Is the AI Boom Built on Funny Money?

What Investors Should Understand About the Financial Architecture Behind the AI Boom

Introduction: The Trillion-Dollar Question

The artificial intelligence sector has attracted unprecedented capital, with investments reaching astronomical figures. OpenAI alone has committed $1.4 trillion in commitments over the next eight years, a figure larger than the GDP of most countries.

Consider a simple analogy: I give your company $1,000, and tomorrow, you give my company $1,000 back. We both book the transaction and claim we made $1,000 in revenue. On paper, we're growing. In reality, no new value was created. This simple exchange is a template for the kind of transactions that can inflate an entire sector.

This analysis examines several accounting considerations that sophisticated investors should understand when evaluating AI-related financial statements. These are not allegations of wrongdoing, auditors at major firms like Deloitte (OpenAI's auditor) and the Big Four continue to sign off on these statements. Rather, this is an exploration of the structural complexity that makes AI financials challenging to evaluate, and why greater transparency may benefit all stakeholders.

Fair-minded readers should note: many of these accounting treatments are standard practice for high-growth industries. The early internet and telecom booms featured similar structures. The question is not whether these practices are legitimate, but whether they create opacity that obscures economic reality, and whether that opacity carries risks investors should understand.

Angela Liu and I discuss the nuances of this article in great detail on the Gaapsavvy podcast, which you can find here, along with a more consolidated version for easier listening here.

1. The Collectibility Conundrum

Are AI Startups Writing Checks Their Business Models Can't Cash?

Public tech companies report massive backlogs and remaining performance obligations (RPOs) tied to AI contracts.

The accounting framework: Under ASC 606, companies must disclose remaining performance obligations, often called 'backlog', representing contracted future revenue. However, this disclosure comes with important nuances that investors should understand:

• Termination clauses matter significantly. If a customer can terminate a contract without substantive cost or penalty, only the noncancelable portion is accounted for under ASC 606. The enforceability period, not the stated contract term, determines what can be reported.

• Non-GAAP backlog disclosures vary. Companies may voluntarily disclose backlog figures outside financial statements that include amounts subject to termination. These should be clearly distinguished from GAAP-compliant remaining performance obligations.

• Collectibility assessments require judgment. ASC 606 requires that revenue recognition occur only when collection is probable. For multi-year commitments from cash-burning startups, this assessment involves significant management judgment.

The financial context: OpenAI's financial trajectory, as reported by multiple outlets including The Wall Street Journal, Deutsche Bank, and HSBC, projects cumulative cash burn of $115-143 billion through 2029 (the range coming from different analysts), with profitability not expected until 2029 or 2030. Deutsche Bank analysts noted this represents losses 'on a scale no startup has operated with in history.'

In November 2025, OpenAI CEO Sam Altman publicly stated the company is 'looking at commitments of about $1.4 trillion over the next 8 years for data center infrastructure. This staggering figure, larger than the GDP of most countries, represents promises OpenAI is making to cloud providers, chip manufacturers, and infrastructure partners.

But here's the critical catch: OpenAI's financial reality doesn't match these commitments

• Projected cumulative cash burn of $115–143 billion through 2029

• Deutsche Bank analysts wrote: 'No startup in history has operated with losses on anything approaching this scale. We are firmly in uncharted territory.'

• HSBC projects a cumulative free cash flow deficit of $207 billion by 2030

• Profitability not expected until 2029 or 2030 at the earliest

• Annual operating losses projected to reach $74 billion in 2028 alone

How can public tech companies be certain that multi-year, multi-billion dollar promises from a company burning cash at this rate are truly 'collectible' over the long term? Unlike Google or Microsoft, which are funded by massive, sustainable revenue engines, OpenAI's commitments rest on continuous fundraising and projected future revenues that may never materialize.

Bull Case: Auditors have reviewed these contracts and their collectibility assessments. OpenAI has secured substantial funding and has access to capital markets. The $1.4 trillion figure represents commitments, not immediate obligations. Additionally, many of these contracts may include milestone-based payments that reduce concentration risk.

