Is AI financing becoming circular?
A Bank for International Settlements analysis maps 1,246 AI firms and 972 investment relationships, finding that commercial links accompanied 16.1% of AI-to-AI deals by count but 46.4% by disclosed value. The pattern is consequential, yet it does not show that the transactions were improper, unprofitable or certain to spread financial stress.
The short version
- What happened
- The BIS assembled a universe of 1,246 AI firms across five supply-chain layers and identified 972 intra-AI investment relationships during 2021–25.
- Who it affects
- AI users may benefit when strategic financing expands scarce computing capacity, but dependence on a small set of linked providers can reduce choice and resilience.
- What remains unproven
- BIS Bulletin 137 is a technical staff publication, not a peer-reviewed journal article, supervisory judgment or allegation of misconduct.
Keep in mind
BIS Bulletin 137 is a technical staff publication, not a peer-reviewed journal article, supervisory judgment or allegation of misconduct.
PitchBook and FactSet are commercial datasets; private deals and contracts may be missing, while 2025 deal records are described as provisional and incomplete.
A relationship is circular when financing and a commercial link coexist at any point during 2021–25, so the definition does not prove that one caused the other or that funds paid for the investor’s products.
Full disclosed transaction value is used when a qualifying investor is present, even though the division of capital among investors and targets is not observed.
The descriptive analysis does not estimate defaults, investment returns, competition effects, consumer harm or the probability of a system-wide financial shock.
Source analysis · 5 linked sources. Primary analysed source date: 1 October 2026. Earliest linked supporting source: 23 February 2026.
Read the full report · 10 min The evidence, methods, implications and limitations
The BIS finds financing and supply links overlapping at scale
The Bank for International Settlements published a new map of financial and commercial ties among AI firms on 1 October. Its question is narrower than whether the AI investment boom is a bubble. The authors ask how often an investment between AI companies overlaps with a supplier-customer relationship and what that overlap could mean for interpreting demand, exposure and financial stability.
Their starting universe contains 1,246 firms classified into five layers of the AI supply chain: compute, infrastructure, data tools, models and applications. Investment information comes from PitchBook records covering equity and debt transactions. Commercial relationship data come from FactSet, supplemented by manual checks intended to reduce omissions. The study period is 2021 through 2025.
The matching exercise produced 972 intra-AI investment relationships. The authors call a relationship circular when an AI investor and an AI target also had a commercial supplier-customer link, in either direction, at some point during the same five-year period. This is a relationship-level definition rather than proof that money from a particular financing round was contractually tied to a purchase.
On that definition, 16.1% of AI-to-AI deals by count were circular, while those deals represented 46.4% of disclosed AI-to-AI deal value. The gap is central to the report: overlapping commercial and investment relationships were a minority of deals, but they were disproportionately associated with the larger disclosed transactions.[1][2][3]
The headline percentages describe different denominators
Several figures in the report can sound contradictory if their denominators are blurred. Among deals in which AI firms invested, 28.7% of disclosed value went to other AI firms. Looking from the target side, 55.2% of incoming disclosed investment value for AI firms came from other AI firms. Those percentages measure the AI-to-AI share of investment flows, not the share that also had a commercial link.
The 16.1% and 46.4% figures then examine only AI-to-AI deals and ask how many also met the report’s circularity definition. The first is a share of deal count; the second is a share of full disclosed deal value. A reader should not interpret either number as the percentage of all AI funding, all company revenue or all spending on computing infrastructure.
Direction also matters. The BIS reports that 64% of circular relationships involved an AI firm investing in a company to which it also supplied products or services. It also finds that 73% of circular investment relationships originated in the compute or infrastructure layers, where expensive and specialised inputs can make long-term customer and financing ties economically attractive.
These counts are relationships and deals, not independent firms. One company can appear repeatedly, and a multi-investor round can connect several parties. The report says a deal involving multiple AI investors enters the deal calculations once; if any AI investor also has a commercial relationship with the target, the full disclosed transaction value is classified as circular.[1][2]
Why suppliers and customers may finance one another
Overlapping ties are not inherently suspicious. A cloud or chip supplier may know more about a customer’s usage and growth than an outside lender does. Financing that customer can secure demand for expensive capacity while helping a young model or application business fund the infrastructure it needs before revenue catches up.
