The routes by which an AI setback could spread are on the public record. Who would bear the loss is not.

A widely read Fortune commentary argued in October that the US economy has not concentrated so heavily on one theme since the railroad era, and cited the BIS for the claim that more than half of investment in AI firms comes from other AI firms.

Concentration is not in doubt. The useful question is what happens if something goes wrong.

That question is narrower than the bubble debate, and more practical: if one part of the AI economy has a setback, by what route could it reach a company that is not in AI at all?

We traced it channel by channel, across funding, capacity, demand, valuation and credit, from 698 public sources: company filings with the SEC, central-bank and multilateral publications, and official statistics. Every source was retrieved and the specific claim confirmed on the page. The first thing that work did was refine the claim we started from. The BIS figure is 55.2%, from a staff bulletin covering 2021 to 2025 and counting full round values; it is not a measure of circular financing as such.

We specified 22 channels through which a setback could travel. Fifteen are well evidenced as to existence, seven are plausible but thinly evidenced, and four are opaque in whole or part. We mapped 87 financial and commercial relationships among 49 entities, 85 of them verified at source, and the most consequential appear in notes to the accounts rather than in announcements. One private, funding-dependent company that publishes no financial statements touches 57.5% of the disclosed relationship value.

The public record lets the obligor-side stock of off-balance-sheet and contingent obligations be summed at about US$1.48–1.53 trillion. That is a maximum, not a loss estimate, and most of it is the forward rent of hyperscalers, most of them highly rated.

So far, it has stayed contained

In July 2026 AI equities fell sharply, leverage amplified the fall and one leveraged AI-focused fund failed. There were no defaults, no bank losses and no pull-back in consumption, and valuations outside AI held. The IMF's global stress test found banks would remain above regulatory minimums under an AI-led correction. For the non-AI core of the mid-market we found no measured exposure to date.

Two things surprised us

Our view, for what it is worth

We went in looking at concentration and came out looking at debt. In the railroad busts, defaults reached about 36% of corporate bond par value in 1873–75 and the debt sat with banks; in the telecom bust of 2000–02 it sat with bondholders and vendors, and banks were largely spared. Where the debt sits decided how far each bust travelled. Today that is the fact the public record shows least clearly: lessors' notes, vehicle bonds, private-credit funds and the banks that lend to them. On the public record, we cannot say how far a larger setback would travel, and we doubt that anyone yet can. A board can still ask:

These are questions, not advice. The decision stays with the reader, and if the public record misses something you know, we would like to hear it.