AI infrastructure is booming — but who’s funding who?
By Leeja van Bezouwen, Credit Analyst and Victor Verberk, CEO/CIO
Key Takeaways
- AI infrastructure deals use complex, compliant accounting structures, but hidden risks remain.
- Circular financing and long-term commitments can expose suppliers to credit and refinancing risks.
- Many ecosystem players rely on external funding, making them vulnerable if financial conditions tighten.
- Risks are correlated across equity, leases, and receivables, amplifying potential impact.
- Historical parallels show that compliant accounting can mask vulnerabilities until stressed.
Introduction
Over the past year, announced AI infrastructure commitments have exceeded USD 1 trillion. The headline numbers are staggering. But from a credit perspective, the more important question is how these transactions are structured and where the credit risk actually sits.
We examined the accounting behind the major AI infrastructure deals: lease classifications, special purpose vehicles, remaining performance obligations, equity stakes, and depreciation policies. Our conclusion is that while accounting treatments appear broadly compliant with current standards, the structures do embed vulnerabilities that are not immediately visible in reported figures. These include circular financing arrangements, duration mismatches between obligations and revenues, and correlated exposures across multiple balance sheet lines.
This note sets out what we found and why it matters for credit investors.
1. What the accounting shows
A natural starting point for our review was whether the scale and complexity of recent AI infrastructure transactions has led to premature or aggressive revenue recognition. Veterans in the industry might remember similar concerns during the dot com period in the early 2000s.This concern has been raised frequently in market commentary, particularly given the involvement of long-dated contracts, leasing structures and counterparties that are not yet profitable..
Our review did not identify evidence of premature revenue recognition. The accounting treatments we examined appear to comply with current standards, including the revenue recognition requirements under ASC 6061. Lease arrangements are generally structured as operating leases, with revenue recognised over the lease term rather than upfront. Remaining performance obligations are disclosed as future contractual commitments and do not result in immediate revenue recognition. Equity investments in customers are recorded as assets, separate from product sales. So far so good.
However, the structures do raise questions.
- First, there is increasing circularity in how capital flows through the ecosystem. In several cases, Graphics Processing Unit (GPU) suppliers have made equity investments in customers, who then use that capital to purchase or lease GPUs from the same supplier. Each step in this chain is accounted for correctly in isolation. Taken together, however, the effect is that suppliers are, to some extent, helping to finance their own revenue growth. The income statement reflects product sales, but not the extent to which the supplier contributed to the customer’s funding capacity.
- Second, a substantial amount of debt is held in special purpose vehicles, often collateralised by GPUs. These vehicles can sit outside the consolidated balance sheet while still supporting core activities. The collateral depreciates rapidly, with some companies assuming useful lives of five to six years while many market participants view two to three years as more realistic. If asset values decline faster than debt amortises, covenant headroom erodes and refinancing risk increases. While these structures are accounted for appropriately at inception, credit risk can re-emerge at the sponsor level if conditions deteriorate.
- Third, several large cloud infrastructure contracts are structured around remaining performance obligations from customers that are not yet profitable. RPOs do not accelerate revenue recognition, but they can create a sense of visibility that rests on important assumptions. These obligations often extend over many years and are large relative to customers’ current cash generation. At the same time, suppliers may already be committing substantial capital, with cash receipts occurring much later. This creates a timing mismatch: capital deployed upfront, revenue and cash arriving years later. When delivery timelines extend due to construction or supply-chain constraints, this gap widens. With a significant share of future activity depending on a small number of counterparties with uncertain cash flows, the quality of that backlog deserves scrutiny.
It is, however, worth noting that many of these commitments are structured around milestones or delivery schedules rather than being unconditional. Capital is often deployed in stages, and purchase commitments may only become binding as specific deployment targets are met. In some cases, customers can reportedly cancel GPU orders until relatively close to delivery. This provides some protection against over-commitment, but primarily for the customer. Milestone structures primarily affect the timing of investment. They do not remove the underlying dependence on customers being able to fund these commitments over the long term.
“Milestone structures affect timing, not the underlying dependence on customers’ ability to fund commitments.”
