Is AI Demand Becoming Circular?

The short answer is yes, with an important limit.

More AI infrastructure is being supported by transactions in which the investor, supplier, customer, landlord and guarantor roles overlap. The scale is no longer anecdotal. In its latest quarterly filing, NVIDIA reported $99 billion of equity investments, $25 billion of equity-investment commitments and $36 billion of cloud-service commitments. It also disclosed guarantees capped at $105 billion for a site intended to be leased to OpenAI.

Those arrangements make some reported orders, backlogs and forward revenue less independent than they first appear. Capital supplied or supported by an ecosystem participant can return to that participant as revenue.

But that is not the same as saying the revenue is fictitious, the infrastructure will go unused or end-customer demand is absent. AI suppliers are still delivering real chips and capacity. External customers are still signing contracts. NVIDIA’s latest Data Center revenue grew 117% year over year to $89.0 billion. The analytical problem is narrower: how much of today’s demand evidence reflects independent willingness to pay, and how much reflects capital and risk being recycled within the ecosystem?

What circular financing means

In plain language, circular financing occurs when a company helps fund a customer or infrastructure partner, and that recipient then uses part of the capital to buy the funder’s products or services.

The circle need not be a literal transfer of the same dollars. It can involve an equity investment, a loan, a purchase guarantee, a capacity backstop, a long-term lease, a warrant or several contracts across different entities. The economic link matters more than the legal form.

A simple example looks like this:

The financing can be economically useful. It can bridge a timing gap between enormous upfront construction costs and years of future cash flow. It can also let specialist operators build capacity that a customer needs but does not want to own.

The problem arises when the financing is presented as independent demand confirmation. A supplier-assisted order proves that a transaction was contracted. It does not, by itself, prove that unaffiliated end customers will ultimately generate enough cash to support the entire chain.

Vendor financing and cross-investments are now material

The cleanest current illustration is NVIDIA’s expanding role beyond selling accelerators.

NVIDIA disclosed that selected AI clouds buy its infrastructure while NVIDIA commits to use cloud services from them. Those commitments, typically six years long, totalled $36 billion at 26 July 2026 and decline as third parties or NVIDIA consume capacity. The company may also share in revenue generated by third-party customers.

That structure can solve a real constraint. An AI cloud may have customer demand but lack the balance sheet to finance a large cluster. NVIDIA’s commitment makes the project easier to fund and can reserve capacity for its own research.

It also couples several signals that an observer might otherwise count separately: NVIDIA records infrastructure sales; the cloud provider records capacity and backlog; NVIDIA carries a service commitment; and future third-party revenue may flow back through a revenue share. Counting each as independent proof of demand would exaggerate the evidence.

Other agreements make the overlap even clearer:

• NVIDIA and OpenAI announced a letter of intent under which OpenAI would deploy at least 10 gigawatts of NVIDIA systems and NVIDIA intended to invest up to $100 billion progressively as capacity was deployed. A stated intention is not a completed investment.
• Amazon agreed to invest $50 billion in OpenAI while OpenAI and AWS expanded a compute agreement by $100 billion over eight years. The partnership also gives AWS distribution rights for an OpenAI enterprise product and provides OpenAI with roughly two gigawatts of Trainium capacity.
• NVIDIA invested $2 billion in CoreWeave in January 2026. CoreWeave buys NVIDIA systems, while a separate agreement requires NVIDIA to purchase residual unsold CoreWeave capacity, subject to its terms, through April 2032.

None of these transactions is automatically improper or uneconomic. Strategic customers and suppliers have financed one another for decades. The relevant issue is disclosure and interpretation: a purchase supported by the seller or by a related investor carries less independent information about final demand than an arm’s-length cash purchase.

Private credit and SPVs change where the risk sits

AI infrastructure is increasingly financed outside the ordinary unsecured corporate balance sheet. Private-credit funds, asset managers and banks can lend to a special-purpose vehicle, or SPV, that owns the servers, land or data centre and is repaid from a specific customer contract.

The structure can be sensible. Long-lived infrastructure can be matched with long-dated capital, and project risk can be separated from the operator’s other assets. It can also create bankruptcy remoteness, giving lenders a defined pool of collateral and contracts.

