The debt behind the AI boom, read every evening

The market we watch

Last price, change on the day and distance from the one-year high.

The AI stocks, layer by layer

Last year, 100 a year ago, weekly close; each layer is the equal-weight average of its stocks, against the S&P 500. Chips: Nvidia, AMD, Broadcom, TSMC, ASML, Micron, SK Hynix, Arm, Marvell. Hyperscalers: Microsoft, Alphabet, Amazon, Meta, Oracle. Neoclouds and servers: CoreWeave, Nebius, IREN, Supermicro, Dell, Arista. AI software: Palantir, ServiceNow, Salesforce, Snowflake.

Where each AI stock stands

The 24 AI stocks we follow are — below their one-year high on average. The layer furthest from its high is — (—), the closest — (—).

Change over one and three months and distance from the one-year high. SK Hynix trades in Seoul, in won.

Map of the AI debt

Billions of dollars, from the latest filings with the SEC; for OpenAI and Anthropic, which are not listed, from prospectuses and the press. Free cash is operating cash minus capital spending, over the period shown. The numbers are re-read from the SEC every evening after a new quarterly report; a model reads the filing for the leases already signed but not yet on the balance sheet, mostly data centres, the largest hidden commitment of the build-out. Type, rating, risk and notes are written by Claude, a model made by Anthropic.

Who holds the AI debt, and who pays if it breaks

The risk does not stay with the companies building the data centres. It ends up with those who need long, safe yields: pension funds and life insurers. It gets there by three roads: long bonds, where the hyperscalers are now over a fifth of the index; private credit; and off-balance-sheet vehicles such as Meta's Hyperion. Vanguard and BlackRock mostly manage clients' money: they lose fees, the real losses fall on savers. Blackstone is the most exposed on its own account, because it owns and finances data centres.

From public sources of September and October 2026. The figures measure different things (total private credit, data-centre portfolios, index weights), so they do not add up. How much data-centre debt pension funds hold is published by no one: it is the most opaque point.

Four bubbles lined up at their peak

AI is — months past the peak, at —. At the same month the Nasdaq Telecom stood at —, the Nasdaq at —, the railroads at —. Three years after its peak the Nasdaq Telecom was worth —.

Each curve is 100 in the month of its high; five years before on the left, three years after on the right. US railroads: index of railroad stocks 1857-1929 (NBER, Macaulay 1938). Dot-com: Nasdaq. Cables and telecom: Nasdaq Telecom, the index of the companies that laid the fibre. AI: the US chip index (SOX), at its June 2026 high for now.

Before the peak AI tracks the Nasdaq and telecom almost evenly, but the final climb was steeper than both. After the peak it falls in step with cables and telecom, the bubble in which suppliers financed their customers and the debt sat with those building the network, as today. The railroads are the useful exception: after their 1909 high they did not crash, they stayed flat for years, because they were companies that collected cash. That is the open question about AI: how much of the revenue is real cash.