Finally, some supply agreements are supported by warrants or equity grants to customers. While these arrangements are accounted for correctly, they affect the underlying economics of the transaction. When equity is granted as part of a commercial relationship, its value effectively reduces the price the supplier receives for the goods or services provided. As a result, headline deal values can give an inflated impression of the revenue ultimately earned, even though the accounting treatment itself follows established rules.
2. Where resilience is tested
“Compliance does not necessarily imply resilience.”
The accounting structures described above are broadly compliant with current standards. Compliance, however, does not necessarily imply resilience. Many of these arrangements rest on the assumption that AI applications will ultimately generate sufficient revenue to support the infrastructure being built around them. If that monetisation path proves slower or weaker than expected, the accounting consequences could be significant.
In such a scenario, pressures would likely emerge across multiple parts of the balance sheet at the same time. Equity stakes in customers could face impairment as valuations adjust or funding conditions tighten. Leased hardware would remain on suppliers’ balance sheets, still depreciating, but with weaker residual value assumptions and reduced re-leasing potential. Receivables and lease payments would require renewed assessment of collectability, increasing expected credit losses. For suppliers with large remaining performance obligations tied to a small number of counterparties, capital may already have been deployed to build capacity that does not convert into revenue, leading to underutilised assets and potential write-downs. Where debt is held in SPVs secured by GPUs, falling asset values could trigger covenant pressure or require sponsor support, effectively pulling risk back onto the balance sheet.
“Pressures would likely emerge across multiple parts of the balance sheet at the same time.”
These risks are amplified by the current state of profitability across the ecosystem. At present, profitability within the AI infrastructure chain is concentrated primarily at the GPU supplier level. Downstream, the picture is different. Many neoclouds have yet to demonstrate sustained profitability and continue to operate with limited or negative cash generation. AI labs are generating revenue but remain loss-making, with profitability targets several years away. The startups building applications on top of these models are typically not yet profitable either. Across each layer, purchases are funded largely through equity raises or debt rather than operating cash flow. As a result, the ecosystem depends, directly or indirectly, on continued access to external capital. If funding conditions tighten, the ability of customers to meet their commitments would come under pressure across multiple layers simultaneously.
Duration mismatch adds a further dimension. Some infrastructure providers have committed to lease terms extending 15 years or more, while customer contracts typically run five years or less and may not be renewed. If AI demand slows or key customers reduce their commitments, the provider faces fixed obligations without matching revenue streams. Unlike equipment that can be redeployed or written down, lease commitments are contractual and often carry significant termination penalties.
Stepping back, many of the characteristics described above are typically associated with equity or venture capital investments: long development cycles, uncertain monetisation, and a wide dispersion of outcomes. What is notable in the current AI infrastructure build-out is that a growing share of this risk is being financed through debt-like instruments. For credit investors, this matters because debt absorbs downside without participating in upside, making the alignment between risk profile and funding structure a key consideration.
“A slowdown in AI monetisation would flow simultaneously through equity, leases, receivables and infrastructure assets.”
The broader point is that these exposures are correlated. A slowdown in AI monetisation would not affect a single revenue line or asset class, but would flow simultaneously through equity investments, lease portfolios, receivables and infrastructure assets. For credit investors, this concentration of risk around a shared set of assumptions, spread across multiple accounting lines, warrants close attention.
3. Why this rhymes with history
The dynamics described above are not unique to AI. Similar structural patterns have appeared in earlier periods of rapid infrastructure build-out driven by transformative technologies. The closest parallel is the telecom and data infrastructure expansion of the late 1990s and early 2000s.
That period included well-documented cases of accounting misconduct and inappropriate revenue recognition at individual companies. Those failures were specific and severe.
At the same time, the broader sector also exhibited vulnerabilities that were structural rather than fraudulent. Large amounts of capital were deployed ahead of proven demand, supported by long-dated contracts, vendor financing, special purpose vehicles and optimistic assumptions about utilisation and pricing. In many cases, accounting treatment for these structures was compliant with prevailing standards at the time. Vulnerabilities only became apparent when demand growth fell short of expectations or funding conditions tightened. Stress then surfaced not primarily through revenue restatements, but through impairments, refinancing challenges, balance-sheet consolidation and deteriorating credit metrics.
“Vulnerabilities only became apparent when demand growth fell short or funding conditions tightened.”
