn wherever they earn it. The difference between them is net factor income from abroad, the profits, interest and dividends flowing in minus those flowing out. For an aging economy that has invested overseas for decades the gap is large: Japan's net income from abroad has run at roughly 5 to 6% of its GDP in recent years outside the course, which is why its "lost decades" look different in GNP terms.
The same GDP can be counted from production (sum of value added), from income (what that value added is paid out as) and from expenditure (who bought the final output). They agree by construction, because unsold output is counted as inventory investment, as if the producer bought it from itself.
The lecturer's rule: decide which side you are on before you start a sentence, and never mix them. Rearranging the expenditure side gives the identity he called the most important formula for a country's external balance. National saving is what is left of income after consumption and government purchases, and it can only be invested at home or lent abroad:
A trade surplus is not a sign of competitiveness and a deficit is not a sign of weakness; each is the mirror of a saving-investment gap. A tariff is in effect a consumption tax on the importing country's own buyers: the duty is added to the price they pay, and unless the foreign seller cuts its price by the full amount, they bear it. So it can move the trade balance only by moving saving or investment. If the revenue is spent, nothing moves. Set the shares below, then apply a tariff and choose what the government does with the money.
Two readings. The United States invests more than it saves, so it runs a deficit that years of tariffs did not change, because the revenue was spent. China saves more than it invests, so it runs a surplus: near 9% of GDP at the 2007 peak, about 2% afterwards, and about 4% recently. When the surplus fell after 2008 the room was used for investment rather than consumption, which is where Part IV picks up.
Output is one number; it is paid out as wages, as returns to capital, and as production taxes. The split is not a political choice in the first instance. It falls out of prices in the factor markets, and the model behind those prices explains most of what has happened to labour's share in the last thirty years, in China and elsewhere.
Why does labour get the share of output it gets? The textbook answer needs four assumptions: one representative producer, a production function Y = F(K, L) (output Y as a function F of the capital stock K and the number of workers L), fixed supplies of capital and labour, and competitive markets. From those, the answer falls out of hiring decisions. A worker paid ten thousand who produces eight thousand is let go; a worker paid eight who produces ten is hired. So the wage settles at the marginal product of labour, and the rental rate of capital at the marginal product of capital. With constant returns to scale, wages times workers plus rent times capital exhausts output exactly. There is no residual for anyone to fight over, and no excess profit.
The Cobb-Douglas form makes the split one number. With Y = A·Kα·L1−α, where A is a productivity level and α (alpha) is a number between 0 and 1, capital's share of income is α and labour's is 1 − α whatever the quantities. A is the level of technology in the lecture's words, and total factor productivity in the measured sense: everything that raises output from the same inputs, including how well they are allocated and managed. The reason to tolerate such a strong form is empirical: the US labour share sat near two-thirds for decades. Drag L to see diminishing returns, then α to see the split change while the two incomes still add to output.
The lecturer said what he does not believe in this (people work for meaning, not only pay) and what it leaves out (skills, and technical change that favours one factor). He keeps it because it isolates the useful question: if the shares are supposed to be constant, what moved them?
Section 3 was about production: how output splits into wages and capital income. This section follows that split forward to what households can spend, because the two are linked by a chain with three links, and China's story is a break in each of them.
A fall at any link lowers the consumption share, and in China all three moved the same way. The labour share fell as workers moved from farms, where the producer keeps everything, into factories, where output is split with capital. The household income share fell with it, because capital income accrues to firms and to concentrated owners rather than to wage earners. And households saved more of what they did receive, for the two reasons that end this section: migrants have no urban safety net, and families had to buy a home. All three links are in this section; Section 5 then asks whether wages can be legislated back up.
Now the pattern that confuses most discussions of inequality, and the course's central case. Between 2000 and 2009 Chinese households' consumption fell from 47% to 35% of GDP. Over the same years real GDP grew to about two and a half times its size, so real consumption itself still rose by about 80%. Nobody got poorer. What fell was households' claim on each additional yuan.
The largest single reason for the first two links breaking, in the lecturer's own research, is a composition effect. A farmer keeps all of the farm's output and consumes most of it. Moving to a factory multiplies that person's output, but the worker keeps about half; the rest goes to capital, which is concentrated in few hands that save most of it. Under the hukou system the migrant also saves heavily: no urban registration means no schooling for the children, no safety net and no right to buy a home, so the roots stay in the village. Roughly 300 million people made this move. Every one of them consumed more; the economy's consumption share fell anyway.