CoreWeave’s latest 10-Q provides a concrete example. The company reported $35.6 billion of total debt at 30 June 2026 and said certain facilities are held in bankruptcy-remote, special-purpose subsidiaries that own financed infrastructure and related customer contracts. During the first half, it generated $3.7 billion of operating cash, used $14.9 billion in investing activities and raised $14.0 billion from financing activities. Its capital investment is funded through debt and equity issuance, delayed-draw facilities, OEM financing and balance-sheet cash.

The numbers show rapid financing capacity, not current failure. CoreWeave also reported approximately $104 billion of revenue backlog at 30 June and more than $25 billion of additional commitments early in the third quarter. The same filing warned that the company has substantial indebtedness and depends on continued access to capital.

Meta’s Hyperion data-centre joint venture shows a different version. Funds managed by Blue Owl own 80% of the vehicle and Meta retains 20% of an approximately $27 billion development. Blue Owl contributed about $7 billion in cash; Meta received a roughly $3 billion distribution; and part of Blue Owl’s capital was financed by debt placed with PIMCO and other investors. Meta agreed to lease all completed facilities and provided a capped residual-value guarantee for the first 16 years.

That is not hidden debt in the simplistic sense. Meta disclosed the ownership, leases and guarantee. But it does mean that an asset appearing outside Meta’s consolidated property portfolio can still depend heavily on Meta’s future payments and credit support.

What a long-term offtake agreement does

An offtake agreement is a promise to buy output or capacity over time. In energy, it may cover electricity from a new plant. In AI infrastructure, it can cover data-centre space, power, cloud capacity or compute.

Offtake makes construction financeable because lenders can underwrite contracted revenue instead of relying only on a forecast. A take-or-pay contract can require the customer to pay whether or not it uses every unit. A residual-value guarantee can protect the asset owner if a lease ends and the specialised facility is worth less than expected.

That improves funding certainty, but it also moves demand forward. A ten- or twenty-year commitment records future revenue visibility today while leaving the customer responsible for finding productive uses tomorrow.

NVIDIA’s latest filing illustrates the scale. It disclosed guarantees with SB Energy supporting a 4.25-gigawatt Ohio campus intended to host NVIDIA compute under 20-year leases to OpenAI. NVIDIA’s obligation is capped at $105 billion, becomes effective as data centres enter service and declines as OpenAI makes lease payments. The guarantees are conditional and do not cover every site cost or all tenant obligations.

CoreWeave’s agreement with NVIDIA is another form of support: if customer demand does not fully use specified capacity, NVIDIA must purchase the residual, subject to delivery, availability and termination provisions, through 13 April 2032. The backstop reduces CoreWeave’s occupancy risk. It also means that booked capacity under this arrangement is not purely a test of unaffiliated end-customer demand.

How much does this weaken the demand signal?

Enough to change how the evidence should be weighted. Not enough to discard it.

The information content of demand depends on who ultimately bears the cost:

1. STRONGESTIndependent paid usage

   Unaffiliated enterprises or consumers pay, renew and expand without linked financing. Utilisation and cash margin confirm value.

2. STRONGArm’s-length committed capacity

   A creditworthy customer signs a long-term contract without receiving investment from the supplier. Demand is real, but timing and utilisation still matter.

3. MIXEDInvestment plus purchase commitment

   The customer commits to buy while the supplier or platform invests in the customer. The contract matters, but the signals are not independent.

4. WEAKESTVendor-supported capacity or guarantee

   The seller, sponsor or related party absorbs residual demand, financing or asset-value risk. The order says less about external willingness to pay.

No public dataset lets BoomRisk assign the sector to those four buckets comprehensively. Contract terms are private, the same project can contain several layers, and large commitments may be announced years before cash is deployed. A precise statement such as “a third of AI demand is circular” would therefore imply more certainty than the evidence supports.

The defensible conclusion is qualitative: circular structures have become large enough that chip sales, backlog and announced capacity should not be treated as fully independent demand observations. Analysts should trace the source of capital, the ultimate payer, the contract’s cancellation rights and the loss-bearing party before using a deal as confirmation.

The counterargument: financing can reveal demand rather than manufacture it

The bearish interpretation can also go too far.

