AI's Circular Financing: Commodity Model or Bubble Risk

The AI industry is increasingly dependent on circular financing deals to fund data center infrastructure and GPU access, a structure mirroring how commodity industries have financed large-scale projects for decades. In a typical AI circular deal, a startup or smaller data center operator receives GPU allocation and capital from Nvidia, with Nvidia also guaranteeing to purchase any compute capacity the startup cannot use itself. Nvidia may take an equity stake as well. This structure works because GPUs are effectively fungible commodities—two data centers using similar Nvidia chips are interchangeable for most purposes—and because data center development is prohibitively expensive, often costing well over $10 billion（約1.6兆円）. Nvidia has a $1 trillion（約160兆円） order backlog and can resell or redeploy hardware if a borrower fails, making it a credible guarantor.

The model has proven successful in traditional commodities. Japanese development banks and commodity traders financed infrastructure for oil and minerals development starting in the 1960s, committing to future purchases and taking equity stakes. Jamaica used similar structures in the 1980s. Recently, the US Department of War committed to buying all neodymium-praseodymium (rare-earth magnets) from MP Materials for at least $110 per kilogram under a 10-year agreement, an off-take deal that has since been mirrored between MP Materials and General Motors. These structures reduce lender risk by guaranteeing revenue for the producer.

In AI, the approach has scaled rapidly. CoreWeave secured a $6.3 billion（約1兆円） deal with Nvidia structured as a circular commitment, and smaller operators like Firmus obtained $505 million（約810億円） in similar arrangements. Hyperscalers like SpaceX and Meta lease their own data centers to third parties at significant premiums—SpaceX's Colossus 1 and 2 data centers lease for over $2 billion（約3200億円） per month. However, some circular deals now hide obligations off balance sheets. Google backs bonds issued by TeraWulf to finance Fluidstack data centers, which are leased by Anthropic; Google's guarantee ensures the bonds are repaid even if Anthropic cannot pay its lease. Similarly, Meta's $27.3 billion（約4.4兆円） Hyperion data center bond sale and its more recent $12.3 billion（約2兆円） bond sale (marketed by BlackRock) do not appear on Meta's balance sheet as formal liabilities, even though Meta is ultimately responsible if the underlying projects fail to generate revenue.

These arrangements remain serviceable because hyperscalers generate billions of dollars in annual profit and can easily cover defaults. However, two vulnerabilities exist. First, the entire model assumes GPUs and data center capacity remain fungible and resellable. If AI chip standards shift, if technologies emerge enabling efficient local or small-cluster model training, or if demand for massive centralized compute declines, the resale value of current Nvidia infrastructure could collapse and off-take guarantees would provide little protection. Second, Nvidia's exposure is substantial; it has committed up to $750 billion（約120兆円） in guarantees. A broad shift in AI architecture or a slowdown in data center deployment could force Nvidia to absorb losses far larger than the individual deals. Off-balance-sheet guarantees from hyperscalers similarly mask true leverage; if the underlying projects face revenue shortfalls or if hyperscaler margins tighten, these hidden liabilities could become material and destabilize the broader ecosystem.

Circular financing in AI is not inherently bad—it is a well-established strategy in commodity industries. Oil development, minerals mining, and infrastructure projects routinely use guaranteed off-take agreements and equity partnerships to fund expensive upfront infrastructure when individual parties cannot absorb the cost. The analogy holds because Nvidia GPUs are fungible (standardized across all firms), data centers are expensive ($10 billion（約1.6兆円）+), and demand for compute is high with large order backlogs. Nvidia's $1 trillion（約160兆円） order backlog and the willingness of well-capitalized labs to pay premiums for immediate capacity confirm that the underlying product is resellable—a prerequisite for the commodity model to work.

The structure also explains the web of overlapping relationships: OpenAI cancels with Oracle and Meta steps in; SpaceX leases servers to Google; Google invests in Anthropic. Each party takes on risk in exchange for revenue, equity upside, or guaranteed supply. This is how capital gets deployed to build the infrastructure the industry needs.

The danger emerges in two forms. First, technological disruption could render the commodity fungible no longer. If local models, small-cluster inference, or new chip architectures reduce demand for massive centralized data centers, the resale value of current Nvidia infrastructure evaporates and off-take agreements become worthless. Nvidia's own overextension—committing up to $750 billion（約120兆円） in guarantees—amplifies this risk. Second, opaque off-balance-sheet structures mask true leverage. Google's backstop of TeraWulf bonds and Meta's $27.3 billion（約4.4兆円） and $12.3 billion（約2兆円） bond sales hide obligations from standard financial reporting. These are serviceable today because hyperscalers earn billions in annual profit, but the hidden liabilities could surface if earnings pressure increases or if the circular deals themselves fail.

No comments yet. Be the first to share your thoughts!