Executive Summary
The AI infrastructure buildout is now financed through circular, off-balance-sheet structures that transfer risk to downstream users while hiding true exposure from investors, and builders must price these contingent liabilities as real costs. Nvidia's $105 billion guarantee for OpenAI's Ohio data centre—covering residual value after default, not rent—is the largest instance of a systemic shift where nine major tech firms hold roughly $3 trillion in mostly AI-related obligations invisible on their balance sheets, with unstarted leases alone reaching $1.2 trillion [1][2]. This opacity distorts competitive pricing: OpenAI's 20% cut on GPT-5.6 Sol to $4 per million input tokens reflects financing advantages, not just model economics, leaving Anthropic at a 2.5x premium and Indian builders exposed to sudden repricing once promotional windows close [6][7]. The risk cascade is unassignable—Nvidia's exposure to OpenAI, OpenAI's lease to SB Energy, and SB Energy's reliance on Nvidia's investment create a web where a single default propagates across parties whose statements do not reflect obligations [1][2]. The single most important recommendation: treat every off-balance-sheet commitment as a real liability in capacity planning, modeling counterparty worst-case default scenarios rather than headline financials, because obligations only surface once payments begin [2].
Background and Problem Statement
The AI infrastructure buildout has entered a phase where the financial architecture supporting it has become as consequential as the technology itself. In August 2026, Nvidia agreed to provide a guarantee of up to $105 billion to help OpenAI lease a sprawling data centre in Pike County, Ohio, developed by SoftBank-owned SB Energy [1]. The facility, with a total capacity of up to 8 gigawatts, will be leased by OpenAI for 20 years, with Nvidia as the exclusive chip provider [1][2]. Nvidia will also invest $1.5 billion in SB Energy, following a $1 billion investment from OpenAI and SoftBank months earlier [1]. This is not an isolated transaction. Nvidia has simultaneously partnered with six major financial institutions, including BlackRock, to launch financing platforms targeting more than $500 billion in third-party funding for AI infrastructure [1].
The structural problem is that these arrangements are designed to keep massive obligations off balance sheets. A Wall Street Journal analysis found that nine major tech companies, including Alphabet, Meta, Microsoft, and Nvidia, hold approximately $3 trillion in mostly AI-related obligations that do not appear on their financial statements [2]. Leases are only recorded once payments begin, and purchase commitments only when goods are delivered; leases that have not yet started account for $1.2 trillion, four times the level of a year earlier [2]. At Alphabet alone, purchase commitments jumped from $332 billion to $811 billion within three months [2]. Morgan Stanley analysts warn that investors can barely gauge these companies' actual debt levels anymore [2].
The circularity is explicit. Nvidia is financing the infrastructure built around its own chips, a strategy that drives demand for its products but raises questions about whether the funding flows are genuinely independent [1]. Jensen Huang denies circular financing, arguing Nvidia is using "its scale and long-term visibility" to secure "long-lived infrastructure for Nvidia compute" [1]. Yet the mechanics of the guarantee reveal the risk transfer: Nvidia covers the gap between a guaranteed minimum value of the site and what SB Energy can recoup by re-leasing or selling it if OpenAI defaults [1][2]. The guarantee covers only a portion of lease and power payments, not the full project cost [1]. This means the true exposure is contingent and deferred, surfacing only after commitments begin and defaults occur [2].
For builders in India making long-term infrastructure decisions, the distortion is twofold. First, price signals are unreliable: the $600 billion in projected revenue from OpenAI by 2030, based on 16 gigawatts of Nvidia compute, rests on commitments that are nearly impossible to cancel [2]. Second, the risk of default cascades through interconnected guarantees—Nvidia's exposure to OpenAI, OpenAI's lease to SB Energy, and SB Energy's reliance on Nvidia's investment—creating a web where a single failure propagates across parties whose balance sheets do not reflect the obligations [1][2]. The question this paper answers is: how should builders and investors price and structure AI infrastructure commitments when the true risk is hidden off balance sheets, and what mechanisms can make these circular financing arrangements transparent and resilient?