Financing often exists because demand is real but the assets are unusually expensive and long-lived. A cloud provider can have credible customers and still lack sufficient capital to build ahead of them. A vendor backstop can reduce a temporary financing constraint, accelerate deployment and lower unit costs. Long-term contracts can be rational when power, land and leading accelerators are scarce.

There is also contrary operating evidence. NVIDIA’s $89.0 billion of quarterly Data Center revenue is completed revenue, not merely an announced project. CoreWeave’s backlog has broadened, and its annual report said no single customer represented more than 35% of backlog at the end of 2025, down from 85% at the start of that year. In April 2026, Jane Street committed approximately $6 billion to CoreWeave cloud services and made a separate $1 billion equity investment. The equity link makes that signal mixed, but Jane Street is an external operating customer with a disclosed use case.

BoomRisk’s own Semiconductor Demand proxy remains at 1/5, its lowest risk score. The broad US investment-grade spread has widened only 6 basis points over roughly 90 days, leaving the Credit Stress proxy at 3/5 rather than signalling a credit event.

Those facts matter. They argue against treating financing complexity as proof of artificial demand. They do not remove the need to test whether external revenue, utilisation and cash returns eventually catch up with the commitments.

BoomRisk’s assessment: Building, not Intensifying

BoomRisk currently classifies its Financing Structure Watch as Building, with medium confidence.

Building means linked financing, private credit and off-balance-sheet obligations are becoming more material and more visible. It does not mean losses are occurring, that the structures are failing or that a credit crisis has begun.

The assessment is not yet Intensifying because financing remains available, the broad credit proxy is not showing severe stress, large projects have credible counterparties, and several headline amounts are intentions, memorandums or conditional commitments rather than fully deployed capital. Strong semiconductor revenue and expanding external customer relationships provide meaningful contrary evidence.

What would move the assessment to Stable?

The evidence would support Stable if the share of linked financing stopped rising; new projects were increasingly funded by independent customer cash flow or conventional arm’s-length capital; utilisation, renewals and external revenue became clearer; sponsor guarantees and residual-capacity commitments stopped expanding; and issuer-specific credit spreads remained contained.

The live Watch currently uses Easing rather than Stable as a displayed direction. An actual decline in obligations, guarantees or dependence on related-party funding would justify Easing. Stable would mean the current build-up had levelled off without yet reversing.

What would move it to Intensifying?

Intensifying would require more than another large headline. BoomRisk would look for several developments together: vendor or platform capital funding a growing share of purchases; guarantees and take-or-pay obligations hardening faster than external revenue; weaker counterparties relying on serial refinancing; materially wider issuer credit spreads; debt becoming harder to place; declining utilisation or contract renegotiations; and losses moving from an SPV back to its sponsor or supplier.

The distinction is important. More financing is Building. Financing stress becoming mutually reinforcing is Intensifying.

Why this watch remains unscored

Circular financing does not change today’s 3.38/5 composite.

It remains unscored for four reasons:

• No reliable denominator. Public disclosures do not show what share of total AI investment or supplier revenue is supported by related financing.
• Terms are not comparable. An equity stake, a cancellable cloud commitment, a take-or-pay lease and a capped guarantee create different risks.
• Double-counting is a real danger. Credit conditions are already scored through the Credit Stress proxy, while capex and revenue conversion appear elsewhere in the framework.
• No tested thresholds exist. BoomRisk will not invent a neutral number or create 1–5 bands before there is sufficient history to calibrate and backtest them without hindsight.

The watch is still useful. It changes how other evidence is interpreted and identifies where risk may surface first. But context is not a score, and an unmeasurable concern should not be smuggled into the composite through editorial judgement.

The conclusion

AI investment is increasingly supported by circular and closely linked financing structures. That weakens the independence of some purchase, backlog and capacity signals because the capital provider can also be the seller, buyer, guarantor or future beneficiary.

It does not follow that AI demand is artificial.

The decisive test is whether capital raised inside the ecosystem ultimately converts into durable cash flow from outside it. Today’s evidence says the financing loop is Building. It does not yet say the loop is breaking.

BoomRisk is a financial-risk monitoring framework, not an investment recommendation or market forecast. Financing announcements and contractual commitments are not equivalent to deployed capital, recognised revenue or realised investment returns.