top of page
Search

Assessing Risks in AI Infrastructure Finance

Updated: Nov 20

By Didier Vila, PhD Founder and MD of Alpha Matica.


In the development of artificial intelligence infrastructure, companies are allocating significant capital—projected to reach $3-8 trillion cumulatively by 2030 [1]. However, this expansion involves various financial considerations that may affect the sector's sustainability. Based on analyses from institutions such as Goldman Sachs, the International Energy Agency, and J.P. Morgan, this article examines key challenges in AI infrastructure financing, including circular funding structures and off-balance-sheet arrangements.


While these approaches have supported growth, they raise questions about long-term viability [2][3]. As outlined in McKinsey's reports on digital infrastructure, achieving AI advancement requires both technological progress and financial stability [4].

ree

Circularity in Funding and Deals


A notable aspect of AI's financial ecosystem involves circular funding and deals, where hyperscalers invest in AI startups that subsequently direct resources back to the investors' cloud services and hardware. Chipmakers illustrate this through equity investments and vendor financing, providing components at reduced costs with repayment through future transactions or equity [5][6].


According to Goldman Sachs estimates, such arrangements have supported over $400 billion in AI-related infrastructure spending this year [7]. While circularity can align interests and address compute limitations, it may lead to elevated valuations without matching external demand.


This resembles past vendor financing models from the dot-com period, where similar structures contributed to market adjustments when cycles disrupted [3]. If adoption of large language models slows (as McKinsey data indicates, with only one-third of companies reporting scaled AI programs organization-wide), these interconnections could result in asset revaluations and reduced market capitalization [4].


Amortisation and Depreciation Practices


Accounting for AI hardware, such as GPUs with useful lives often limited to three years or less due to rapid technological advancement, introduces complexities [8]. Amortization practices vary, with some firms extending asset lives to support earnings.


For example, Meta announced it was extending the useful life of its network gear to 5.5 years, a change expected to reduce its 2025 depreciation expense by approximately $2.9 billion [9]. Similarly, Amazon extended the useful life of its servers to six years, a move that saved the company nearly $1 billion in a single quarter [10].


This issue arises as hyperscalers' capital expenditures soar [8]. Shorter depreciation periods (1-3 years) increase expense recognition and impact margins. Longer periods (5-6 years), however, may attract regulatory review from entities like the SEC, as they risk overstating profits and asset values by not accounting for the rapid obsolescence of AI hardware [8]. McKinsey's frameworks indicate that varying practices can hinder accurate assessments of economic feasibility, making AI investments dependent on asset longevity assumptions [1].


Off-Balance-Sheet Funding and Debt Levels


To manage balance sheet impacts, AI entities are utilising off-balance-sheet mechanisms and increased debt. Special purpose entities enable acquisition of billions in hardware without immediate liability recording [5][6], while sector-wide debt has surged.


In 2025 alone, U.S. secured debt for data centers jumped 112% to $25.4 billion [11]. In a clear sign of this trend, tech giants like Meta, Oracle, and Alphabet issued a combined $75 billion in bonds and loans in just September and October 2025—more than double the sector's previous annual average [11]. Hyperscalers have accessed credit for data center growth, with Alphabet alone reaffirming $75 billion in capex for 2025 [12].


These methods maintain equity valuations and facilitate expansion but introduce broader risks. Off-balance-sheet approaches can obscure leverage, possibly underrepresenting default likelihoods. This reliance on debt is highlighted by J.P. Morgan projections, which estimate hyperscaler capital expenditures at $342 billion for 2025 (a 62% rise), noting that this spending could be volatile if demand weakens [13].


Additional Considerations: Energy Costs and Supply Chain Issues


AI data centers consumed about 4.4% of U.S. electricity in 2023 and are projected to reach 4.6-9.1% by 2030 [14][15]. This may require $6.7 trillion in global infrastructure while exposing utilities to grid bottlenecks and low reserve margins [1].


Supply chain reliance on concentrated chip manufacturers increases geopolitical risks, potentially raising costs and causing overcapacity in the buildout, alongside limited job creation and up to 12 million U.S. occupational shifts by 2030 [1][16].

Vulnerability

Mitigation Strategies

Circularity

Diversify funding sources; stress-test demand scenarios

Amortisation Conservatism

Standardise depreciation via industry benchmarks (e.g., 3-year)

Off-Balance-Sheet Debt

Enhance disclosure; cap leverage at 3x EBITDA

Energy/Operational Costs

Invest in renewables; optimise efficiency with edge computing

Forecasting/Supply Risks

Adopt agile planning; hedge via diversified suppliers


Implications for Stakeholders


In summary, the sector's future depends on turning investments into clear returns. If AI boosts productivity, these risks may be temporary. If not, the industry could see significant changes. Stakeholders should balance innovation with sound financial management.


For further insights on data center, visit our latest article: Deconstructing the Data Center: A Look at the Cost Structure Igniting the AI Boom!


[1] The cost of compute: A $7 trillion race to scale data centers - McKinsey


[2] Big Tech to report earnings under specter of AI bubble - Reuters


[3] "This Is How the AI Bubble Could Burst" - Plain English with Derek Thompson (Podcast, Sept. 23, 2025) https://podcasts.apple.com/us/podcast/this-is-how-the-ai-bubble-could-burst/id1594471023?i=1000728026459



[5] Musk's xAI nears $20 billion capital raise tied to Nvidia ... - Reuters



[7] Big Tech to report earnings under specter of AI bubble - Reuters (This is the same source as reference [2])


[8] "The AI that we'll have after AI" - Cory Doctorow (citing Princeton CTP research on 2-3 year GPU lifecycles) https://doctorow.medium.com/https-pluralistic-net-2025-10-16-post-ai-ai-productive-residue-5cc95ad345f9


[9] "Meta Platforms extends estimated life of certain servers, network assets" - TipRanks (Note: The link below from 'The Stack' is a direct syndication of this finding, confirming the $2.9B and 5.5-year claims.)


[10] "Longer-lasting servers will save Amazon almost $1 billion" - TechFinitive https://www.techfinitive.com/longer-lasting-servers-will-save-amazon-almost-1-billion/


[11] "Data center debt skyrockets 112% — $25 billion issued as AI boom fuels borrowing frenzy" - The Economic Times https://m.economictimes.com/news/international/us/data-center-debt-skyrockets-112-25-billion-issued-as-ai-boom-fuels-borrowing-frenzy/articleshow/125228221.cms


[12] "Alphabet reaffirms $75 billion spending plan in 2025" - Tech Monitor / Reuters (Note: This is the direct Reuters primary source for the claim.)



[14] "DOE Releases New Report Evaluating Increase in Electricity ..." - Energy.gov (via LBNL report) https://www.energy.gov/articles/doe-releases-new-report-evaluating-increase-in-electricity-demand-data-centers


[15] "Artificial intelligence doubles data center demand by 2030: EPRI" - Utility Dive https://www.utilitydive.com/news/artificial-intelligence-doubles-data-center-demand-2030-EPRI/717467/


[16] "Generative AI and the future of work in America" - McKinsey https://www.mckinsey.com/mgi/our-research/generative-ai-and-the-future-of-work-in-america



 
 
bottom of page