The general lesson outlasts the case. Any time a population shifts from a low-productivity activity where the producer keeps everything to a high-productivity one where returns are split, the average share of the many falls even as each of them gains. It is true of workers moving into industry, of freelancers joining platforms, and of any market where scale concentrates the returns. The decomposition on the course slides puts 60% of the fall in China's household income share on the labour share, and 61% of the labour-share fall on the move out of agriculture . Housing reform broke the third link, the one from household income to household spending. From the late 1990s families bought homes that employers had previously allocated, and in the national accounts a bought home is investment, not consumption, so a decade of rising incomes flowed into I rather than C. Because land is a state monopoly, the land price inside each home was also how local governments financed infrastructure, which is why Section 16 returns to it.
If labour's share is the problem, why not legislate a raise? The answer is one number: the elasticity of substitution between capital and labour, which the lecturer's estimates put at about one for Chinese industry. When labour costs 1% more, firms replace enough workers with machines that employment falls about 1%. Higher pay per worker, fewer workers, the same wage bill, the same share. What does raise labour's share, on the same evidence, is less market power in product markets, because a firm that can charge above cost keeps the markup as profit and profit is capital income, and a shift toward labour-intensive activity, above all services.
The number cuts both ways for anyone running a payroll. At elasticity one, automation adopted in response to rising labour cost does not raise capital's share either; it just replaces the workers whose cost rose. The lecturer's answer to the consumption question was therefore not wages but jobs: let services charge prices that cover costs, because that is where employment will come from, and remove the barrier he called cultural rather than economic, the jobs in demand (care, nursing, teaching, welding) that go unfilled because people do not see them as theirs.
Parts I and II asked how a given output is measured and divided. This part asks what makes the output bigger, which is a different question with different answers: the levers that move a share (wages, transfers, taxes) are not the levers that move the total.
The long-run question is also different in kind from the short-run one. In the long run wages and prices adjust, everyone who wants work has it, and what remains is the capacity of the economy to produce. Solow's model of the 1950s is still the benchmark for that capacity, and it delivers three results the lecturer singled out: saving buys a level and not a growth rate (Section 6), only efficiency raises income per head, which he called the single most important idea of the model (Section 7), and convergence, which he called the most hopeful, because it says a poor country can catch up (Section 8).
What decides how rich a country becomes, and can it get richer just by saving more? The model that answers this tracks one quantity: capital per worker. Lower-case letters are per worker: k is capital per worker and y is output per worker, with y = f(k) a function that has diminishing returns. A fraction s of income (the saving rate) is saved and invested. A fraction δ (delta, the depreciation rate) of capital wears out each period, and if the labour force grows at rate n, new workers must be equipped too. So capital per worker obeys one equation, the tap minus the leak:
The leak grows in proportion to the water; the tap, because of diminishing returns, grows more slowly. So the tub fills to the level where they match and stops. Raise the saving rate and the tap opens, the level rises, and it stops again higher up. Prediction 1: a higher saving rate buys a higher level of income per worker, not a higher long-run growth rate. The growth spike after the switch dies back to zero. Prediction 2: faster population growth lowers capital and income per worker but leaves the long-run growth rate unchanged; a bigger population means a bigger economy, not a richer one. Extra investment cannot escape the steady state, because beyond k* the extra output is less than the extra maintenance.
The escape from the steady state is not more saving but a shift in the production function itself. Section 3 put technology in front of the function as A, which raises output whatever the mix of inputs. Here it is more convenient to put it beside labour instead, as labour-augmenting efficiency E, so that one worker with E = 2 counts as two effective workers, and to let E grow at rate g. The two are the same idea written two ways; the lecture used whichever was convenient. Everything now happens per effective worker (k and y are now capital and output per unit of L·E), break-even investment becomes (δ + n + g)·k, and the steady state stops moving in those units. But that means output per actual worker grows at g forever, and total output at n + g. The single most important idea of the model, in the lecturer's words: sustained growth of income per person comes only from efficiency improvement. Saving cannot do it, population cannot do it.
"Technology" is too narrow a word for E. The lecturer prefers "efficiency": science and engineering, but also education, institutions, and a labour market that gets people into work. It is measured as a residual and has run at about 2% a year in advanced economies since the Industrial Revolution, and essentially zero before it. Set the two rates below and read the three growth rates off the lines.
