Summary

  • BIS Bulletin 137 classifies 46.4% of disclosed AI-to-AI investment deal value as “circular” because investor and target firms also had a supplier-customer relationship during 2021–25.
  • The measure assigns the full disclosed value of a multi-investor round when any AI investor has that link; it does not say how much money was recycled or supplied by that investor.

The headline percentage is large enough to invite a simple story: AI companies finance one another, then the resulting purchases make demand look stronger than it is. The Bank for International Settlements’ Bulletin 137 supports a narrower, more useful finding. Across 1,246 firms grouped into compute, infrastructure, data tools, models and applications, the authors identify 972 investment relationships between AI firms over 2021–25. Of disclosed AI-to-AI deal value, 46.4% fell into their “circular” category; the corresponding share by deal count was 16.1%.

Here, circular describes a relationship map, not a cash route. An investor and target qualify if they had a commercial supplier-customer relationship in either direction at some point in the five-year window. The commercial tie need not have existed when the investment was made. The definition is therefore broader than a financing-round test that requires an investor’s purchase commitment to be part of that same deal.

The valuation rule matters just as much. If a round has several AI investors and any one has a commercial link to the target, the study classifies the entire disclosed round value as circular. It does not allocate the round among investors. So 46.4% is not the connected investor’s share, the value of purchases, a measure of cash returned to investors, or evidence that end customers funded the cycle.

The study still reveals an important concentration. Its chart attributes 73% of circular investment relationships by count to firms in compute or infrastructure. That is a relationship count, not 73% of value. Upstream firms may invest to secure scarce inputs; suppliers may invest in customers; or both sides may trade with one another. Each structure can help finance specific investment while making commercial demand and financial exposure less independent.

The authors combine PitchBook investment data with FactSet commercial links and say they manually checked the data to the extent possible. They also disclose that PitchBook analysts did not review the BIS calculations and that results may differ from PitchBook methodology. The bulletin describes opacity and correlated exposure as possible risks; it does not report a firm failure, realized loss, fictitious revenue or fraud. The evidence is a map of overlapping ties over 2021–25, not a stress test of today’s AI market.

Sources