Hyperscaler AI Capex: The War for Capital

The precedent: 1847, when the railways drained the City

In 1847 Britain discovered that capital has a physical limit.

The railway companies, after the mania of 1845, called in the instalments on subscribed shares just as the Treasury and grain importers were asking the same marketplace for credit. The mechanism was simple: three demands, one pool of savings.

Historical estimates put that year's railway capital calls above £40 million, a share close to 7% of national income. The Bank of England's rate rose to 8% and on 25 October 1847 the government suspended the Bank Charter Act of 1844. The railways were real, the demand for transport was real, and the crisis came anyway, because the problem lay in the financing and never in the asset.

The 2026 pattern: three demands on the same savings

One hundred and seventy-nine years later the structure is identical, the context is different.

Goldman Sachs estimates that AI-related borrowing could reach roughly 1% of global GDP, and this demand comes on top of already heavy public financing needs, as documented in the Hubbis Macro Corner of 14 September 2026^[1]. BlackRock adds the third actor: Japanese investors hold roughly $1.1 trillion of US Treasuries and now find competitive yields at home. The ECB raised its three key rates by 25 basis points on 10 September, and UBS has shifted its base case to two 25-point Fed hikes by year-end, after the August payrolls.

The consensus reads these data as a monetary cycle running late. That reading is wrong in its category, before it is wrong in its numbers.

Three precedents are enough to call it a pattern: 1847, the American crowding-out of 1981-1984, and 2026.

The mechanism: why the long end of the curve ignores central banks

A central bank controls the overnight rate. The long end of the curve prices the savings available against the paper asking to be absorbed.

The American precedent shows it with exact numbers. In September 1981 the 10-year Treasury yield hit 15.8%, the high of the historical series. Federal deficits climbed to nearly 6% of GDP in fiscal year 1983, while private capex was asking for the same savings. Volcker's Fed cut policy rates from 1982, and the 10-year stayed above 10% until 1985.

When sovereign debt, AI capex and Japanese repatriation ask for the same capital in the same quarter, the term premium rises regardless of the path of policy rates. The ECB hike and the two Fed hikes UBS pencils in are the noise. The signal is the build-up of structural demand on the 10- and 30-year segment.

This is a regime change, and it is structural. The 2020-2026 rate cycle closed the thirty-year bull run in bonds, and VAR models calibrated on historical variance are measuring a world that has disappeared.

The debt link: who finances the data centres

The systemic risk of AI capex lives on the liability side, never on the asset side.

Hyperscaler equity valuations are a matter for shareholders. Data centre financing structures are a matter for the system. That is where private credit, leasing and bonds accumulate, priced for revenues that the software has yet to produce.

Eastspring observes that funding structures and balance sheet quality weigh ever more heavily as infrastructure investment broadens. According to the same Hubbis review^[1], Swiss Re estimates that AI data centres and renewables could generate roughly $200 billion in cumulative insurance premiums between 2026 and 2030. An insurance market of this size is born when assets become large enough to require protection from physical risk and counterparty risk.

Lighthouse Canton notes that US debt sustainability holds as long as nominal growth exceeds the effective cost of debt, and that this cushion is temporary. AI capex shortens the life of that cushion, because it pushes the cost of debt upward just as the Treasury refinances.

Asia changes the supply side

Fidelity International sees Japan, South Korea, Taiwan and mainland China using fiscal and industrial policy to convert technology export strength into domestic investment. The Asian AI cycle is shifting from an export cycle to a cycle of domestic reflation, with a policy mix that differs sharply from one market to the next.

The point for global capital is arithmetic. The Asian savings that for thirty years bought Treasuries and Bunds now find domestic uses with positive real yields and political backing. J.P. Morgan Asset Management^[2] already separates, for Asian earnings in the second quarter of 2026, the AI drivers from the traditional ones. The distinction matters because AI has entered company accounts, and therefore domestic demand for capital.

Less Asian savings flowing out means more term premium on Western paper: it is the 1847 mechanism read from the supply side.

Three implications for capital

1. Family offices and sovereign wealth funds: 36-month horizon

Long dollar duration stops being the default safe haven. The reallocation consistent with the pattern favours short maturities, physical gold and real assets in jurisdictions with excess domestic savings. The direction is the one already marked by the dollar's share of reserves, down from 71% in 2001 to 58% in 2024 according to the IMF's COFER data.

2. CEOs, boards and CFOs: 18-month horizon

The narrative that "rates will fall and the cost of capital with them" risks proving false within 18 months. Anyone bringing investors a capex plan built on a falling cost of debt is assuming a cycle inside a regime. The geopolitical risk missing from strategic plans is the competition for capital with governments.

3. Chief Risk Officers: 12-24 month horizon

The scenario missing from VAR models is an orderly, persistent sale of Treasuries by Japanese holders. The term premium rises while policy rates fall. A model calibrated on the 2010-2020 decade assigns this scenario a probability close to zero, and that is the flaw, never the reassurance.

The position, and what would change my mind

My position is clear-cut: the long yields of 2026 are the price of the competition for capital, and central banks are spectators rather than directors.

Two pieces of evidence would prove me wrong. A compression of the US term premium below zero for two consecutive quarters, with TIC data showing Japanese holdings stable or rising. Or a downward revision of aggregate hyperscaler AI capex in 2027 guidance greater than 20%, which would remove one of the three demands from the equation.

Until then, the market has priced the cycle, and left out the structure.

The forecast and what to watch

Expected event: by 31 December 2027, US Treasury TIC data will show Japanese holdings of Treasuries below $1 trillion. The starting point is the level of roughly $1.1 trillion indicated by BlackRock. Confidence: Medium, 60 out of 100. Horizon: 31 December 2027, that is 471 days from today. Verification: TIC series "Major Foreign Holders of Treasury Securities", Japan line.

Kill signal: TIC data for October 2027, published in December 2027, with Japanese holdings at or above $1.1 trillion would refute the thesis.

What to watch:

• The term premium estimated by the New York Fed (ACM model) on the US 10-year: a persistent rise while policy rates fall confirms the thesis.
• The currency-hedged spread between the 10-year JGB and the 10-year Treasury: when the hedged domestic yield beats the US one, Japanese repatriation accelerates.
• Hyperscaler 2027 capex guidance in the January and February 2027 earnings reports: an aggregate cut above 20% takes pressure off the curve.