Beware the Debt-Fuelled AI Boom

If Wall Street giant Goldman Sachs’ crystal ball is to be believed, then by 2030 a staggering $5.3 trillion would have gone into an artificial intelligence (AI) infrastructure frenzy that began in earnest in 2025. In comparison, the cost of building the interstate freeway network (nearly 70,000 kms) in the US over 35 years was just $600 billion in today’s money. The only infrastructure buildout that tops AI’s explosive expansion is China’s modernisation. Starting 2010, China is estimated to have spent more than $15 trillion on various infrastructure projects.

So, whichever way you look at it, the AI boom is a watershed moment in human history. Aside from its disruptive powers, AI has been an unprecedented creator of wealth. Just the ‘Magnificent Seven’ (Nvidia, Apple, Alphabet, Amazon, Meta, Microsoft, and Tesla) have upped their market value from about $7 trillion by the end of 2022 to $22 trillion now. This, of course, doesn’t include the other newly-minted trillion-dollar companies like Broadcom and Micron.

What fuels this frenzy is the belief that AI will replace not just human cognition, but also labour, turning all of us into lotus eaters on the dole. The value of human labour today is estimated at $60 trillion. That, in effect, is the total addressable market (TAM) that the AI industry often talks about.

Until recently, the money that fuelled the AI boom was largely equity, including the cushy surpluses that the tech companies had accumulated over the years. That’s changing. Reuters and London Stock Exchange Group estimate that across 2025-27, the Magnificent Seven is set to add $340 billion of operating cash flow, but $534 billion of capex—that is, $1.57 of extra investment for every incremental $1 of cash flow. How will these so-called hyperscalers bridge the gap? Of course, by borrowing, and Morgan Stanley believes borrowings will set a new record in 2027.

To borrow is not a bad thing in itself. In fact, many savvy CFOs use leverage to increase return on capital and lower tax outgo. But when we borrow beyond our ability to service the debt—like the 2008 global crash proved—it’s usually a ticket to perdition. 

 In the case of AI, there are two potential problems with letting debt fuel the capex boom. One, is the cash flow from the investment and the risk of hardware obsolescence. Two, a lot of the debt is structured in an opaque, circular fashion, with risk seemingly going off balance sheet. Let’s dig into these two risks. 

The single biggest cost of a data centre—between 40% and 60%—is GPUs. Like any hardware, these GPUs have limited life. If these GPUs are used for training of large language models (LLMs), they burn out in two to three years. GPUs used for inference only (which means they are used for eliciting answers to prompts) last longer—as much as seven years.

The life of a GPU matters because it determines depreciation and reinvestment. Michael Burry, the Wall Street investor of the Big Short fame, began shorting Nvidia and Palantir because he believed they were inflating their profits by extending depreciation over six to seven years, instead of two to three. A sensitivity analysis by Goldman Sachs suggests that changing depreciation from five to three years would push cumulative industry depreciation from $3 trillion to $4 trillion—a $1 trillion swing from a single accounting change. 

Amazon is a case in point. In 2025, it decided to accelerate depreciation for a subset of servers from six years to five years, and took a charge of $920 million. Nvidia GPU renter CoreWeave is perhaps the best illustration of investors’ concern on extended depreciation. After this cryptominer pivoted to leasing GPUs last year, it became part of the AI mania. Its IPO opened at $40 a share and touched an all-time high of $183 before short sellers such as Jim Chanos pointed out that CoreWeave’s six-year depreciation policy was unrealistic and that a shorter depreciation cycle would actually balloon its losses. The stock now trades at $70.

 There are other concerns around the data centres. Most of these are funded via special purpose vehicles and the debt is sold like a stable income-producing real estate asset. This creates financing opacity. Consider the latest such deal in the works: Nvidia is in talks with OpenAI to stand guarantee for a $250 billion loan that would allow the ChatGPT maker to lease a 10GW data centre that a SoftBank subsidiary is building. Nvidia will have a separate deal with OpenAI to sell it, reportedly, $350 billion worth of GPUs. The deal isn’t done yet, but the Nvidia stock was down 1% on the news. 

Since a guarantee is not a loan, Nvidia only has to account for a small “fair value” of the guarantee on its books, and not the whole of $250 billion. This becomes a liability only if OpenAI is unable to service the loan it will be taking on the strength of Nvidia’s guarantee. This brings me to the next point. Most of the data centres are built for a single, specific customer. Unlike a commercial real estate with multiple tenants (hence lower risk of default), the lenders to a data centre are putting all their eggs in one basket. Many of these contracts allow the customer to terminate the lease if there is chronic performance failure. Flipping a data centre built for, say, OpenAI to an Anthropic isn’t easy. That’s because every data centre is precisely engineered to a customer’s specific needs, keeping in mind the kind of chips they use, power density, cooling method, and network shape. Should a tenant decide to walk out, then this data centre becomes a stranded asset. 

Ultimately, a data centre’s ability to service its borrowings will depend on the demand for compute. A recent survey by the Federal Reserve Bank of Atlanta reveals that the perceived AI productivity gains are larger than the measured gains. However, the executives polled expect the gains to double this year. Let’s hope they are right.

 If the expected income from these massive data centre investments fails to materialise, then the world may face a crisis bigger than the 2008 financial meltdown. India, with its underexposure to AI, is relatively safe. But that doesn’t mean its stock market won’t feel the shock waves from Wall Street—when they do come.