Three implications survive contact with data, which is how a model earns trust. The capital-output ratio should be roughly constant, and it is; China's rose about 0.5% a year over 1978 to 2007, consistent with an economy converging up to its steady state. Real wages should grow with output per worker, so the labour share should be stable, which holds in the US. And the real interest rate should be roughly constant, around 2% in advanced economies. One more idea from the discussion is worth keeping: an economy's speed is set by its slowest input, not its fastest. AI may race ahead while energy, capital or education hold the whole back. Whether US debt of about 40 trillion dollars is sustainable turns on the same question. A debt burden is measured against GDP, so it shrinks on its own whenever the economy grows faster than the interest on the debt, and efficiency growth is what decides the growth side of that race.
The most hopeful idea in growth theory, told through 1985: the lecturer arriving in the United States to highways and supermarkets that did not exist in China, when Chinese income per head was about one-fiftieth of American. Far below the steady state, a little investment adds a lot of output, so, other things equal, poorer economies grow faster. Singapore, independent in 1965, now exceeds US income per head.
Two forces do the work. One is moving up to the steady state with a given technology. The other is learning, which shifts the production function itself: knowledge flows no matter what barriers are built, and imitation, though scorned, is how development works. Both fade as the gap closes, which is why China's growth has gone from 10% to 5% and cannot go back. And both need the ability to learn. An illiterate population cannot catch up; education, research and development (R&D) and openness switch convergence on. Convergence is therefore conditional, not universal: controlling for saving, population growth and human capital, income gaps close about 2% a year, and only a handful of countries have escaped the middle-income trap.
Why China's rate was so large for so long, in the lecturer's account, is that the potential had been suppressed and was released all at once: widespread basic schooling built up before 1978, urbanisation from about 15% to 60% that freed 300 million people from subsistence plots, markets allowed to work, infrastructure built by decree, students sent abroad on Deng Xiaoping's bet that even a 5% return rate would be a win, entry to the World Trade Organization (WTO) and the state-enterprise reform of the late 1990s. Other countries did not have the suppression, so the recipe does not transfer. In the same spirit he rejected the 2008 argument that China could grow 8% for twenty more years because Japan, Singapore, Taiwan and Korea had done so from one-fifth of US income: their dependency ratios fell for decades while China's bottomed in 2010, and China's participation rate was already 60%. That view, he said, is the source of many of today's problems: over-ambitious targets, inefficient investment, and a real-estate sector built for a brighter future than arrived.
Growth can be decomposed. Start from the same production function as Section 3, with one change: labour L becomes human capital H, the headcount weighted by schooling, so that a better-educated workforce counts as more labour. With Y = A·Kα·H1−α, output growth is the weighted sum of input growth plus whatever is left, the Solow residual, which is efficiency growth. The lecturer uses a variant that measures capital by the capital-output ratio rather than the stock. It gives efficiency double weight, because higher efficiency also induces firms to invest, and it credits market-driven investment to A and H; what remains in the capital-output term is investment that happened for other reasons, such as stimulus.
Applied to China it tells the whole story of the slowdown. From 1978 to 2007, with growth of 10%, about 60% came from efficiency, 35% from human capital (a younger population, and college intake rising from under 300,000 to five million) and 5% from a rising capital-output ratio: textbook growth. From 2008 to 2015 human capital's contribution collapsed, efficiency's contribution fell from about 6 to 2.6 percentage points, and most of the 8.7% came from capital deepening, which requires ever more investment and cannot last. Choose a period, then build your own economy from the three sources.
Why efficiency growth fell after 2008: the catch-up effect diminishing, a smaller agricultural sector to draw on, a service sector whose productivity grows slowly, a falling trade-to-GDP ratio, a falling participation rate as retirement at 60 met more years in school, and a rising investment rate. From 2015 the campaign against overcapacity, debt and zombie firms raised efficiency's share again, at a lower growth rate; then COVID.
Across the world, investment and efficiency growth go together: technology arrives embedded in new equipment, investment has spillovers, high efficiency attracts capital, and a good business environment drives both. China is the exception. Across its provinces, higher investment rates go with slower efficiency growth. The reading is an inverted U: up to a point more investment raises efficiency, beyond it investing faster means choosing worse projects. China invests over 40% of GDP, about double the average of the Group of Twenty (G20) economies without China, and may sit past the peak. The lecturer's emblem was a bridge in northern China built on flat land over a river the government planned to dig later, paid with an IOU, to meet a growth target.
Solow has one kind of capital; reality has private structures and machinery, human capital, R&D and public infrastructure, with different horizons and different gaps between private and social returns. Who allocates between them, the market or the state, is the industrial-policy question, taboo for too long and now everywhere. Against: governments may not be able to pick winners, and politics may decide who is picked. For: markets under-invest where private investors ignore social returns. The lecture left it there, as an open trade-off rather than a verdict.
