AI Infrastructure Financing Loops

The phrase “circular financing” should describe a traceable flow of money, guarantees, or reciprocal commitments. By mid-2026, company filings show several such links. They establish coupled financing and demand, but not yet that revenue is artificial, contracts will fail, or the system is a credit bubble.

Two disclosed financing networks

Amazon, Anthropic, and OpenAI

Amazon's 2026 Q2 10-Q links capital and commercial capacity in unusually explicit terms:

• Amazon and Anthropic expanded their AWS relationship by more than $100 billion over ten years, including obligations tied to AWS chip performance.
• Amazon made $10 billion of Anthropic preferred-stock investments in Q2 and described a facility of up to $20 billion. Availability under that facility increases as Amazon reaches compute-capacity delivery milestones; $15 billion remained after the Q2 investments.
• Amazon had invested $28.7 billion in OpenAI preferred stock by June 30 and funded the remaining $21.3 billion commitment after quarter-end. It also disclosed an AWS commercial arrangement involving AWS chips and a joint model-services collaboration.

These facts satisfy the first two arrows in a possible loop: Amazon provides capital, and the model companies commit to Amazon infrastructure. The filing does not by itself show the share of purchases funded by Amazon, the margin on those services, or how much capacity is consumed by independent end users.

NVIDIA, AI clouds, and project guarantees

NVIDIA's fiscal 2027 Q2 10-Q discloses a broader set of links:

• $279 billion of supply and capacity commitments, $25 billion of equity-investment commitments, and $36 billion of AI-cloud service commitments as of July 26, 2026;
• agreements in which AI clouds buy NVIDIA systems while NVIDIA commits to purchase cloud capacity, with those commitments declining as third parties or NVIDIA use the capacity;
• maximum gross guarantee exposure of $108.5 billion, including a conditional $105 billion cap related to a 4.25-gigawatt OpenAI campus with 20-year leases;
• memoranda with capital providers intended to mobilize more than $500 billion over time, with possible limited residual-value support from NVIDIA.

The gross guarantee cap is not a current loss or balance-sheet liability of the same amount. The exposure starts as facilities enter service, falls as lease payments are made, and remains subject to contract conditions. It is nevertheless direct evidence that a supplier can support the land, power, financing, and capacity that enable customers to buy and deploy its products.

The transaction to look for

A potentially reinforcing loop might involve:

1. a supplier or investor provides equity, debt, a guarantee, or purchase support to a customer;
2. the customer uses that capacity to buy the supplier's product;
3. those purchases support the supplier's reported demand or valuation;
4. the financing remains exposed to the customer's ability to refinance or monetize the purchased assets.

Each arrow needs a contract, filing, offering memorandum, or explicit company disclosure. The Amazon and NVIDIA filings now provide that evidence for several arrows. The remaining question is whether independent customers ultimately pay enough to cover the assets and obligations.

Ordinary structures versus fragile ones

Feature	Usually less fragile	Worth closer review
customer base	diversified external customers	one or two related anchor customers
contract	cancellable or matched to end demand	long take-or-pay term with uncertain downstream use
debt	long maturity, fixed rate, strong coverage	short maturity, floating rate, refinancing dependence
collateral	reusable asset with liquid secondary market	specialized, rapidly depreciating equipment
guarantee	limited and clearly disclosed	broad supplier, sponsor, or parent guarantee
revenue	cash sale to an independent buyer	sale funded by the seller or a closely related investor

“Take or pay” is not automatically abusive; it can finance real infrastructure by making demand bankable. The question is who ultimately bears the loss if the forecast demand does not appear.

Why hardware duration matters

Data centers and power infrastructure can have long useful lives, while accelerators may lose economic competitiveness much faster. A financing package can therefore combine long-lived buildings, power contracts, and short-lived compute equipment. Analysis should separate those assets rather than assigning one depreciation story to the entire project.

Useful checks include:

• accounting depreciation versus expected economic life;
• resale value and alternative users for the equipment;
• upgrade obligations in customer contracts;
• debt maturity relative to the revenue contract;
• who owns the residual asset after termination.

Working classification

Label	Current assessment
coupled financing and demand	supported by primary disclosures
circular revenue	plausible in specific arrangements, but not established without transaction-level funding and revenue-recognition data
broad credit bubble	not established; the reviewed filings do not show a system-wide default or refinancing break
loss-transfer risk	material where suppliers guarantee sites, buy unused capacity, or depend on counterparties to reimburse them

Evidence record for one loop

party_a: supplier or investor
party_b: customer or project company
capital_flow: equity | loan | guarantee | purchase support
purchase_flow: product or service
source_documents: []
recourse: unknown
maturity: unknown
related_revenue_recognition: unknown
independent_end_demand: unknown

If the source documents are empty, the loop remains a hypothesis. When documents do exist, record the maximum exposure, activation conditions, independent end demand, revenue recognition, and who holds the residual asset. A large headline commitment without those terms can overstate both safety and danger.

My reading · Opened notes are remembered in this browser.