On October 1, 2026, the Bank for International Settlements released an eight-page research brief titled “Circular Relationships Among AI Companies.” The report examines how money flows within the AI industry: a cloud service provider invests in a model company, which then uses the funds to purchase cloud services; a model company invests in a compute provider to secure future computing resources; and two companies both purchase from each other and hold equity stakes in one another.
This relationship has precedents in traditional industries, but AI has scaled it up significantly. Chips, data centers, cloud computing, and foundational models require continuous massive capital investment and long order cycles, making it difficult for suppliers and customers to build partnerships based solely on single purchase contracts. Equity stakes, loans, long-term procurement commitments, and compute reservations are now bundled into the same deals. Investment decisions drive orders, which in turn support revenue and valuation, gradually forming a network across the industry that connects both capital and commercial flows.
Money in the AI industry is beginning to circulate along the supply chain.
Figure 1 | Servers, networking, and cooling equipment form the foundational infrastructure for the earliest AI investments.
BIS compiled a list of 1,246 AI companies spanning five segments: computing power, infrastructure, data tools, models, and applications. It then integrated funding and investment data from PitchBook with supply chain relationships from FactSet, followed by manual verification of each entry. The study identified 972 internal investment relationships among AI companies between 2021 and 2025.
Among these, when AI companies act as investors, 28.7% of the investment funds flow to other AI companies; from the perspective of funding recipients, 55.2% of the investment received by AI firms comes from other AI companies. Capital is flowing frequently within the industry, with upstream giants, model companies, and infrastructure providers often appearing on both investor and customer lists.
The report defines a "circular relationship" as a connection between two AI companies that have both an investment relationship and a supplier-customer relationship during the same period. This definition focuses on enduring business relationships over a five-year period and does not require that a specific investment be explicitly tied to a specific purchase in a contract. It better reflects real-world industry networks, as companies may first make an investment and sign a procurement agreement months later, or begin collaboration first and later solidify long-term ties through equity arrangements.
This network will also reduce businesses' dependence on individual customers. The higher the revenue contributed by a customer, the more important ownership relationships and long-term procurement commitments become, and the more management needs to explain where future demand will come from and whether contracts can be upheld during industry downturns.
This will also affect how the market prices revenue quality and customer concentration.
II. How the Three Cyclic Relationships Are Formed
Figure 3 | Three Common Types of Circular Investment Relationships Among AI Companies
The first relationship is funded by the supplier. Chip manufacturers, cloud service providers, or data center operators invest in their own customers, who then purchase computing power, chips, or cloud services. Funds and products flow in the same direction. Suppliers gain earlier access to operational data and can observe actual usage and growth rates more directly than traditional financial investors; meanwhile, customers, after securing capital, are better positioned to sign long-term contracts.
The second relationship is funded by the client. Model companies or large tech firms invest in key suppliers, with capital flowing upstream; in return, chips and computing power services flow back downstream. The core here is resource assurance. High-end chips, data center power, and dedicated facilities all have long lead times for construction; an investment can secure more stable production capacity and enable clients to participate earlier in suppliers’ expansion plans.
The third relationship is more complex, with both parties purchasing each other’s products or services. The model may be optimized for a specific chip, the cloud platform integrates the model into its enterprise products, and the model company, in turn, procures training and inference resources from the cloud platform. Capital ties link these contracts together, encouraging both parties to invest more in customized development, specialized equipment, and long-term operations and maintenance.
In economics, this type of investment is commonly described as "specific investment." It refers to assets built for a particular partner, whose value declines significantly when transferred to another client. Examples include a data center cluster designed for a specific model or software stacks rewritten for a particular chip. Equity relationships help stabilize cooperation expectations and reduce the incentive for either party to renegotiate terms after the facilities are built.
Three: Public collaboration has already demonstrated this network.
Figure 2 | Wikimedia server room, where computational demand ultimately translates into substantial deployment of racks and equipment
Microsoft and OpenAI are among the most prominent examples. In 2023, Microsoft announced a multi-year, multi-billion-dollar investment, designating Azure as OpenAI’s exclusive cloud service provider. The funding supports model development and computational expansion, while the cloud services contract channels a significant portion of those expenditures back to Azure. Additionally, Microsoft distributes model capabilities through its own products, creating a single partnership that encompasses investment, procurement, infrastructure, and product distribution channels.
Amazon and Anthropic have adopted a similar partnership. Amazon announced it will invest up to $4 billion, while Anthropic has designated AWS as its primary cloud provider and uses Trainium and Inferentia chips to train and deploy models. For Amazon, the investment brings model collaboration and increased cloud usage; for Anthropic, funding, chips, and customer access are integrated into a single long-term partnership framework.
Large-scale collaborations also involve embedded technology migration costs. Once a model is trained and optimized for a specific type of chip, inference tools, data pipelines, and monitoring systems are all built around that environment. Switching cloud platforms impacts engineering teams and product stability. Investment relationships extend the collaboration timeline, encouraging both parties to jointly shoulder the initial adaptation costs.
The compute rental company CoreWeave demonstrates the connection between chip suppliers and new cloud infrastructure. NVIDIA provides CoreWeave with its GPU ecosystem and also participates in its financing and expansion arrangements. CoreWeave continues to build data centers, while NVIDIA gains a stable, large-scale deployment environment. Their collaboration directly impacts equipment procurement, capacity utilization, and the next round of capital expenditures.
Each of these transactions has a unique contractual structure and is at a different stage of development. The commonality is clear: the most scarce resources in the AI industry are concentrated in the hands of a few companies; spot purchases alone cannot cover multi-year construction and product cycles, making capital relationships an extension of supply contracts.