2. The Consolidation Shell Game

In October 2025, OpenAI completed its restructuring into a Public Benefit Corporation, and Microsoft's stake was formally disclosed for the first time: approximately 27% of OpenAI Group PBC, valued at roughly $135 billion.

Here's what makes this remarkable: at 27% ownership, Microsoft is above the traditional 20% threshold that typically triggers equity method accounting, yet they are still not consolidating OpenAI's financials.

Microsoft's October 2025 10-Q filing confirms: 'The investment is accounted for under the equity method of accounting, with our share of OpenAI's income or loss recognized in other income (expense), net.'

This is not full consolidation, and that distinction matters enormously. Under equity method accounting, Microsoft only recognizes its proportional share of OpenAI's net income or loss as a single line item. Under full consolidation, Microsoft would have to:

• Merge OpenAI's entire balance sheet with its own, all assets, all liabilities, all debt

• Eliminate intercompany revenues (the billions flowing back and forth)

• Show OpenAI's full operating losses in Microsoft's operating income

• Consolidate OpenAI's $1.4 trillion in infrastructure commitments and obligations

Instead, Microsoft keeps OpenAI's projected $115+ billion in cumulative cash burn, its massive debt obligations, and its full scale of operating losses completely off Microsoft's consolidated financial statements. The equity method creates a convenient partition, a single line item that masks the true financial exposure of this deeply intertwined relationship.

If the two companies were fully consolidated, the $250 billion Azure commitment and reciprocal revenue-sharing arrangements would have to be eliminated as internal transactions. Instead, Microsoft can book Azure revenue from OpenAI as real external revenue.

The fact that Microsoft, with a 27% stake, $13+ billion invested, board observation rights, and deeply intertwined operations, is still using equity method rather than consolidation should raise questions about what the combined picture would actually look like.

The Bull Case: The equity method is appropriate under GAAP when an investor has 'significant influence' but not 'control.' Microsoft does not control OpenAI's board or operations, the OpenAI Foundation maintains control of the for-profit entity. Auditors have reviewed this determination. The 20% threshold is a presumption, not a bright-line rule, and control, not ownership percentage, is the ultimate determinant.

Transparency Requests: Given the depth of operational integration (exclusive cloud provider, revenue sharing, IP rights), does equity method accounting provide investors with a complete picture of Microsoft's economic exposure to OpenAI?

3. Circular Revenue and Intercompany Flows

Understanding the Money Flows

The Microsoft-OpenAI relationship involves substantial bidirectional revenue flows. For example:

• OpenAI committed to purchase $250 billion in Azure cloud services

• Microsoft pays OpenAI approximately 20% of Azure OpenAI Service and Bing revenue

• OpenAI pays Microsoft approximately 20% of ChatGPT and API revenue

• Leaked documents suggest OpenAI spent $8.67 billion on Azure inference in the first nine months of 2025

The Accounting Question: ASC 606 requires that reciprocal arrangements be evaluated for commercial substance. In arrangements where parties exchange goods or services of similar fair value, there may be questions about whether the transaction price reflects the fair value of what is being exchanged. Microsoft's 10-Q references 'reciprocal revenue-sharing arrangements' but does not provide gross versus net figures.

The Bull Case: These are arm's-length transactions between legally separate entities. Both parties receive genuine value, OpenAI gets compute infrastructure; Microsoft gets AI model access and revenue share. The circular nature reflects legitimate business interdependence, not accounting manipulation. Both companies' auditors have reviewed these arrangements.

Transparency Requests: What is the net economic transfer between these entities after accounting for all reciprocal flows? Should public companies provide greater disclosure of gross versus net revenue from related parties with significant ownership stakes?

4. GPU Depreciation and Useful Life Assumptions

Michael Burry, the investor who predicted the 2008 financial crisis and was portrayed in 'The Big Short', has argued that hyperscalers are 'understating depreciation by extending useful life of assets,' estimating this could understate depreciation by approximately $176 billion between 2026 and 2028. This has sparked significant debate.