What benchmark says a consumption share is too low? The golden rule. The objective is not output but consumption. In the steady state investment equals break-even investment, so consumption per worker (c*, the star marking a steady-state value) is the gap between the output curve and the break-even line, c* = f(k*) − (δ + n + g)·k*. Save nothing and there is no capital and nothing to consume; save everything and there is nothing to consume either. The best saving rate is where the gap is widest, which is where the slope of the output curve equals the slope of the line: MPK = δ + n + g, where MPK is the marginal product of capital, the extra output from one more unit of capital, which is also the return on investment.
That gives a test any investor can run with three ratios. If the return on capital net of depreciation exceeds the economy's growth rate, the economy is below the golden rule and should save more; if it falls short, it is saving too much and should consume more. The US numbers from class: capital about 2.5 years of GDP, capital income 30% of GDP, depreciation 10% of GDP, so MPK is 12%, δ is 4%, and 8% against growth of 3% says the US saves too little. For China the lecturer stated only the sign: the net return to investment is below its 5% growth, so it invests too much and should consume more. The golden-rule rate itself is unknown; "it is easier to tell which side of it you are on."
How fast can an economy grow if everything goes right? The lecturer, who sits on the expert panel that advises China's five-year plans, splits GDP growth into labour-productivity growth plus labour-force growth and treat them separately. Labour-force growth comes from demographics and participation, which are projectable. Productivity growth is read off history: plot productivity growth against productivity relative to the United States for eighteen economies, East Asian and Western European, and the upper edge is a surprisingly tight downward-sloping band. That edge is the potential; most economies sit well below it.
Its track record is the reason to take it seriously. Using data to 2008 it projected productivity growth for 2009 to 2014 that matched what happened almost exactly, while GDP ran ahead because of the stimulus. For the 13th plan it gave 6.36% against the 6.5% the doubling target needed, and helped stop a higher target being set. The projection below is the 2015 vintage from the course slides. It has not been redone for the fertility drop since 2021, which reaches the labour force only around 2040, and the stated purpose of the work is to make planners more conservative: "if everything goes right, this is the potential."
Two things follow for anyone forecasting China. The productivity band implies about 4.5% potential for the 15th plan, and the plan carries no numerical growth target, because of uncertainty, while keeping the 2035 goal of a moderately advanced economy, read as about 20,000 dollars per head against 14,000 to 15,000 today. And if the potential is realised, China's economy would be about 1.8 times the United States by 2050 at purchasing power parity, while income per head would be about half of America's. Size and wealth are different questions, which is Prediction 2 again.
Solow has no price in it at all. The short run is where prices and wages are sticky, demand can fall short of capacity, and policy has real effects. The tool is the IS-LM model (investment-saving, liquidity-money), old but, in the lecturer's words, still relevant, and the question is what China's stimulus after 2008 did.
The goods market: Y = C(Y − T) + I(r) + G, where C(Y − T) says consumption depends on disposable income (income Y minus taxes T), I(r) says investment depends on the real interest rate r, and G is government purchases. One equation, two unknowns, so it traces a relation between the interest rate and output rather than a solution: a lower rate raises investment, and output must rise to meet it. That downward-sloping line is IS. A rise in government purchases shifts it right by the multiplier 1/(1 − MPC), where MPC is the marginal propensity to consume, the fraction of an extra unit of income that households spend: the government's yuan becomes a firm's income, then a household's, which spends 0.7 of it; the next household spends 0.7 of that, and so on. Summing 1 + 0.7 + 0.49 + … gives 1/(1 − 0.7), so with China's marginal propensity to consume of about 0.7 the naive multiplier is 3.3, "too large to be reasonable, and we will see why."
The money market: people hold liquid money to transact, so demand rises with income; they pay for it in forgone return, so demand falls with the interest rate. With the real money supply fixed by the central bank, each level of income has one rate that clears the market, and higher income needs a higher rate. That upward-sloping line is LM (L for liquidity, M for money; the real money supply is M/P, the money stock M divided by the price level P). Interpret the interest rate as a price, the price of liquidity, and much becomes easier. Where the two lines cross is the only pair of income and interest rate that clears both markets at once.
Now the multiplier shrinks. More G raises income; more income raises money demand; with the money supply fixed the rate rises; the higher rate cuts private investment along IS. Government spending crowds out private investment. Drive it below, and watch the naive point and the actual equilibrium separate.