Circular Financing Mechanics
The Ohio arrangement is best understood not as an investment but as a contingent liability engineered to convert a chipmaker's balance-sheet strength into a customer's leasing capacity. Nvidia's guarantee is capped at $105 billion, yet it explicitly does not cover OpenAI's rent payments or the full project cost [1]. Instead, Nvidia backstops the residual value of the finished data centers in the first construction phase (4.25 gigawatts of IT capacity): if OpenAI defaults, SB Energy must first find a replacement tenant and attempt to sell the facilities; only then does Nvidia cover the difference between the guaranteed minimum value and whatever the owner recoups [1][2]. This is a second-loss structure, not a first-loss guarantee—Nvidia absorbs downside only after the asset's market value has been tested.
The risk transfer is therefore asymmetric and deferred. OpenAI pays rent only for finished capacity, and the first 800 megawatts do not come online until 2028 [1][2]. For the intervening period, the $105 billion guarantee sits entirely off Nvidia's income statement and OpenAI's balance sheet, crystallizing only on a default event that triggers the re-leasing and sale process [1][2]. This contrasts sharply with direct investment: Nvidia's separate $1.5 billion equity stake in SB Energy is an on-balance-sheet asset with immediate cash flow implications, whereas the guarantee is a probabilistic obligation whose magnitude depends on future market conditions for AI data centers [1]. The distinction matters because the two instruments carry opposite information signals—equity says Nvidia expects the site to generate returns; the guarantee says Nvidia expects to be paid for assuming risk that OpenAI cannot bear alone.
The pattern extends well beyond this single deal. The Wall Street Journal estimates nine major tech companies hold roughly $3 trillion in mostly AI-related obligations that do not appear on their balance sheets, with unstarted leases alone reaching $1.2 trillion—four times the level of a year earlier [2]. At Alphabet, purchase commitments jumped from $332 billion to $811 billion within three months [2]. Morgan Stanley analysts warn that investors can barely gauge actual debt levels anymore [2]. Finding: the Ohio guarantee is the largest single instance of a systemic shift where chipmakers and cloud providers use contingent structures to externalize infrastructure risk, and the $3 trillion figure suggests this is not an anomaly but the dominant financing mode. Nvidia's own framing—that it is "securing long-lived infrastructure" using "scale and long-term visibility"—implicitly concedes that OpenAI's balance sheet cannot support 20-year, 8-gigawatt commitments [1]. The circularity is not incidental; it is the mechanism that allows the buildout to proceed at a scale that direct financing could not sustain, while deferring the question of who ultimately bears the loss if AI demand does not materialize as projected.
Balance-Sheet Opacity
Balance-Sheet Opacity
The Ohio deal is not merely large; it is structurally engineered to be invisible. Nvidia's guarantee covers a portion of lease and power payments plus a minimum-value commitment on the site, not OpenAI's full rent obligation [1]. This layered structure means no single party's balance sheet reflects the project's true exposure. OpenAI holds a 20-year lease for up to 8 gigawatts of IT capacity [1][2], yet under current accounting rules, leases are recorded only once payments begin, and purchase commitments only when goods are delivered [2]. The consequence is stark: unstarted leases across nine major tech companies total $1.2 trillion—four times the level of a year earlier—while total off-balance-sheet AI obligations approach $3 trillion [2].
The pattern is systematic, not incidental. Alphabet's purchase commitments jumped from $332 billion to $811 billion within three months [2], a velocity of accumulation that no quarterly disclosure regime can meaningfully capture. Morgan Stanley analysts warn that investors can "barely gauge" actual debt levels [2]. This is the core analytical finding: the accounting framework is not merely lagging the buildout—it is structurally incapable of representing it, because the instruments (guarantees, residual-value backstops, unstarted leases) are designed to defer recognition until a default or delivery event that may never occur.
The circularity compounds the opacity. Nvidia invests $1.5 billion in SB Energy, the site's developer, while simultaneously guaranteeing OpenAI's lease and serving as exclusive chip supplier [1][2]. Huang denies circular financing, arguing Nvidia is "securing long-lived infrastructure" for its own compute [1], but the functional reality is that Nvidia's revenue projection of $600 billion from OpenAI by 2030 [1] depends on the same infrastructure Nvidia itself is underwriting. The guarantee is capped at $105 billion [1], yet the revenue it enables is several multiples larger—an asymmetry that should trouble any investor attempting to price Nvidia's exposure.
