US-China Decoupling: Who Pays for the $184 Billion AI Push

The technological contest between the United States and China over AI is usually read as a race between models and chips. The structure that funds it tells a different story.

Tokyo, 1953: postal savings become industrial policy

By 1953 Japan had already assembled the machine that would pay for thirty years of growth. It was called Zaitō, the Fiscal Investment and Loan Program: households' postal deposits flowed into the Treasury and came out as direct credit to steel, shipyards, chemicals.

Three words describe the mechanism: administered rates, closed capital accounts, guided credit.

Japanese households earned modest real returns on their deposits. Heavy industry received capital at a political price. The Bank of Japan closed the loop with window guidance, allocating lending quotas to banks sector by sector.

The system held as long as foreign demand absorbed the output. Once the industrial surplus exceeded what the world wanted to buy, guided credit went looking for other outlets and found them in urban land and equities. On 29 December 1989 the Nikkei hit its all-time high. The following decade went to working off that machine's liabilities.

Seoul, 1973: the second run of the same mechanism

In January 1973 Park Chung-hee announced the heavy and chemical industrialisation plan. Six designated sectors, among them steel, shipbuilding, electronics, chemicals and machinery.

The National Investment Fund, created in 1974, collected deposits from savings institutions and directed them toward those sectors at administered rates. Korean households paid for their own heavy industry, unaware of the share.

The mechanism was identical to Tokyo's, twenty years later and in far more of a hurry.

The reckoning came in 1979: double-digit inflation, excess capacity in the shipyards, a stabilisation programme imposed in April. Korea closed 1980 with its first post-war contraction in output. Industry survived and became competitive; savers paid the bill, through years of negative real returns and a devaluation of the won.

Two precedents remain a hypothesis. The third turns it into a pattern.

Beijing, 2026: $184 billion and the same architecture

Stanford's AI Index 2026 estimates that China's state guidance funds poured $184 billion into AI-related companies between 2000 and 2023. The figure is taken from an op-ed published by the South China Morning Post on 24 September 2026^[1], written by a voice with ten years of factory work behind it.

Guidance funds, zhengfu yindao jijin, are hybrid vehicles: local public capital that attracts private co-investment and directs it toward the sectors designated by the plan.

The form is private equity. The function is Zaitō.

The funding channel repeats that of the two precedents: deposits that pay little, capital controls, a banking system that lends where the plan points. Anyone living in China has few alternatives between bank deposits and property, and the second stopped working in 2021.

The mechanism: capex arrives before revenue

Financial repression is an invisible tax. It transfers income from depositors to designated borrowers, and its yield grows with the gap between the administered rate and the market return.

Applied to iron and ships, it produces exports. Applied to data centres, it produces compute capacity waiting for a customer.

This is the difference from 1953 and 1973. Japanese steel and Korean ships faced foreign demand ready to absorb them. Chinese compute faces an external market already segmented by American export controls, and a domestic market whose spending power is compressed by the very mechanism that pays for the facilities.

Technological decoupling between the United States and China closes the external outlet. Financial repression closes the internal one. Both closures have the same accounting origin.

The position: the $184 billion compresses the demand it is meant to serve

This desk's thesis is blunt. That $184 billion is the final stage of a financial repression model, and the structure paying for Chinese AI capex compresses the consumption that would validate it.

Calling it a cycle is a category error. It is a regime change, rooted in an architecture that has been running since the 1990s.

Household consumption's share of Chinese output remains among the lowest of the major economies, and the gap has held for twenty years. On 20 September 2026 the New York Times devoted an analysis of China's AI economy^[2] to the subject. The point raised in the Hong Kong paper remains the most direct one: ordinary people have little money in their pockets.

A model that pays for facilities by implicitly taxing those facilities' end customers runs into an arithmetic limit.

The serious objection, and what would change this reading

The strongest objection runs like this: AI is an intermediate good, so the demand that matters is corporate demand, and Chinese firms export. Should automation raise productivity in export industry, compute would find its customer beyond the border.

The objection carries weight. It holds on condition that destination markets stay open, and export controls, tariffs and European data rules all work the other way.

What would change this reading: hukou reform extended to first-tier cities, a national pension system funded by transfers from the central budget, liberalisation of deposit rates. Three measures that would shift income from firms to households in a structural way. Rhetorical signals in that direction have existed for years; the measurable transfers remain modest.

Three implications for capital

First, 36-month horizon: Chinese equity exposure should be split between producers of capacity and owners of demand. Whoever sells compute inside a plan gets guaranteed orders and administered margins. Whoever sells to the Chinese consumer faces a market the plan itself is cooling.

Second, 24-month horizon: Chinese deflation becomes an export good. Excess capacity with no domestic outlet looks for prices elsewhere, and enters European inflation models as a favourable supply shock, with effects on rate differentials.

Third, 36-month horizon: the credit risk tied to AI capex requires two separate maps. In the United States the financing runs through private credit and neocloud vehicles. In China it runs through the balance sheets of local governments and systemic banks. The first structure reprices fast; the second accumulates quietly and then lands on the public balance sheet.

The market has yet to price the second.

The forecast

By 31 December 2027 household final consumption expenditure as a share of Chinese GDP will remain below 45%, even with new consumption support programmes announced by Beijing.

Confidence: 72%. Horizon: 462 days, to 31 December 2027. Verification: the annual data from China's National Bureau of Statistics. Kill signal: a reading at or above 45% for the year 2027.

What to watch

Three indicators will show which of the two readings holds.

• New guidance funds announced by the provinces in 2027: acceleration or consolidation.
• The urban household saving rate in the People's Bank of China's quarterly surveys.
• Export prices for Chinese servers and modules: a rapid decline signals capacity looking for an outlet.

The first measures the input. The second measures confidence. The third measures the discharge, and it reaches European and American balance sheets before the other two do, in the form of imported deflation rather than industrial competition.