BIS Counts Whole Rounds in Its 46.4% Circular AI Financing Estimate

The Bank for International Settlements says 46.4% of disclosed AI-to-AI deal value from 2021 through 2025 involved firms that also traded with each other. The figure is striking. Its construction is broader than a measurement of money returning to an investor through sales.

One commercially linked AI investor can qualify a financing round with several investors. The BIS then counts the full disclosed value of that round, not the linked investor's contribution. A commercial tie observed anywhere in the five-year window is enough, even if it did not exist when the investment closed.

That makes 46.4% a measure of market-structure overlap. It is not an estimate of circular cash flow.

The published result reproduces only at the graph level

The BIS Bulletin starts with 1,246 AI firms across compute, infrastructure, data tools, models and applications. PitchBook financing data and FactSet commercial-link data produce 972 intra-AI investment relationships for 2021 through 2025.

Among AI-to-AI deals, 16.1% by count meet the BIS circularity rule. Those deals represent 46.4% of disclosed value. The denominator matters: this is not 46.4% of all AI financing. It is the disclosed value of AI-to-AI deals in the study.

The accompanying graph workbook reproduces 16.1% and 46.4%. It does not publish the underlying deal list needed to rebuild either share. The 1,246-firm universe and 972 relationships appear in the paper, not as auditable rows in the workbook.

The gap between the count share and value share still reveals something. Dividing each published value share by its deal-count share implies that a circular-classified deal was, on average, about 4.5 times the disclosed size of another AI-to-AI deal. That calculation uses the BIS classification and full-round values. It does not show whether many large rounds or a few exceptional ones create the difference.

No concentration table answers that question. The public files provide no top-five share, deal-value distribution or investor allocation. A claim that a handful of mega-rounds explains 46.4% would therefore outrun the evidence. So would a claim that the result is broadly distributed.

One test spans five years, another counts every dollar

The BIS defines an investment relationship as circular when the firms also have a supplier-customer relationship at any point from 2021 through 2025. This is a relationship-level test across a window, not a deal-date test.

The distinction can change classification. A 2021 investment can qualify because the same pair traded in 2025. That may identify a durable strategic link, but it cannot establish that the investment financed the contemporaneous purchase of the investor's product.

The second issue is allocation. The paper says a multi-investor deal enters once. If any AI investor has a commercial relationship with the AI target, the entire disclosed transaction value enters the circular numerator. The authors say they do not observe how the round's value was distributed among its investors or, in some deals, among targets.

A stricter sensitivity would require two joins. First, each commercial relationship needs a start and end date that can be matched to the deal date. Second, the round needs investor-level contribution amounts. Neither field is in the public workbook, and the paper says the second is not observed. The requested deal-date and linked-check re-estimate is therefore not calculable from the released material. Its honest output is unavailable, not a lower percentage invented from incomplete disclosure.

This also separates three economic structures that the headline number combines. In the first, a supplier finances its customer, so capital and goods move in the same direction. In the second, a customer finances its supplier, so the flows run in opposite directions. In the third, the firms supply each other.

The paper reports that 64% of circular relationships involve an investor that supplies the target. That first type can support a revenue-recycling concern. The other types can reflect input security, bargaining power or reciprocal specialization. The release does not publish the counts needed to divide the remaining 36% between customer-finances-supplier and reciprocal structures.

Upstream concentration is visible, cash recycling is not

The BIS also reports that 73% of circular investment relationships originate with compute or infrastructure firms. This composition does reproduce from the workbook's relationship counts. Compute investors account for 45 relationships and infrastructure investors for 54, out of 136 plotted relationships. The result is 72.8%, rounded to 73%.

Those 136 observations are not the 972 total relationships and are not a deal-value total. A deal with several circular AI investors contributes several relationships to this graph. The result locates overlap upstream, where chips and cloud capacity are scarce. It does not quantify how much of a supplier's investment later became revenue.

The bulletin compares AI with a 2006 Journal of Finance study. That paper examined more than 10,000 US customer-supplier relationships from 1988 through 2001. It found a customer held a disclosed block of at least 5% in its supplier in 3.31% of the 10,211 relationships with ownership data.

That benchmark is useful evidence that direct customer ownership of suppliers was uncommon in a broad historical sample. It is not an apples-to-apples baseline for the BIS result. The older study tests only customer stakes in suppliers, uses a 5% ownership threshold and excludes supplier-finances-customer structures. The BIS includes equity and debt investment, all three commercial directions and a five-year relationship window. Comparing 3.3% with 16.1% or 46.4% as if they measured the same unit would manufacture a trend.

The policy implication survives the measurement limits. Financing and commercial exposures are concentrated in the same upstream network. A target's failure can damage both an investor's stake and its future sales, while a supplier shock can impair the customer that financed it.

Supervisors and investors still need a narrower table before treating 46.4% as circular cash. Deal dates should be matched to dated contracts. Linked investors' checks should be separated from the rest of each round. Purchase commitments, guarantees and disbursements should be reported apart from headline transaction values. Until then, the BIS estimate maps where financial and commercial relationships overlap. It does not trace where each dollar went.