His concerns are backed up by the following:

• Microsoft extended server useful life from 4 years (2020) to 6 years (2022), saving $3.7 billion in operating income

• Meta extended server depreciation to 5.5 years, reducing depreciation expense by $2.3 billion in the first nine months of 2025

• Amazon shortened useful life for some servers to 5 years in 2025, citing 'the increased pace of technology development, particularly in AI and machine learning'

• Nvidia CEO Jensen Huang joked that when Blackwell ships, 'you couldn't give Hoppers away'

The Bear Case: Technical analyses suggest GPU useful life of 1-3 years due to thermal stress and rapid obsolescence. Princeton's Center for Information Technology Policy noted: 'The accounting subsidy creates a window of roughly three to six years where reported costs are artificially low. That window is precisely when the market structure for AI applications is being determined.'

The Bull Case: Nvidia responded to Burry's critique with a detailed memo disputing his claims, arguing that its GPUs maintain significant value years after release. CoreWeave CEO Michael Intrator noted that A100 chips from 2020 remain 'fully booked' and H100s were re-leased at 95% of original price. Under ASC 250, changes in useful life are treated as changes in accounting estimates, not corrections of errors, and require new information suggesting assets will remain in service longer.

Transparency Requests: Given the divergence between Amazon (shortening life) and Meta (extending life), what specific engineering data and utilization metrics support each company's assumptions? Should companies disclose sensitivity analyses showing earnings impact of alternative useful life assumptions?

5. Systemic Risk and Interdependence

The Financial Cascade if Just One AI Giant Falls

The AI infrastructure buildout has created deep interdependencies:

• OpenAI depends on Microsoft ($250B Azure), SoftBank ($40B funding), CoreWeave, Oracle, and continuous capital markets access

• Microsoft depends on OpenAI models for Copilot, Azure AI services, and competitive positioning

• CoreWeave carries $14 billion in debt with OpenAI and Microsoft as major customers

• S&P Global warned SoftBank's 'financial condition will likely deteriorate' due to OpenAI investment

If a major AI company fails to achieve profitability, cascading effects could include: impairment of multi-year infrastructure commitments, write-downs of backlog and remaining performance obligations, losses for lenders in the AI infrastructure stack, and evaporation of revenue growth that justified premium valuations.

The Bull Case: AI is demonstrating genuine productivity gains. ChatGPT has 700+ million weekly users. Enterprise adoption is accelerating. The 'winner-take-all' dynamic means survivors will capture enormous value. Microsoft's core business generates massive cash flow that can absorb OpenAI-related volatility.

Transparency Requests: What is the aggregate cross-company exposure in the AI infrastructure ecosystem? How would a stress scenario at one major player cascade through interconnected balance sheets?

6. Capitalized Software and AI Development Costs

Under ASC 350-40 (internal-use software) and ASC 985-20 (software for sale), companies must determine when development costs should be capitalized versus expensed. For AI companies, this creates significant judgment around model training costs.

The accounting framework:

• Preliminary project stage costs are expensed (planning, evaluation, research)

• Application development stage costs are generally capitalized

• Post-implementation costs (maintenance, training users) are expensed

The AI-specific challenge:

Deloitte's guidance on generative AI costs notes that 'because generative AI is essentially a form of software, we believe that general software development accounting considerations apply.' However, determining when a foundation model transitions from 'preliminary research' to 'application development' involves significant judgment, particularly given ongoing debates about technological feasibility of frontier AI.

Additionally, the distinction between 'training costs' (developing the model) and 'inference costs' (running the model) affects whether costs are capitalized as development or expensed as operations. OpenAI reportedly separates these, with training costs largely covered by Microsoft credits while inference is paid in cash.

Transparency Requests: What portion of AI company R&D spending is being capitalized versus expensed? How do companies determine when an AI model has achieved 'technological feasibility'? Are capitalization policies consistent across the industry?

7. Off-Balance Sheet Financing and SPV Structures

According to the Financial Times, Oracle, Meta, xAI, and CoreWeave have pushed more than $120 billion of AI data center financing off their balance sheets using Special Purpose Vehicles (SPVs). This financing structure allows companies to secure computing capacity while keeping large debts out of their core financial statements.

