Artificial Indigestion

The idiom ‘bite off more than you can chew’ means to take on more work, responsibility, or commitments than you can realistically manage or complete. It’s a good expression to describe what’s happening in the AI world today. We’re witnessing the biggest capex splurge in the history of the world as Big Tech hyperscalers race to build data centre capacity to run their AI models.

The prevailing narrative promises an era of unlimited computing power and untold wealth for those able to harness it. But there’s one big question hanging in the air that no-one seems to have a clear answer for: Who’s going to pay for this?

Hidden in the financial footnotes of the top hyperscalers and chipmakers lies $3 trillion of off-balance sheet commitments, special purpose vehicles and vendor credit backstops, a number that’s increasing rapidly. It’s important to understand that these liabilities don’t appear on their balance sheets.

Nvidia, for example, carries only ~$10 billion in reported corporate debt, but holds over $530 billion in off-balance-sheet supply commitments, residual value guarantees, and credit backstops. To help underfunded ‘neoclouds’ buy its GPUs, Nvidia provides minimum collateral value guarantees to private lenders. It has also partnered with Apollo, BlackRock, and Blackstone on a $500 billion compute-financing platform, backstopping up to $125 billion to lower borrowing costs for data centre developers.

Simply put, Nvidia is arranging half a trillion dollars in loans so clients can buy its hardware, then guaranteeing the debt if they can’t pay. ‘Circular financial engineering’ is a good way to describe this practice, and almost every major player in the AI space is engaged in some form of it on an unprecedented scale.

This positive feedback loop has resulted in an earnings bonanza for hardware suppliers connected to the AI infrastructure trade. But are the record profits and stock prices of these businesses sustainable? The central question for investors is whether AI end-user revenues can grow fast enough to justify all the capital expenditure and cover the mountain of debt it’s created, especially as interest rates rise.

The macro numbers reveal a staggering monetization gap. Analysis by Bain & Company indicates that the global AI infrastructure buildout requires $6 trillion in AI-related revenue by 2031 to justify the data centres currently under construction. However, current consumer subscriptions and enterprise software productivity gains are tracking to deliver only $1.2 trillion to $1.8 trillion, leaving a massive $4.2 trillion global revenue shortfall that companies like OpenAI and Anthropic must solve by creating entirely new product categories.

But the bills start coming due even before 2031. There’s a critical debt maturity wall between 2027 and 2029, when hundreds of billions in private credit notes will face mandatory repayment or refinancing (at much higher projected interest rates). If enterprise AI adoption scales at a normal corporate software pace (20%–30% annually) rather than the projected hockey-stick curves, or if monetization stalls, a default-driven ‘Subprime Data Centre’ crisis becomes a real hazard. Under this scenario:

  1. AI labs fail to meet non-cancellable lease obligations.
  2. Lenders trigger vendor credit guarantees, legally shifting tens of billions in senior debt back onto the balance sheets of Nvidia, Broadcom and other hyperscalers.
  3. Credit rating downgrades ripple across Big Tech, while millions of liquidated GPUs flood the secondary market, severely eroding hardware pricing power and gross margins.

Who’s going to pay for this? Unless they’re able to chew and swallow what they’ve already bitten off, it could well be the chipmakers and hyperscalers themselves, along with their private credit enablers.

Nvidia especially carries a unique risk in that its revenue and profitability are directly tied to its contingent liabilities. The primary revenue engines of businesses like Google, Meta and Microsoft are independent of AI infrastructure spend. If AI monetization stalls, they’ll be saddled with a lot of debt, but their revenue streams won’t be directly affected. Nvidia however will be hit on both the debt and the revenue fronts simultaneously.

Granted this is a bearish scenario, but analysts across the board are starting to wake up to the reality that this isn’t a low-probability outcome by any stretch. Michael Burry of The Big Short fame is among several notable short sellers taking significant positions against the AI trade. Even the baseline expectations of the major Wall Street banks are starting to look a little less rosy.

This matters because we’re talking about the biggest companies in the world, rather than some niche corner of the market (like gold miners). These are the companies that have dominated the major cap-weighted benchmark indices for more than a decade. Nine out of the ten biggest weightings in the S&P 500 are Big Tech, accounting for more than a third of the index. Nvidia alone accounts for 8%. Roughly half of the index is directly connected to the AI supply chain. It’s a similar story on the global scale, where US Big Tech and the AI trade dominate global benchmark indices.

