The $3 Trillion AI Infrastructure Binge Hiding in the Footnotes
Big Tech’s AI infrastructure spending refers to the massive long-term commitments hyperscale technology companies are making to chips, energy, leases, and data centers to power artificial intelligence services, much of which appears in filing footnotes as future obligations instead of on balance sheets as current liabilities. Nine technology giants have together disclosed about $3 trillion in such commitments largely tied to AI infrastructure, a number that already dwarfs their recent capital expenditure and yet still understates the true economic risk investors are taking on. The obligations include $1.2 trillion in leases that have not started and $1.9 trillion in purchase commitments for chips, energy and data-center infrastructure. Alphabet alone reported $811 billion in purchase and contractual commitments as of June 30, while Meta disclosed $347 billion in leases that have yet to commence.
Accounting Games: How Data Center Debt Stays Out of Sight
The real problem is not just the size of AI infrastructure spending, but how it is being presented. These trillions in obligations are disclosed in filing footnotes yet are not recognized as liabilities on the face of corporate balance sheets. Investors who skim headline figures can walk away believing Big Tech’s data center debt is manageable, when in reality the companies have “gorged themselves” on off-balance-sheet liabilities, including backstops, uncommenced leases and purchase commitments. Michael Burry alleges that hyperscalers are using overly long depreciation schedules for rapidly aging AI chips and servers, assigning useful lives of five to six years to hardware whose economic cycle is closer to two or three years. He argues this practice could suppress depreciation expenses by $176 billion between 2026 and 2028 and inflate earnings figures for companies such as Oracle and Meta.
Model Convergence Could Turn Today’s Assets Into Tomorrow’s Stranded Costs
All of this might be survivable if AI demand were guaranteed and returns obvious. They are not. If AI demand disappoints, these companies could be left paying for expensive infrastructure that fails to generate sufficient returns. At the same time, AI strategies are shifting: Chamath Palihapitiya expects enterprises to move toward model-agnostic AI platforms within the next 36 months as the cost gap between open models narrows and businesses “realize the importance of data/IP leakage”. As model costs converge, he argues companies will increasingly use independent third-party harnesses or control planes that work with multiple AI models. That shift raises a hard question: why lock in multiyear chip and data center commitments built around today’s architectures if tomorrow’s AI stack becomes interchangeable, cheaper, and controlled by a different layer entirely?
The Coming Compression: When Spending, Earnings and Financing Collide
Burry’s warning is blunt: hyperscalers are marching toward a three-way “compression” where demand weakens, earnings fall as expenses catch up, and financing tightens as capital becomes more expensive. He has focused on circular financing among hyperscalers, AI labs and chipmakers, arguing that capital moving within this loop can artificially reinforce revenue and demand. Meanwhile, he asks a simple question that boards prefer to dodge: “When does the spending for AI data center buildout actually end?”. The answer matters because these companies have already amassed hundreds of billions in off-balance-sheet exposure, from construction assets that do not begin depreciating until placed in service to forward purchase commitments such as Nvidia’s $182 billion. When investor skepticism collides with opaque accounting and still-unclear AI investment returns, the earnings party can end abruptly.
What Investors Should Demand Before the AI Bill Comes Due
The AI boom is being sold as inevitable progress, but the financial story is far messier. Enterprises may well pivot to model-agnostic platforms and third-party control planes that reduce lock-in to any single AI provider. That future could be good for companies positioned as neutral orchestrators, yet brutal for hyperscalers stuck servicing data center debt that no longer matches where the value in AI resides. Investors should stop treating these commitments as distant abstractions and start viewing them as real, looming claims on future cash flow. Off-balance-sheet AI infrastructure spending, aggressive depreciation assumptions, and opaque data center debt structures are not minor footnote quirks; they are core to Big Tech financial risk and to whether AI investment returns will ever justify the scale of the bet. Until those questions are answered in plain numbers, the market is mispricing risk hiding in plain sight.
