The Industry That Funds Itself: BIS Bulletin 137 Puts Hard Numbers on Circular AI Financing

For two years, “circular financing” has been the AI boom’s favorite dirty word — invoked every time a chipmaker invests in a lab that turns around and buys the chipmaker’s silicon. The debate has run on anecdotes: a $100B commitment here, a credit facility there, stitched together by analysts into a narrative reminiscent of the telecom vendor-finance era. On 1 October 2026, the Bank for International Settlements replaced anecdote with arithmetic. BIS Bulletin No 137, “Circular relationships among AI firms,” is the first systematic central-bank measurement of how much of the AI investment complex is actually funding itself — and the headline number is stark: 55.2% of all investment that flowed into AI firms between 2021 and 2025 came from other AI firms.

The measurement nobody had done

The five-author team — Jon Frost, Rudraksh Kansal, Kumar Rishabh, Vatsala Shreeti and Leanne Si Ying Zhang — built the dataset from scratch. They classified a universe of 1,246 AI firms into five supply chain layers (compute, infrastructure, data tools, models and applications), extracted deal-level financing information from PitchBook, mapped commercial supplier-customer relationships from FactSet, and ran extensive manual checks to make sure no links were missed. The result: 972 distinct intra-AI investment relationships formed between 2021 and 2025.

Crucially, the BIS defines circularity more carefully than the media does. Rather than asking whether a single deal bundles financing with a purchase commitment, the Bulletin calls a relationship circular whenever two AI firms that share a supplier-customer link at any point in the five-year window also hold an investment relationship — in either direction, by equity, debt, M&A or joint venture. That broader lens captures three structures: a supplier investing in its customer (capital and goods flow the same way), a customer financing its supplier (they flow in opposite directions), and reciprocal arrangements where goods flow both ways.

The findings, measured across that relationship graph:

• 28.7% of AI firms’ outbound investment deals (by deal value) targeted other AI firms — the ecosystem buys itself on the way out, too.
• 55.2% of incoming investment in AI firms came from other AI firms.
• 16.1% of AI-to-AI deals by count — but 46.4% by deal value — also involved a commercial supply chain relationship between investor and target. The circular deals are the big deals.
• 64% of circular relationships follow the type-1 pattern: the investor also supplies its target. The vendor-financing shape dominates.

One comparison in the paper quietly lands the hardest. In a broad sample of more than 10,000 US customer-supplier relationships, economics research (Fee, Hadlock and Thomas, 2006) found direct equity stakes in trading partners accounted for just 3.3% of relationships. In the AI compute-and-cloud submarket, 15.2% of customer-supplier relationships also carried financing ties — roughly five times the norm. Whatever this is, it is not ordinary industrial practice.

Why AI does this to itself

The Bulletin is not a polemic; most of its pages are devoted to explaining why these structures are economically rational. Suppliers make unusually efficient lenders because they can observe their customers directly — a chipmaker can see a model lab’s compute consumption and growth trajectory far more cheaply than an outside bank ever could (the classic trade-credit argument of Burkart and Ellingsen). Financing a customer also stabilizes demand for the supplier’s own product, which matters when the product is a GPU cluster and the customer needs extraordinary capital before generating any revenue.

The reverse direction — a customer financing its supplier — is about securing scarce inputs. In a market where memory chips and advanced lithography have only a handful of providers, an equity stake is both an allocation guarantee and a seat at the table on strategic decisions, including who else gets supply. And because AI inputs are deeply relationship-specific (data centers configured to particular workloads, chips tailored to particular models), firms face “hold-up” risk on every sunk investment; partial ownership aligns incentives the way full vertical integration would, at lower cost. The BIS notes the same patterns in pharmaceuticals, biotech, Japanese keiretsu manufacturing and — pointedly — 1990s telecommunications.

The geography of the data confirms the theory: 73% of circular investment relationships originate from firms in the compute and infrastructure layers, where inputs are most critical and most customized. Downstream layers, where inputs are closer to commodities, show far less overlap.

Three macro risks, one Lucent ghost

Where the economics ends, the financial-stability analysis begins, and the Bulletin is blunt about three transmission channels.

First, demand becomes partly endogenous. When a supplier finances a customer, part of the supplier’s reported revenue growth is nothing more than its own capital coming home. The parallel the BIS draws is exact: Lucent and Nortel financed network operators in the late 1990s so operators could buy Lucent and Nortel equipment, booking both the sales and the loans as assets — until operator revenue failed to materialize and both sides of the ledger collapsed at once. “Such dynamics may also play out in AI if revenue growth and end user demand fall short of firms’ expectations,” the authors write, warning of significant volatility amid AI-inflated global equity valuations.

Second, double exposure amplifies contagion. An investor that is also a supplier is exposed to the same counterparty twice — a shock to the customer simultaneously writedowns the equity stake and kills future product revenue. Because circular deals concentrate among a small number of very large upstream firms, adverse shocks propagate through commercial and financial channels simultaneously, with private credit and special purpose vehicles adding hidden leverage on top.

Third, opacity defeats supervision. Many of the firms are private and disclose little; even public ones bury terms in structures that mix cash with long-term purchase commitments and residual value guarantees (RVGs) — contingent obligations that sit off balance sheet and only materialize in a downturn, precisely when the guarantor can least absorb them. Deal values that make headlines, the BIS cautions, may differ substantially from disbursed amounts. And because the firms span sectors and jurisdictions, no single supervisor can see the whole web.

The week’s timing was not subtle

The Bulletin landed the same week the archetype deal surfaced in public filings: Reuters reported on 1 October that Broadcom has agreed to lend Anthropic up to $42 billion via a convertible note — about one-third of Anthropic’s $125.2 billion, five-year TPU commitment — deepening a relationship that already spans compute supply, equity and now vendor credit. The BIS paper does not name the parties, but it did not have to. It also builds on the BIS’s own running tally of alarm: Working Paper 1367 estimated the AI race drives over-investment roughly 50% above the socially efficient level, and the 2026 Annual Economic Report chapter — pointedly titled “Progress and peril” — flagged the trillion-dollar hyperscaler capex wave.

For investors, the actionable core of Bulletin 137 is a single reframing: when you read that AI demand is “skyrocketing,” ask whose capital is producing the sky. Until end-user revenue can carry the load on its own, the industry’s most important customer remains itself.

Bottom line

BIS Bulletin 137 converts the circular-financing debate from vibe to dataset. A majority of the money arriving at AI firms now departs from other AI firms; nearly half of that deal value travels between suppliers and their own customers; and the structure is concentrated, leveraged, opaque and self-reinforcing. None of this predicts a crash — vendor financing can persist for years while end demand catches up. But it tells supervisors, lenders and IPO buyers exactly where the fragility lives if it doesn’t.