The reverse relationship can also make economic sense. A customer may invest in a supplier to obtain scarce chips, memory, data-centre capacity or other critical inputs. Where equipment is customised or only a few firms can provide it, financing can align incentives and reduce the risk that either side walks away after the other has made a costly, relationship-specific commitment.
The BIS’s earlier mapping of the AI supply chain is important context. Compute, infrastructure, data tools, models and applications are not interchangeable markets. Upstream capacity can be capital-intensive, geographically concentrated and difficult to substitute quickly. A commercial link accompanied by investment may therefore be a rational response to genuine bottlenecks rather than an attempt to manufacture demand.
That is why the report does not classify circularity as misconduct. It documents a structure and discusses mechanisms. Whether any individual deal improves competition, locks out rivals, distorts revenue or shifts risk depends on its terms, the parties’ alternatives and the actual flow of goods, services and cash—details the aggregate study cannot adjudicate.[2][4]
The financial-stability concern is double exposure and opaque demand
The concern begins when the same counterparty relationship carries both commercial and financial exposure. If a supplier invests in a customer and sells it computing services, disappointment at the customer can reduce the value of the investment and the supplier’s future revenue at the same time. What looked like two sources of growth can become one concentrated risk.
Circular structures can also make demand harder to read. A supplier’s sales may partly reflect capacity purchased by a customer that the supplier helped finance. That does not make the sale unreal, but it complicates the distinction between independent final demand and demand supported by capital flowing within the same ecosystem. Investors and supervisors may need more detail before treating booked revenue as evidence of broad end-user adoption.
The report points to the late-1990s telecommunications boom as a caution rather than a forecast. Equipment vendors financed network operators that bought their products; when operator revenue disappointed, vendors faced both financing losses and falling orders. AI markets differ in technology, contract structure and participants, so the analogy cannot establish that the same outcome will occur.
Related BIS modelling examines how competitive investment races, debt and circular stakes could amplify a downturn. That working paper is a calibrated theoretical exercise, not an observed stress test of the 972 relationships in the new Bulletin. Combining the two sources is useful for identifying plausible channels, but not for converting a structural map into a probability of crisis.[2][5]
Important measurement limits keep the findings descriptive
The report relies on two commercial databases and extensive manual checks, but private transactions and commercial contracts can remain undisclosed. FactSet may record that a supplier-customer relationship exists without showing its complete value, duration or exclusivity. PitchBook deal records can be provisional, and the report says 2025 information is incomplete.
Disclosed deal value is also a blunt measure. For a financing round with several investors, the data do not show how the total was divided among them. The BIS therefore uses the full disclosed transaction value when a qualifying circular connection is present. This can overstate the amount attributable to the investor that also has a supply relationship.
Timing is broad. A commercial relationship at any point from 2021 to 2025 can make an investment relationship circular even if the supply link and financing did not begin together. That choice helps capture strategic relationships, but it cannot prove that one caused the other or that investment proceeds were spent on the investor’s products.
The analysis does not report a counterfactual group showing what would have happened without circular links, nor does it estimate default probability, return on investment, market power or consumer harm. It identifies concentration and possible transmission channels. Claims that these deals are already a bubble, a cartel or a hidden failure would go beyond the evidence.[2][3]
What the findings mean for people, firms and regulators
For investors and lenders, the practical lesson is to look through headline deal values. Due diligence should separate cash actually contributed by each party, purchase commitments, cloud credits, equipment guarantees, leases, special-purpose vehicles and any conditions tying financing to future orders. The risk sits in the contract architecture, not in the word circular alone.
For AI customers, concentrated financing and supply ties may have two opposing effects. They can bring infrastructure online faster and help new services reach users. They can also reduce choice if a model developer becomes dependent on one cloud, chip or data-centre provider. Switching costs and the ability to move workloads matter as much as the announced investment amount.
For workers and communities hosting data centres, a demand reversal could affect construction, energy planning and employment even when the underlying technology remains useful. Conversely, assuming every related-party deal is fragile could deter capacity that supports productive applications. Better disclosure helps policymakers distinguish genuine build-out from exposures that may amplify stress.
The next evidence should be more granular and longitudinal. Useful follow-up would identify investor-level contributions, contract-linked purchases, maturity and collateral, revenue generated outside the financing network, and how relationships change under weaker demand. Stress testing across firms and jurisdictions would show whether the mapped links transmit losses or simply reflect ordinary strategic partnerships.[1][2][3][5]