The AI Boom Is Partly Funding Itself: BIS Finds 55.2% of Investment Into AI Firms Comes From Other AI Firms

Economists at the Bank for International Settlements have put a number on the accusation that the AI boom is partly a closed loop: between 2021 and 2025, 55.2 percent of the investment value flowing into the AI companies they studied came from other AI companies. The figure comes from BIS Bulletin No. 137, "Circular relationships among AI firms," published on October 1 by a team led by Jon Frost.

The sample covered 1,246 AI companies across five layers of the AI supply chain — computing, infrastructure, data tools, modeling and applications — and identified 972 investment relationships in which one AI firm put money into another. The comparison the authors offer is stark: across a broad sample of US companies, equity stakes held by a business partner account for just 3.3 percent of customer-supplier relationships, and in computing and related infrastructure the figure is 15.2 percent.

The striking part is not how many of these deals there are, but how big they are. Transactions that combine an investment with a commercial supply-chain relationship account for only 16.1 percent of AI-to-AI deals by count, yet 46.4 percent by value. On the outgoing side, 28.7 percent of the deal value AI companies deployed went to other AI firms. In 73 percent of circular relationships the investor is a computing or infrastructure company, and the most common arrangement — 64 percent of cases — is a supplier investing in a customer that buys its products.

Nvidia and CoreWeave are the study's emblematic example: Nvidia holds a stake in the compute provider, sells it chips, and has committed to repurchase up to $6.3 billion of unused capacity by 2032. Similar structures run through the sector, from hyperscalers that fund model labs and sell them cloud capacity, to chipmakers extending credit for their own silicon.

The authors are careful not to call these arrangements irrational. A supplier can observe how intensively its hardware is actually being used and how fast a customer is growing — information an outside lender would have to pay for — and financing a customer also stabilizes demand for its own products. Going the other way, customers invest in suppliers to lock in scarce inputs such as high-bandwidth memory during shortages.

The risk is that revenue becomes harder to read. When a supplier finances its customer, part of the supplier's growth is funded by its own capital, and analysts, lenders and regulators cannot easily separate organic end-user demand from demand the sector created for itself. The bulletin points to the late-1990s telecom bubble, when equipment vendors such as Lucent and Nortel lent operators money to buy their own gear — and were left holding both worthless loans and collapsing sales.

If expected revenue fails to materialize, a supplier that is also a shareholder in its customer loses twice: the equity depreciates and the sales decline. Private credit and special-purpose vehicles built to finance data-center infrastructure add leverage that is difficult to measure, and because circular exposure is concentrated in a small number of large companies upstream, shocks could propagate simultaneously through commercial and financial channels.

None of this makes the structures illegal, and the BIS does not call for a ban. What it does argue is that disclosure has not kept up with the loop: clearer reporting on supplier investments, supply commitments and cross-holdings would let markets price AI demand with far more confidence than they can today.