Circular AI Investment: What the BIS Numbers Reveal

In October 1847 the British government suspended the Bank Charter Act. English railway companies were calling in capital in instalments, and shareholders paid those instalments with bank credit. Demand for rails came from the same hands that financed it.

When the banks asked for cash, the circuit stopped within weeks. Nearly one hundred and eighty years later, the same shape returns, measured in supervisory tables.

Circular investment in artificial intelligence now has an official figure, and the figure describes a closed loop.

The figure the BIS put in a table

On 1 October 2026 the Bank for International Settlements published eight pages on circular ties among artificial intelligence firms. Between 2021 and 2025, 55.2% of capital flowing into AI firms came from other AI firms, as documented in BIS Bulletin 137^[1]. And 28.7% of the value of the deals these firms close targets another AI firm.

The paper carries the names of five economists from the monetary department: Jon Frost, Rudraksh Kansal, Kumar Rishabh, Vatsala Shreeti and Leanne Si Ying Zhang. Alongside the text, the BIS publishes the Excel file with the data behind the charts.

Anyone managing capital should download that file before reading the press commentary.

Financial outlet 24/7 Wall St. picked up the figure on 3 October 2026, with a headline on the AI money machine^[2].

The financier is also the seller

The decisive passage in the Bulletin concerns the overlap between capital and supply. 16.1% of AI-to-AI deals by count, and 46.4% by value, include a commercial relationship between the party investing and the party receiving the investment.

The mechanism runs like this. The compute provider invests in the lab; the lab uses that money to buy compute; the revenue returns to the income statement of the investor. The same euro appears three times: as capital, as spending and as turnover.

The authors acknowledge the economic logic of these ties: they serve to lock in critical inputs and to reduce information asymmetries between those who invest and those who build. In the same paragraph they flag the price: macroeconomic risks and growing opacity.

That opacity has a practical effect on supervision. A supervisor looking at consolidated accounts sees demand; the relevant figure lives in the related-party disclosures, and it arrives months late. The BIS asks supervisors to look there.

Three precedents make a pattern

In Tokyo, between 1985 and 1989, cross-shareholdings between Japanese banks and industry held the equity market's free float in place. The Nikkei touched 38,915 points on 29 December 1989 and saw its cycle low around 7,000 points in March 2009.

In New Jersey, between 1998 and 2000, Lucent Technologies lent money to the telephone operators that bought its equipment, and booked those sales as revenue. The stock lost over 90% of its value between 1999 and 2002.

In all three cases the sequence matches. Internal financing inflates reported demand, and reported demand justifies new debt. External credit arrives last and leaves first.

Three precedents are enough to call it a pattern.

Cycle or structure

The distinction weighs on allocation. Capital circularity belongs to the cycle, with a historical duration of between three and seven years, and it reabsorbs once external customers pay in cash. Compute concentration belongs to the structure, measures in decades and outlives the reabsorption.

The market treats the two as one thing, and prices both as growth. The BIS separates them, and the separation changes the job of anyone writing risk models. The International Monetary Fund treats artificial intelligence investment as a line item in global growth; the BIS shows how much of that line runs in a circle.

A portfolio built on the cycle buys the dip and waits. A portfolio built on the structure buys fabs, energy and lithography, and accepts a ten-year horizon.

Three implications for capital

The BIS measurement translates into three moves with different horizons.

1. Revenue quality: measure the share of turnover that comes from related counterparties.
2. Private credit: map portfolio exposure to data centre debt.
3. Physical sovereignty: fabs, lithography and energy remain the bottleneck.

The first move concerns revenue quality: ask every portfolio company what share of turnover comes from customers who are also shareholders or suppliers. The figure exists in the contracts and is missing from the press releases. Whoever obtains it before the rest gains six months of advantage.

The second concerns the private credit financing data centres, where the risk concentrates far from equity valuations. A Chief Risk Officer measuring AI exposure with the beta of tech stocks is looking in the wrong place.

The third concerns the story chief financial officers take to investors. Growth described as market demand, and composed half of related counterparties, ages badly within eighteen months.

The three moves carry different horizons: twelve months for the first, twenty-four for the second, thirty-six for the third.

Where this desk stands

AI capex risk lives in the credit that funds it, and Bulletin 137 makes it countable. The market prices valuations; the fragility sits in the balance sheets of those lending money to compute centres.

A counterargument exists, and it deserves respect. Circular relationships serve to secure scarce inputs: the party investing in the customer secures the supply of chips, energy and capacity, as the integrated steelmakers of the twentieth century did. This reading holds as long as final demand pays in cash.

Two data points would change this position. A share of AI-to-AI investment falling below 40% in the next BIS update would change the picture. Together with double-digit growth in revenue from third-party customers, it would describe a sector walking on its own legs.

The prediction

Prediction: by 30 September 2027 either the BIS or the Financial Stability Board publishes a second paper on bank exposure to AI firms and data centres. The paper calls for dedicated supervisory data. Confidence: 72%. Horizon: 358 days.

Verification indicator: the publications section of the BIS website and the FSB report calendar. Falsification signal: 30 September 2027 arrives with zero BIS or FSB publications on credit exposure to AI firms and data centres.

What to watch:

Three indicators hold up or dismantle this reading over the next twelve months. They need to be read together, because each one taken alone comes out ambiguous.

• The share of revenue at large compute firms that comes from customers they hold stakes in, quarter by quarter.
• Spreads on private credit tied to data centres, compared with generic high yield.
• The next BIS update on the share of AI-to-AI investment, expected above or below 40%.

The divergence between reported demand and cash collected always resolves. The question is how, and on which balance sheet. Anyone reading Bulletin 137 now has a year of margin over the risk models that still ignore circularity.