The comparison with AI is not intended to suggest that the same outcomes will occur, nor that similar misconduct is present today. The technology, use cases and market structure are different, and some participants, particularly the largest hyperscalers, are better capitalised. The relevance of history lies in showing how rapid infrastructure expansion can create correlated exposure to the same underlying demand assumptions across multiple balance sheets, and how downside risk often emerges through accounting and financing channels rather than through obvious breaches of rules.
4. Implications for credit investors
Whether AI investments will ultimately yield the expected returns remains uncertain. As credit investors, our focus is on how this infrastructure build-out is being financed and where credit risk is accumulating along the way. Much of the AI infrastructure build-out relies on long-dated commitments, front-loaded capital deployment and counterparties whose ability to pay depends on continued access to external funding rather than internally generated cash flows.
Private credit continues to play a growing role in funding these structures. The associated risks, uncertain monetisation paths, rapid technological obsolescence and potentially limited recovery values therefore require careful assessment when expressed through debt instruments. Debt absorbs downside risk without participating in upside, making the alignment between risk profile and capital structure critical.
For us, this reinforces the importance of looking beyond headline yields and compliant accounting outcomes. The more fundamental questions are where cash ultimately comes from, how resilient asset values are under stress, and how exposures interact when multiple counterparties depend on the same underlying demand materialising.
These are the questions we will continue to ask.
- ASC 606 is the US GAAP standard governing how and when companies recognise revenue from contracts with customers. ↩︎
Important Information
Osmosis Investment Management NL B.V. (Osmosis NL) is licensed as an Alternative Investment Fund Manager (AIFM) under the Alternative Investment Fund Managers Directive (AIFMD) and the Dutch Financial Supervision Act (Wet op het financieel toezicht, Wft) and authorized to provide discretionary portfolio management services. Osmosis NL is subject to supervision by the Dutch Authority for the Financial Markets (AFM).
This document and any marketing communication are intended solely for Professional Clients as defined in the Wft. It is not directed at, nor intended for distribution to, any person in any jurisdiction where such distribution would be unlawful.
The information provided is for general information purposes only and does not constitute investment advice, a recommendation, research or an offer or solicitation to buy or sell any financial instrument. It is not tailored to individual circumstances or investment objectives.
The views expressed are as of the date of publication of this document and may change without notice. Although this information is obtained from sources believed to be reliable, no representation or warranty is made as to its accuracy or completeness. Osmosis NL accepts no liability for any direct or indirect loss arising from use of this information.
Past performance is not a reliable indicator of future results. No representation or warranty is made that any account or investment will achieve results similar to those shown. Actual results may differ substantially due to factors such as market conditions, timing and pricing of trades, portfolio composition, fees, and client circumstances. Investments can fall as well as rise in value and may result in the loss of capital. Forecasts, projections, or targets are for illustrative purposes only and are not guaranteed in any way.
Any investment examples included herein are for illustrative purposes only and do not constitute a recommendation to buy or sell any specific security. There is no assurance that such investments will remain in the strategy or have ever been held. Case studies have been selected on a non-performance basis as indicative of the investment approach and process.
Benchmark information is provided for comparison purposes only. Indices are unmanaged, not available for direct investment, and do not reflect the deduction of fees or expenses, which would reduce returns. Past benchmark performance is not a reliable indicator of future results, and the referenced benchmarks may not be appropriate for all investors.
If reference is made to an investment fund, please refer to the relevant fund’s prospectus or offering documents with more details on investment objectives, costs, and risks before making any final investment decisions.
Scenarios and performance presented are estimates based on past data and current market conditions and are not exact indicators of future results. Actual outcomes will vary depending on market performance and the duration of investment.
Clients are encouraged to consult their own legal, tax, accounting, and other professional advisers before making investment decisions and to promptly inform Osmosis NL of any changes to their investment objectives or financial situation.
For Australian Investors: Osmosis NL is a Corporate Authorised Representative (CAR 001316961) of Eminence Global Asset Management Pty Ltd (EGAM) (AFSL holder 305573). Where Osmosis NL provides financial services in Australia, it does so as an authorised representative on behalf of EGAM. The information and materials contained in this document have been prepared for accredited wholesale clients only, as defined by Corporations Act 2001 (Cth) and in accepting the content of this document, you warrant that you are such an investor.