Over Half of A.I. Investment Comes From Other A.I. Firms

When Nvidia announced in September 2025 that it would invest $100 billion in OpenAI to help fund data center construction, Bernstein analyst Stacy Rasgon wrote that the action would "fuel 'circular' concerns."

A year on, the central bank for central banks has done the research.

More than half of the money invested in artificial intelligence companies between 2021 and 2025 came from other A.I. companies, according to economists at the Bank for International Settlements. Of the value of deals among A.I. 46.4% went to companies that were also trading partners of the investor, the bank said in a report published on Oct. 1. Overall, 55.2% of the money flowing into A.I. firms came from other A.I. firms.

The findings come as the industry's biggest backers negotiate how to pay for the next round of buildout.

Nvidia has agreed to work with Apollo Global Management, BlackRock, Blackstone, Brookfield Asset Management, Goldman Sachs and KKR to backstop more than $500 billion in third-party capital for A.I. infrastructure and data centers, media reports said. Nvidia may backstop up to $125 billion, or one-fourth of potential deals, these reports added. The backstop would apply if asset values fall by more than expected.

Part of that financing is a guarantee for an OpenAI data center project in southern Ohio. Nvidia would guarantee less than $120 billion, The Wall Street Journal reported on Aug. 14. An SEC filing three days later put the figure at up to $105 billion. The Journal also reported that the guarantee had been scaled back from an initial $250 billion after investors raised concerns about Nvidia's risk exposure. Nvidia was also reportedly considering a separate deal to finance OpenAI's purchase of $350 billion in Nvidia chips.

The sums are large against what the BIS expects the industry to spend.

It says the largest A.I. firms collectively plan to spend hundreds of billions of dollars on A.I. infrastructure in 2026, and that A.I. capital spending has become macroeconomically significant in many economies. Nvidia, the supplier at the center of the reported talks, has a market capitalization of about $5 trillion.

Economists call the pattern circular financing. A chipmaker takes a stake in a customer, and the customer spends part of what it raised on the chipmaker's chips. Or a customer buys into a supplier whose output it cannot do without.

Jon Frost and his four co-authors wrote their definition around the relationship between two firms, not around a single transaction. A pair counts as circular if one invested in the other between 2021 and 2025 and the two also had a supplier-customer relationship at any point in that window.

They started with 1,246 A.I. firms, sorted into five layers of the supply chain. Compute, infrastructure, data tools, models and applications. Deal records came from PitchBook and commercial links from FactSet, and the economists say they added extensive manual checks to catch relationships the databases missed. The search turned up 972 investment relationships between A.I. firms.

All in the Family

Of the deal value A.I. firms put out as investors, 28.7% went into other A.I. firms. Of the deal value A.I. firms took in, 55.2% came from other A.I. firms. The count and the dollars tell different stories on circularity. Only 16.1% of A.I.-to-A.I. deals by number involved a commercial link between the two sides. By value the share is 46.4%, implying that the so-called "circular financing" deals are much larger than the rest.

Of the circular relationships, 64% involved an A.I. firm that invests in a company and also sells to it.

“These numbers suggest that a significant portion of financing and commercial relationships in AI are self-referencing.” — BIS Bulletin No 137

The Logic

The authors do not treat the practice as a scandal, and they lay out why companies do it.

A supplier sees how hard a customer is using its product, which gives it a better read on that customer's growth than a bank has. A model developer burns through compute long before it earns commensurate revenue, and the chipmaker or cloud provider selling the compute can watch the consumption in real time. Financing the customer also keeps demand for expensive products steady. The theory the BIS economists lean on is Burkart and Ellingsen's 2004 work on trade credit.

The reasoning for the financing is also about scarcity.

When only a few firms make an input and the buyer cannot make its own product without it, an equity stake can secure supply. It can also give the buyer a say in whether the supplier sells to its competitors. The BIS points to acute shortages of memory chips as the sort of bottleneck that invites this.

A third explanation covers inputs built for one customer. Data centers are configured to a client's workloads with chips designed, manufactured and delivered to a particular model. Once the money is sunk, either side can try to renegotiate in its own favor. A loan or a partial ownership stake aligns incentives, and the BIS calls it a partial substitute for full vertical integration. Similar examples appear in biotechnology, telecommunications and manufacturing.

What sets A.I. apart is how many of those conditions exist in one industry at once. It has startups whose prospects are hard to judge, customized inputs, a handful of providers for critical equipment such as photolithography machines and specialized chips, and financing needs so large that outside capital may be costly or unavailable.

Outside A.I., equity ties between trading partners are uncommon.

In a sample of more than 10,000 U.S. customer-supplier relationships, Fee, Hadlock and Thomas found direct equity stakes in 3.3% of them. In A.I. compute, cloud and related infrastructure, 15.2% of customer-supplier relationships also had financing relationships.

The ties get closer as you move upstream. Compute and infrastructure suppliers originate 73% of circular investment relationships. In the data tools and models layers, where inputs are closer to commodities, they are rarer.

What Lies Beneath

The BIS identifies three main risks.

The first is demand that is partly of the industry's own making. When a supplier finances a customer, some of the supplier's revenue growth traces back to its own capital rather than to final demand. Investors, lenders and supervisors then struggle to tell how much of the boom is organic.

The authors looked back to the late 1990s for an example.

Lucent and Nortel financed network operators so the operators could buy their equipment, and the vendors booked both the sales and the loans as assets. When operator revenues failed to materialize or slowed, the operators could neither repay the loans nor keep buying. The vendors lost money and sales together.