Why are prices sticky at all? Wages are set in contracts for a period, and customers punish price changes that have no visible reason, war being the exception. So the model holds the price level fixed first, then lets it move. Above the economy's natural rate of output (Ȳ, "Y-bar", the output it produces at full employment) people work overtime, costs rise, and firms must raise prices; below it there is slack and prices fall. China today, in the lecturer's reading, is below: demand too weak, downward pressure on prices.
That adjustment is what connects the short run to the long. Start at the natural rate and print money: the rate falls, output rises, the economy overheats, and prices rise until real money balances are back where they started. Output, the rate, consumption and investment all return; only the price level is higher. Money is neutral in the long run. Below the natural rate, monetary policy has real effects, because the same expansion fills slack instead of overheating. A negative demand shock (a war ending, tariffs, a foreign crisis) works in reverse: output drops, prices fall over time, real balances grow, the rate falls, and private investment rises until output is back, with a new mix, less government and more private investment. That is the peace dividend.
China's 2009 stimulus is the case the whole course was built toward. The investment rate, already high after WTO entry, was pushed to about 47% of GDP in 2010 and 2011. The lecturer, who has served as an independent bank director, described the mechanism: after the late-1990s clean-up moved bad loans off the banks, they had become market-minded and reluctant to lend to local governments. In late 2008 regulators signalled to bank boards that a bank refusing a local-government loan would have to explain the refusal in writing. Bank credit expanded by 13% of GDP in the first quarter of 2009. Where US federal grants went to repaying state debt, Chinese local governments took the signal to the bank: "if you ask me to do one, I will do three."
In the growth accounting of Section 9 the cost is visible: efficiency's contribution fell from about 6 to 2.6 percentage points, and growth was bought with a rising capital-output ratio. The course slides spell out the channel and put a number on it . Subsidised infrastructure is intensive in unskilled labour and capital; it bid up unskilled wages and the cost of capital, crowded out market-oriented firms that use educated labour, raised debt, and lowered the return to capital. The estimated loss of consumption relative to a no-stimulus path peaked near 12% in 2010 and faded only by 2016.
The lesson was learned too well. The centre is now reluctant to stimulate at all, central-government debt is only about 23% of GDP, and the lecturer's view is that stimulus should be stronger after the real-estate correction, but through the six networks the 15th plan names (the power grid, data centres, logistics, underground utilities, water, communications) rather than roads and housing. The old channels are exhausted.
None of the above happens by itself; it is decided by officials, and their incentives explain the investment bias better than any theory of demand. Local governments control land, licences and regulation. Leaders are judged on GDP growth. A prefecture leader serves about three years in a post, so a project that lifts GDP now and leaves debt to a successor is attractive, and debt is not in the contract. Major taxes are collected where goods are produced, not consumed, so a city competes for factories and protects its firms, never its consumers. Investment can be enticed; consumption is a household decision.
The same horizon shapes the plans. Officials aim to finish most of a five-year plan by year four. Targets, once set, bind: the 2012 pledge to double GDP by 2020 bound every plan after it and was nearly met despite COVID. When rigidity bites it bites hard: an energy cap translated from a carbon target met booming exports, power was rationed and factories shut for months to hit it, and the indicator was dropped afterwards. The proposal from the course is to make household consumption growth a performance indicator and to collect tax by place of consumption, which digital payments make feasible; the obstacle is data.
Section 16 explained what officials choose to build. This one explains how they treat the firms that build it, and why a country can grow fast with rules that look terrible on paper.
In 2012 the World Bank's Doing Business survey ranked China only 91st of 185 economies. The government asked how an economy with such a poor business environment could grow so fast, and the Bank had no answer. The lecturer's research does: the indices measure rules on paper, read by consultants who simulate a start-up in Beijing and Shanghai, not implementation or informal practice. Local governments are intensely business-friendly but cannot help everyone, so they help selectively. A firm in a five-year-plan industry is a natural choice. In one city every leader from the party secretary to the vice-chairs of the consultative conference carried a portfolio of firms whose problems with land, approvals and regulation were escalated up the chain until solved. He would now call it the exercise of discretionary power rather than preferential treatment.
The evidence, since investment agreements are commercial secrets, comes from registration data traced to ultimate shareholders: private firms with a state entity among their owners are more productive and get cheaper credit. The system's costs are on the slides too: unfairness, corruption, over-support of favoured sectors leading to overcapacity in solar panels and electric vehicles, a bias for producers over consumers, local protectionism, and a weak incentive to improve the formal rules . The direction now: rules are improving, so there is less demand for discretion, and the risk to an official of exercising it is rising. The lasting point for anyone entering a market is that the written rules and the operating rules are different documents, and growth follows the second.