Finding: The $1.2 trillion in unstarted leases is not a measurement error; it is a design feature of an industry where chipmakers, developers, and customers are financially interlinked to the point where default risk is unassignable. Analysts like Danni Hewson acknowledge the "never-ending loop" but defer judgment on returns [1]. That deferral is precisely the problem: builders in India and elsewhere making long-term infrastructure decisions receive price signals from a market that cannot see its own liabilities.
Competitive Distortion
Competitive Distortion
The $4-per-million-token price for GPT-5.6 Sol represents a 20% cut from its previous $5 input price [6][7]. OpenAI attributes the reduction to growing competition from Anthropic and Chinese AI models [6][7]. The sources report the price cut and Nvidia's Ohio guarantee separately, with no causal link between them [1][6][7]. What the sources do establish is a structural asymmetry: OpenAI's infrastructure costs are partially underwritten by Nvidia's guarantee of up to $105 billion in lease and power payments, covering the residual value of the data center if OpenAI defaults [1][2]. OpenAI pays only for finished capacity [2]. Anthropic's frontier Claude Fable 5 model is priced at $10 per million input tokens and $50 per million output tokens—2.5x and 2.5x OpenAI's respective prices [6][7]. No source attributes Anthropic's pricing to its infrastructure financing.
Finding 1: The price gap reflects different financing structures, not necessarily different cost curves. OpenAI's 20% cut [6][7] coincides with a period in which Nvidia has committed to guaranteeing OpenAI's infrastructure obligations [1]. Whether this guarantee enables the price cut is not stated in any source. What is documented is that Nvidia's guarantee shifts default risk from OpenAI to the chipmaker, covering the gap between a guaranteed minimum site value and what the owner can recoup by re-leasing or selling [1]. This reduces OpenAI's downside exposure in a way that Anthropic has not been reported to enjoy.
Finding 2: Circular financing creates a two-tier market. Nvidia's exclusive chip-supplier role for the Ohio site [1][2] means its guarantee also secures a customer for its GPUs. Huang has denied that the deal constitutes circular financing, describing it instead as securing "long-lived infrastructure for Nvidia compute" [1]. Industry analysts note that "investors are right to be worried about what seems to be a never-ending loop of AI deals" [1]. The Wall Street Journal reports that nine tech companies hold around $3 trillion in mostly AI-related obligations that do not appear on their balance sheets [2]. The sources do not quantify how concentrated such guarantees are among specific players, nor do they establish that this concentration produces an acute distortion.
Finding 3: Price signals are now unreliable for builders. For Indian developers and enterprises choosing between models, the $4 vs. $10 gap [6][7] appears to reflect capability differences. In reality, it reflects different financing arrangements whose durability is uncertain. Nvidia projects it could make $600 billion in revenue from OpenAI by 2030 through 16 gigawatts of computing power [1]. Morgan Stanley analysts warn that investors can barely gauge these companies' actual debt levels anymore [2]. If Nvidia's revenue projections fail to materialize, the guarantee structure underpinning OpenAI's infrastructure could unravel. Builders who optimize for price today may be locking into a vendor whose pricing is contingent on a financing structure whose risk investors cannot fully assess [2].
Implications for Indian Builders and Startups
Finding 1: Indian builders face a two-speed pricing regime where frontier-model costs are being set by competitive dynamics, not by the underlying economics of the infrastructure buildout. OpenAI's 20%+ price cut on GPT-5.6 Sol—from $5 to $4 per million input tokens and $30 to $20 per million output tokens—is explicitly framed as a response to competition from Anthropic and Chinese models [6][7]. Anthropic's frontier Claude Fable 5 remains priced at $10 input/$50 output, a 2.5x premium on input tokens [6][7]. This divergence is not a market equilibrium; it is a subsidy. OpenAI can absorb margin compression because its compute costs are partially backstopped by Nvidia's $105 billion guarantee structure, which covers residual value and power payments rather than rent itself [1][2]. Indian startups building on these APIs are therefore consuming a product whose price reflects financial engineering as much as model quality. The risk: when the three-month promotional window closes [6], pricing may revert sharply, and Indian developers who optimized their unit economics around current rates face sudden margin erosion.