How it works:

• SPVs, separate legal entities, raise debt to build data centers or purchase GPUs

• Tech companies lease capacity from these SPVs rather than owning the assets

• Because the tech company doesn't control the SPV, the debt remains off its balance sheet

• Private capital from Pimco, BlackRock, Apollo, Blue Owl, and major banks provides the funding

Key examples:

• Meta's 'Beignet Investor' SPV raised $30 billion for the Hyperion facility, including ~$27 billion in loans

• CoreWeave created an SPV for its $11.9 billion OpenAI contract, later expanded to $22+ billion

• xAI is raising $20 billion including up to $12.5 billion in debt through an SPV structure

In Meta's Beignet deal, Meta owns 20% of the SPV and provided a 'residual value guarantee', meaning if the data center value falls below a set level at lease end and Meta doesn't renew, it must repay investors. UBS analysts noted that SPV financing 'actually increases the outstanding liabilities of technology companies, implying that their overall credit quality may be worse than what current models indicate.'

The Bull Case: These structures are legal and have been used in many industries. They allow capital-intensive buildouts without tying up corporate balance sheets. The assets (data centers, GPUs) serve as collateral. Risk is transferred to sophisticated institutional investors who understand the structures.

Transparency Requests: What is the aggregate economic exposure of hyperscalers to their off-balance sheet AI infrastructure? How should guarantees and residual value commitments factor into credit analysis?

Conclusion: The Case for Greater Disclosure

This analysis does not allege fraud or accounting violations. Major auditing firms, Deloitte, EY, PwC, KPMG, continue to sign off on these financial statements. The accounting treatments described are generally permissible under GAAP.

 The question is whether current disclosure practices provide investors with sufficient transparency to evaluate:

• The true economic substance of circular revenue flows between related parties

• The extent of 'soft backlog' versus firm commitments in reported RPOs

• The aggregate off-balance sheet exposure through SPV structures

• The sensitivity of earnings to alternative depreciation assumptions

• The systemic risk created by concentrated interdependencies

 The SEC is already examining these issues. Chief Accountant Kurt Hohl recently noted that auditor independence frameworks may no longer be 'fit for purpose' given AI alliances that have Big Four firms both auditing and selling products built on the same platforms.

The AI boom may well prove transformative and sustainable. But investors deserve to make that assessment with full visibility into the financial architecture supporting it. Greater transparency is not a bear case, it is a precondition for informed capital allocation.

 A Note on Sources

 This analysis relies on a combination of public filings, audited financial statements, and investigative reporting. Certain figures, particularly OpenAI's internal projections, derive from leaked documents reported by The Wall Street Journal, The Information, and other outlets. While these leaks have been widely cited and not disputed by the companies involved, readers should note they represent unverified reporting and may not reflect current projections.

 Public company filings (10-Q, 10-K, 8-K) cited herein are available through SEC EDGAR. Analyst estimates from Deutsche Bank, HSBC, and others represent professional opinions, not verified facts.

 Key Sources

• Microsoft 10-Q, October 30, 2025 (equity method disclosure, OpenAI investment)

• Microsoft 8-K and Blog, October 28, 2025 (27% stake, $250B Azure commitment)

• Sam Altman, X post, November 6, 2025 ($1.4T commitments, $20B ARR)

• Financial Times, December 2025 ($120B off-balance sheet AI financing)

• The Wall Street Journal, November 2025 (OpenAI financial projections)

• Deutsche Bank, HSBC Global Research (OpenAI cash burn analysis)

• Deloitte, 'Technology Spotlight: Accounting for Generative AI Software Products'

• ASC 606 (Revenue Recognition), ASC 350-40 (Internal-Use Software), ASC 250 (Accounting Changes)

• Princeton CITP, October 2025 (AI chip lifespan analysis)

• CNBC, November 2025 (GPU depreciation debate, Burry/Nvidia response)

• Seeking Alpha, November 2025 (OpenAI auditor confirmation - Deloitte)