Four: 46.4% of the investment amount is also linked to business relationships.
Figure 4｜Key data on AI company investment relationships from 2021 to 2025
BIS statistics show that 16.1% of investment transactions between AI companies, by count, also involve commercial supply relationships; by disclosed transaction value, this proportion rises to 46.4%. Large transactions exhibit a more pronounced overlap with supply chain collaborations. Among circular relationships, 64% involve investors who also supply products or services to the invested companies.
The upstream segment is the most concentrated. 73% of circular investment relationships originate from computing power or infrastructure companies. Chips, cloud platforms, and data centers control expensive resources with long lead times for development; model companies must secure supply in advance, while upstream firms seek assurance of long-term demand. Both sides share part of the risk, leading naturally to collaboration in financing.
BIS also cited comparative data from traditional industries. Among more than 10,000 U.S. customer-supplier relationships, the proportion of direct equity holdings in trading partners was approximately 3.3%; in the AI computing, cloud, and related infrastructure submarkets, this proportion reached 15.2%. The capital requirements, concentration of supply, and degree of customization in the AI industry collectively increase the frequency of capital linkages.
The methodology is also worth noting. The BIS calculates based on the total amount of public transactions, and the data cannot be broken down to show each investor’s actual contribution within a single funding round. The 46.4% figure reflects a strong overlap between large funding rounds and business relationships, but it does not mean that an equivalent proportion of funds was designated for purchasing the investors’ products. Business collaborations have become a key contextual factor in AI funding.
Five: Order growth may include demand driven by capital.
The most direct financial impact of a circular relationship is that the source of demand becomes harder to distinguish. After investing in a customer, the supplier sees the customer use the new funds to purchase the supplier’s products, resulting in immediate revenue for the supplier, while simultaneously holding an equity or debt stake in the customer on its balance sheet. Revenue, expected investment returns, and customer valuation are all interconnected through this same relationship.
As long as end users continue to pay, this structure enables the industry to expand more rapidly. Model companies secure computing power in advance, data centers gain long-term orders, and chip manufacturers can plan production accordingly. Problems arise when final demand falls short of expectations: customers reduce purchases, suppliers first lose orders, then face equity impairments, loan losses, or guarantee obligations.
In the late 1990s, the telecommunications equipment industry offered a similar experience. Equipment vendors such as Lucent and Nortel provided financing to network operators, who then purchased their equipment. During the peak of network construction, sales and loan assets grew simultaneously; as end-user revenue slowed, operators' ability to repay debt and make purchases declined together, leaving equipment vendors under dual pressure from falling sales and financial losses.
The AI industry has a broader range of connections. Some infrastructure is financed through private credit, special purpose vehicles, and long-term asset residual value guarantees, with risks potentially dispersed among public companies, private enterprises, credit funds, and project companies. It is difficult to fully reconstruct actual capital contributions, future commitments, and guarantee obligations by examining only the investment amounts or procurement contracts disclosed by a single company.
Contract pricing also influences external assessments. Long-term compute agreements may include upfront payments, minimum usage requirements, price discounts, and expansion options, while equity investments could alter both parties' expectations regarding future orders. When analyzing the quality of an AI company’s customers, it is essential to present cash revenue, contract liabilities, remaining performance obligations, and purchases from affiliated partners on the same table to understand the term of orders and true payment capacity.
Six: Regulators and investors need to see the complete contract chain.
Cyclical investment serves real industrial demands. Regulatory focus will center on whether information is fully disclosed. When a company discloses an equity investment, investors must also be aware of whether, during the same period, there are minimum purchase volumes, cloud service credits, reserved computing power, asset residual value guarantees, or exclusivity arrangements. Revenue recognition, customer concentration, contingent liabilities, and investment impairments may all be affected by these terms.
Competition issues will also enter the discussion. Companies that control chips, cloud platforms, or model access points can accelerate product collaboration by investing in key customers and suppliers, thereby influencing other companies’ access to computing power, model integration, or customer acquisition. Different regulatory authorities—competition agencies, securities regulators, and banking supervisors—in various jurisdictions see different parts of the same network; cross-departmental information sharing will become increasingly important.
For the market, a more practical approach is to examine financing alongside orders. Questions such as how many new customers were brought in, how much revenue came from portfolio companies, how long procurement commitments last, and who is financing the related assets are more indicative of the quality of industry momentum than tracking financing valuations alone.
This BIS briefing breaks down the AI investment boom into a set of verifiable relationships. Compute providers, model companies, and application platforms are exchanging capital for demand, resources, and partnership stability. Over the coming years, growth in the AI industry will continue to be driven by this network, and markets will increasingly ask: For every dollar of revenue, how much comes from end customers versus capital invested upstream within the supply chain?
This article is intended solely for legal, policy, and industry research purposes, aiming to provide an objective analysis of digital finance, stablecoins, digital assets, and related regulatory developments. It does not constitute any form of investment advice, legal opinion, tax advice, or other professional guidance, nor does it constitute any recommendation, promotion, or solicitation of financial products, digital assets, or business projects. The regulatory rules, market data, and institutional information referenced in this article are primarily sourced from publicly available materials and may be subject to change due to evolving laws, regulations, market conditions, or project developments. Readers are advised to independently assess the information in light of the latest public disclosures and to comply with applicable laws and regulations in their respective jurisdictions. The author and the publishing platform shall not be held liable for any investment, trading, or other commercial decisions made based on the content of this article.