Suppliers financed their customers on a large scale in that boom. Before the 2001 crash, Cisco's and Nortel's customer financing exceeded 10% of their annual revenues, and the top five telecom equipment makers' financing exceeded 123% of their combined earnings, according to Fortune.

Such dynamics, the authors write, "may also play out in AI if revenue growth and end user demand fall short of firms' expectations." They add that with A.I.-driven equity valuations already high, the result could be significant financial market volatility.

Second, the counterparty risk is doubled. An investor that also supplies its target loses the value of its stake and its future product sales if the customer stumbles. Because circular deals are concentrated among a small number of very large firms in the upstream layers, a shock could travel through commercial and financial channels at the same moment. Private credit and special purpose vehicles, increasingly used to fund A.I. infrastructure, can add hidden leverage on top.

The third risk is not immediately visible. Many of the firms are private and disclose little. Even listed ones strike deals that mix cash with long-term purchase commitments and guarantees on asset values.

Residual value guarantees are one example. The guarantor pledges to cover any shortfall in what an asset, such as chips or data center equipment, is worth after a fixed period. Those commitments sit off the balance sheet and come due in a downturn, when the guarantor is least able to pay. The firms also span sectors and jurisdictions, so no single supervisor sees the whole web. Reported deal values may not capture the full picture, and headline figures can differ substantially from the amounts actually disbursed.

Who Holds the Bag?

The concentration of S&P 500 gains in a narrow set of A.I. and A.I.-adjacent firms has created a systemic vulnerability for the global financial system. Passive investing amplifies it, as does the Federal Reserve's tightening response to A.I.-driven inflation pressures that have arrived well ahead of broad productivity benefits.

A group of 42 A.I. stocks have driven 67% of the S&P 500's price gains since the start of 2024, according to The Kobeissi Letter, a market commentary platform. They contributed 22 percentage points of the index's 33% advance. The other 458 stocks added 11 points. The same group accounted for 58% of earnings growth and 81% of growth in capital expenditure and research and development.

The Magnificent Seven made up 34.9% of S&P 500 market capitalization in October 2026, according to The Motley Fool. A.I. stocks make up 49% of the index, the second-highest concentration on record. The top 10 stocks account for about 40%, and concentration at these levels has not been seen in decades outside extremes such as the late-19th-century railroad boom.

Torsten Slok, chief economist at Apollo, wrote on June 8 that "there are not yet signs that AI is boosting revenues or profit margins outside the Mag 7," and that profit margins are flat for the S&P 493, the index excluding the Magnificent Seven. Current prices therefore embed very high expectations for diffusion that have yet to materialize broadly.

Passive funds held 55% of U.S. fund assets at the end of 2025, up from 53% a year earlier, with $19.4 trillion in passive mutual funds and exchange-traded funds, against $16 trillion in active funds. U.S. retirement assets reached $51.2 trillion on June 30, according to the Investment Company Institute, including $15.0 trillion in defined contribution plans and $19.9 trillion in individual retirement accounts. A large share of those plans, and of target-date funds, is benchmarked to or heavily weighted toward the S&P 500 and similar market-cap indexes. Households held a record 48.23% of financial assets in equities in the second quarter, up from 44.66% in the first, according to Federal Reserve data.

Pension liabilities, retirement savings and other set-it-and-forget-it capital are therefore structurally long a handful of highly correlated, richly valued A.I. names. A sharp de-rating would reach household wealth, corporate pension funded status and institutional balance sheets with few natural hedges, whether it came from slower-than-expected monetization, competitive disruption, regulation, energy constraints or a simple rotation.

The Federal Reserve raised the federal funds target range by 25 basis points on Sept. 16, to 3.75% to 4.00%, in a 12-0 vote. Chase's summary of the meeting cites energy prices as the main driver, with crude above $100 a barrel in early September amid the Iran conflict. Fed Chairman Kevin Warsh told an International Monetary Fund meeting in Thailand that A.I. investment is a "shock" to prices and demand, with wealth and productivity gains "over the horizon."

The buildout adds to near-term price pressure through demand for chips, power, construction and related goods, and higher rates weigh on the non-A.I. majority of the economy and on leveraged or rate-sensitive parts of the market while the mega-cap A.I. cohort has so far held up. Policy is aimed at cooling the broader economy to offset A.I.-related inflation, yet the equity market, and the passive and pension capital tied to it, stays heavily exposed to the sector driving those pressures.

The boom also feeds itself. Capital spending funded by strong cash flows and easy access to equity markets supports further investment and higher valuations. Energy and physical-infrastructure limits could cap how fast the buildout scales, and extreme market-cap concentration has often preceded relative underperformance or volatility for the leaders.

Sources

• Frost, J, Kansal, R, Rishabh, K, Shreeti, V and Zhang, L S Y (2026): "Circular relationships among AI firms", BIS Bulletin, no 137, 1 October.
• Burkart, M and Ellingsen, T (2004): "In-kind finance: a theory of trade credit", American Economic Review, vol 94, no 3. Cited via BIS Bulletin No 137.
• Fee, C, Hadlock, C and Thomas, S (2006): "Corporate equity ownership and the governance of product market relationships", Journal of Finance, vol 61, no 3. Cited via BIS Bulletin No 137.
• OECD (2003): "After the telecommunications bubble", OECD Economic Outlook, no 73. Cited via BIS Bulletin No 137.
• Rungcharoenkitkul, P (2026): "The AI investment race", BIS Working Papers, no 1367. Cited via BIS Bulletin No 137.