There is a pattern in how the AI boom is being financed that Bloomberg documented in a detailed investigation in March 2026 and that has since become one of the most debated structural questions about the AI investment cycle. It works like this: Microsoft invested more than $13 billion in OpenAI over several years. OpenAI committed to spending $250 billion on Azure cloud services. Oracle is constructing $300 billion in Stargate data center infrastructure for OpenAI under long-term contracts. Nvidia invested $30 billion in OpenAI’s most recent $122 billion funding round. OpenAI continues to be one of Nvidia’s largest chip customers. Nvidia has also taken stakes in CoreWeave and other neoclouds, which are in turn Nvidia’s infrastructure customers.
The money moves from investor to startup and back to the investor’s products and services through a loop that is self-reinforcing while AI demand is growing and potentially self-amplifying in ways that obscure where genuine end-user demand ends and engineered demand begins.
Bloomberg was careful to distinguish circular deals from fraudulent round-trip transactions. The circular structure is not illegal or inherently inappropriate. It can be a winner for all involved if things go well: Company A buys a stake in Company B, giving Company B more money to invest and expand so that, in the end, it needs more of Company A’s products and services. The critique is not that the deals are improper. It is that they can inflate valuations based on committed spending that originates from within the same small circle of companies rather than from organic end-user demand, and that at the scale these deals have reached in 2026, the valuation inflation risk has become large enough to matter for the AI investment thesis.
The Scale That Has Changed the Nature of the Risk
The Bloomberg investigation was published in January 2026, when the circular deal count was significant but not yet extraordinary. In March 2026, OpenAI closed a $122 billion funding round at an $852 billion valuation, with Amazon investing $50 billion, Nvidia investing $30 billion, and SoftBank investing $30 billion. Amazon’s $50 billion commitment is structured partly as compute credits, with OpenAI simultaneously committing to spend $100 billion on AWS over eight years. Nvidia’s $30 billion investment goes into a company that is one of Nvidia’s largest GPU customers. SoftBank committed $3 billion annually to OpenAI technology deployment across its own portfolio of companies. Critics note that when you strip away the compute commitments and conditional tranches, the actual cash injection is substantially smaller than the headline suggests, and the $852 billion valuation assumes revenue growth that has never been sustained by any technology company at this scale.
The Amazon-Anthropic structure, announced April 20, mirrors the Microsoft-OpenAI pattern at even larger scale. Amazon invested $5 billion in Anthropic with options for up to $20 billion more. Anthropic simultaneously committed to spending more than $100 billion on AWS over the next decade, securing up to 5 gigawatts of compute capacity. Amazon is investing in a company that will spend 20 times the investment back on Amazon’s infrastructure. The circular structure at that ratio is not just a financial relationship between two companies. It is Amazon using an investment vehicle to guarantee $100 billion in AWS revenue over a decade, structured in a way that counts as a funding round for Anthropic’s valuation rather than as a customer prepayment from Amazon’s revenue perspective.
What the Circular Structure Reveals About AI Economics
The circular deal dynamic reveals something specific and important about the economics of frontier AI development that the deal structures themselves obscure. Frontier AI model development requires compute at a scale that no company can finance from current operating cash flows. OpenAI reported $8.5 billion in annualised losses despite $5 billion in revenue. Anthropic is spending at rates that require continuous external financing. The only way these companies can access the compute they need to remain competitive is through financing relationships with the cloud providers and chipmakers that supply that compute. Those suppliers have a strong incentive to provide financing because the AI labs are their largest and most strategically important customers. As a result, the market has evolved into a circular financial structure in which a small group of players controls both demand and supply.
The risk is not that this architecture will collapse. The genuine end-user demand for AI products and services is real and growing. The risk is that the valuation levels it has produced for AI companies, and the capital allocation decisions those valuations drive in infrastructure investment, data center development, and chip procurement, are partially based on committed spending that would not exist if the circular financing structure unwound. Janus Henderson, the asset manager, described the wave of AI deals as more like a virtuous circle that helps line up suppliers, builders and customers to meet exploding demand for computing power. That framing is fair when demand is growing.
The Stress Test the Market Has Not Faced
The analytical question is what happens to the virtuous circle if demand growth moderates, if AI product monetisation timelines extend further than current models project, or if one of the major participants in the loop encounters financial difficulty that forces it to reduce its commitments. For the AI infrastructure market, that is not a theoretical question. It is the most important stress test that $785 billion in annual hyperscaler capex has not yet faced.
The infrastructure market that is most exposed to the circular deal dynamic is not the one that built the first $200 billion in capacity. It is the one building the next $600 billion on the assumption that the demand the circular financing has front-loaded is as real as the committed spending makes it appear. Tonight Nvidia earnings will tell us whether that demand is converting to revenue at the pace the financing structures assumed. That is a much more important question than whether Nvidia beats consensus by $800 million.