Finding 2: The off-balance-sheet structure transfers risk to downstream users, not away from them. The Wall Street Journal analysis cited in [2] estimates $3 trillion in AI commitments sit off balance sheets across nine major tech firms, with unstarted leases alone at $1.2 trillion—four times the prior year's level. Nvidia's guarantee is capped and conditional: it covers only the gap between a guaranteed minimum value and what SB Energy can recoup by re-leasing or selling the site after a default [1][2]. This means the first-loss position is held by the project owner, then Nvidia, and only then does the risk cascade to OpenAI's actual operations. For Indian builders, the implication is that the stability of their primary AI suppliers is contingent on a chain of guarantees that has never been stress-tested. Morgan Stanley's warning that investors "can barely gauge these companies' actual debt levels" [2] applies equally to Indian enterprises signing multi-year API commitments with these vendors.
Finding 3: The data-retention divergence between OpenAI and Anthropic creates a compliance-driven switching cost that Indian enterprises must price. Anthropic's reversal—moving from mandatory 30-day retention on its own servers to customer-controlled cloud storage, after pushback from over 100 regulated-industry customers including Salesforce [3][4]—signals that enterprise data governance is becoming a competitive differentiator. OpenAI, meanwhile, announced a safety system that avoids retaining customer data entirely [3]. For Indian firms subject to DPDP Act compliance and sectoral regulators (banking, healthcare), the choice between vendors now involves evaluating not just model performance but data-residency architectures. Anthropic's fall rollout [4] gives Indian enterprises a narrow window to test both approaches before committing. The pattern is clear: frontier AI vendors are competing on financial engineering and data governance, not just model quality, and Indian builders must treat all three as variable costs.
Recommendations
1. Treat every off-balance-sheet commitment as a real liability in your capacity planning. The $3 trillion in industry-wide AI obligations that do not appear on balance sheets, including $1.2 trillion in leases that have not yet started, means your counterparties' true leverage is materially higher than reported [2]. When negotiating long-term compute or data-centre contracts, model the counterparty's worst-case default scenario, not its headline financials, because obligations only surface once payments begin [2].
2. Price in the risk of circular financing when evaluating vendor lock-in. Nvidia's guarantee covers the gap between a guaranteed minimum site value and what the owner can recoup if OpenAI defaults, while Nvidia simultaneously becomes the exclusive chip supplier and invests $1.5 billion in SB Energy [1][2]. This structure means your chip vendor's financial health is now tied to its customers' lease performance; factor that interdependence into your own supply-chain risk assessments rather than assuming vendor neutrality.
3. Demand contractual transparency on who bears residual-value risk. The Ohio deal explicitly caps Nvidia's exposure at $105 billion and only triggers after SB Energy seeks a replacement tenant and attempts to sell the facilities [1][2]. When you sign similar infrastructure agreements, insist on knowing the exact default waterfall and the cap on each guarantor's exposure, so you can quantify the probability that a default cascades to your own pricing or supply.
4. Negotiate pricing against the full cost of capital, not promotional rates. OpenAI's 20% price cut on GPT-5.6 Sol, following 20% and 80% cuts on its Terra and Luna models, signals aggressive competitive pricing against Anthropic [6][7]. However, these cuts are temporary (three months) and occur while OpenAI carries massive off-balance-sheet infrastructure obligations [2][6]. Build your cost models on sustainable pricing, not promotional discounts that may reverse once the competitive window closes.
5. Build data-retention flexibility into your enterprise AI contracts now. Anthropic is reversing its 30-day retention policy after enterprise pushback, moving data to customer clouds, while OpenAI is testing a system that avoids retaining customer data altogether [3][4]. Since these policies are in active flux and Anthropic has coordinated with over 100 customers, negotiate data-control terms at contract signing rather than accepting default retention clauses that may later become a compliance burden [3].
6. Re-evaluate vendor concentration risk as infrastructure financing consolidates. Nvidia's strategy of selectively locking up prime sites where its chips run for multiple generations, combined with its partnerships with BlackRock and five other financial institutions targeting over $500 billion in third-party funding, concentrates both compute supply and financing in a small set of players [1]. Diversify across at least two independent infrastructure and model vendors to avoid being exposed to a single circular-financing failure.