China's Central Bank Is Buying the Risk on AI Capex

Seoul 1973: the state buys the credit risk

In 1973 the South Korean government of Park Chung-hee launched its heavy and chemical industry plan. The National Investment Fund collected compulsory savings and channelled them to the banks, which lent to steel, shipyards and petrochemicals at rates decided by the government.

The mechanism came down to three words: the state picks, the bank lends, the public balance sheet absorbs.

For six years it worked. By the end of the decade Korea had a competitive heavy industry and a mass of non-performing loans that cost two cycles of bank restructuring.

Japan had pulled the same lever earlier: the Bank of Japan's window guidance, through which the central bank told commercial banks how much to lend and to whom, right up to 1991. France had done it from 1945 with the Treasury circuit, which obliged the banking system to hold government paper and fund reconstruction at an administered rate.

Three precedents are enough to call it a pattern. The fourth carries the date 29 September 2026.

On 29 September 2026 Beijing pulls four levers

On Tuesday 29 September 2026 the People's Bank of China cut the one-year pledged supplementary lending rate by 0.25 points, to 1.5%, and extended the facility to six infrastructure networks: Caixin Global documents it^[1].

In the same announcement the central bank added 200 billion yuan to the relending quota for science and technology, which rises to 1,400 billion, and raised the support ratio on eligible loans to 100%.

The quota for agriculture and small business grows by 500 billion, up to a combined ceiling of 4,850 billion. The private enterprise quota rises by 300 billion and reaches 1,300 billion.

On 28 September the State Council had called for a timely adjustment of monetary instruments. Reuters read the package together with the mortgage subsidies (29 September dispatch^[2]), and the Asian financial press framed it as a growth support move (Kaohoon International^[3]).

The number that matters is the 100% ratio

Relending works like this: the commercial bank extends the loan to an eligible borrower, brings that credit to the central bank and obtains refinancing at a concessional rate.

The support ratio says how much of that loan the central bank refinances. At 60%, the commercial lender covers the rest with its own deposits and funding, and pays the market price on that share. At 100% the friction disappears.

Credit risk formally stays with the bank. The cost of funding passes in full to the central bank's balance sheet, which sets it by decree.

The price of money for the chosen sector therefore stops depending on the interbank market. This is the point the market has yet to price: a 1,400 billion yuan quota refinanced in full amounts to a dedicated capital channel, immune to the rate cycle and to lenders' risk appetite.

American AI capex lives on private credit

In the United States, compute capex is funded elsewhere: hyperscaler corporate bonds, private credit, special purpose vehicles that own the building and the cooling, vendor financing on the hardware.

The chain is long and private. Every link prices risk on its own account, and every link can shut when the price changes.

The BIS and the IMF have flagged in their financial stability reports that the fragility sits with whoever funds the data centres, more than with equity valuations.

Two architectures for the same industrial goal. Beijing puts the state balance sheet at the start of the chain, with a rate written by hand. Washington keeps it at the end, as the implicit insurer in case of accident.

The difference decides who pays first when compute demand disappoints, and how fast the funding stops.

How a relending quota becomes a public liability

The transformation happens through accounting and through politics.

Accounting: the refinancing sits on the central bank's asset side, and any loss erodes the capital of an entity that belongs to the state. Politics: when a programme carries the name of a national priority, widespread borrower default becomes a matter of government, never a market outcome.

The Korean precedent shows it precisely. The non-performing loans from the 1973 plan ended up on the public balance sheet over the following decade, through the industrial rationalisation programmes of the 1980s.

Then there is the effect on price. Full refinancing at a concessional rate makes it rational for a bank to lend at the thinnest margin, because the cost of funding is known and fixed. Private capital assessing the same project in China competes with a cost of money decided around a table.

This is a regime change in funding structure, never a cyclical adjustment.

This desk's position, and what would refute it

With the ratio at 100% and a quota of 1,400 billion yuan, Beijing has nationalised the funding risk of artificial intelligence, while Washington leaves it to private credit: this divergence weighs more than the gap between the models.

The argument rests on a verifiable fact. The price of capital for the chosen sector in China now follows from an administrative decision, published, dated.

Three facts would change my mind. First: disbursement data showing marginal drawdown on the sci-tech quota, a sign that the constraint sits in project demand, more than in the cost of credit.

Second: a revision of the support ratio below 100% within twelve months, which would downgrade the move to a tactical experiment. Third: an explicit public guarantee on American capex, which would close the divergence from the opposite side.

Three implications for capital

1. Family offices and sovereign funds, 36-month horizon. The risk premium on Chinese tech champions embeds a cost of capital that is now administered. The practical consequence concerns selection: suppliers inside the eligible perimeter enjoy funding that their competitors pay for at full price.

2. Boards of directors, 18-month horizon. Geopolitical risk in strategic plans looks at export controls. The lever that moves margins arrives from the funding side: a competitor with funding at 1.5% sustains prices that erode the gross margin of anyone borrowing at market rates.

3. Chief risk officers, 24-month horizon. VAR models calibrated on historical variance treat the Chinese cost of credit as a market variable. What is needed is a scenario where that cost stays fixed by decree while global rates rise.

For anyone carrying the macro narrative to investors, a warning applies: the technology advantage story holds up poorly when the competitor competes on the cost of capital.

The forecast, and what to watch

By 30 June 2027 the People's Bank of China takes the science and technology relending quota beyond 1,600 billion yuan. Confidence: 70 out of 100. Horizon: 270 days. Verification: the central bank's releases on structural monetary policy instruments.

Falsification signal: at 30 June 2027 that quota stands at or below 1,400 billion yuan in the central bank's official documents.

What to watch:

• the support ratio in the coming releases: a move below 100% inverts the reading
• actual drawdown against the ceiling, published in the quarterly monetary policy report
• the arrival of similar instruments elsewhere: a European or Japanese development bank refinancing compute capex at an administered rate

The divergence between administered capital and private capital always resolves. The question is how.