To demonstrate the extent of this dominance, consider that the market-cap weighted MSCI AC World index has returned more than double its equal weighted counterpart (the exact same companies weighted equally) over the last 10 years, and that ~75% of that outperformance is attributable to US Big Tech alone. In other words, a very small group of massive US tech companies has been driving global stock market returns, while the average company’s return has been comparatively pedestrian. We’ve been arguing for some time that this trend can’t continue forever. This speculative debt binge may well be the catalyst for a reversal.

Who are the winners in a scenario like this? Defensive industries, non-US markets, small- & mid-cap stocks. Basically, everyone who benefits from cheap excess computing power who aren’t carrying the bag for the upfront infrastructure cost.

It’s instructive to look at the run-up to the dot-com bubble and its aftermath. In 2000, index concentration was extreme. (Today it’s even worse.) Cap-weighted indices outperformed their equal-weighted counterparts significantly in the late 90s. Telecom companies spent billions building fibre-optic cables but didn’t have enough consumer demand to monetize them immediately. The decade that followed is famously referred to as the ‘lost decade’ for US mega-cap equities (returns were negative), while ‘old economy’ sectors, international & emerging markets, and smaller companies delivered double-digit annualized returns. And the internet was a significant tailwind for them.

While there are differences between the current AI trade and the dot-com bubble, there are enough similarities for us to ask: Might we be on the verge of another lost decade for US mega-caps? The risk is very real. If this happens, investors who’ve piled into popular thematic ETFs and market trackers are going to be sorely disappointed.

But what if all this AI investment actually pays off and the technology successfully delivers on its transformative promises? Does it follow that US Big Tech will continue to outperform everything else? While a successful AI expansion validates the massive capital outlays of the hyperscalers and chipmakers, market history and financial mechanics show that the primary beneficiaries of a technological revolution often shift as the technology matures. In a fully realized AI bull case, economic value naturally diffuses across the broader market in several ways:

  • The Shift from Infrastructure to Downstream Productivity: In almost every historical infrastructure boom – from railroads and electricity to the internet – the upfront buildout creates immense initial wealth for infrastructure providers, but the long-term compounding gains shift to downstream adopters. If AI models successfully automate white-collar workflows and industrial processes, non-tech sectors like healthcare, financial services, logistics, and manufacturing will see substantial margin expansion by lowering operational overhead, all without having to carry billions of dollars in data centre debt or hardware depreciation.
  • Valuation Baselines and Multiple Compression: US Big Tech companies are already trading at premium valuation multiples, meaning much of their projected future earnings growth is already priced into their stock prices. If a mega-cap tech firm grows earnings rapidly but its valuation multiple compresses as its cloud compute business matures into a utility-like model, its stock returns may be modest. Conversely, small-cap companies or non-US equities trading at steep valuation discounts that experience AI-driven productivity boosts can enjoy both earnings growth and multiple expansion, generating superior relative stock returns.
  • Compute Deflation and Commodity Dynamics: If the AI bull case succeeds because algorithmic breakthroughs and efficient chip architectures make computing power dramatically cheaper, AI compute becomes a pervasive, low-cost commodity. Cheap compute is a massive tailwind for application developers, enterprise end-users, and consumers, but it can eventually squeeze the premium gross profit margins of the cloud hyperscalers and chip designers who built out the physical infrastructure.
  • Mathematical Concentration Limits: US Big Tech already accounts for more than a third of the S&P 500’s total market capitalization. For these few companies to indefinitely outperform the rest of the global stock market in a thriving, growing economy, market concentration would have to reach unprecedented levels where a handful of firms capture an unfeasible share of broad GDP. Market capital historically diffuses as productivity gains spread throughout the rest of the corporate ecosystem.

In short, the bull case for AI can be completely right about the technology’s economic impact while still ushering in a market environment where capital-light software builders, non-tech enterprise adopters, small- & mid-cap stocks, and international markets generate stronger relative investment returns than the mega-cap infrastructure builders carrying the upfront capital costs.

What made US Big Tech so successful for so long was that they were able to generate high returns on relatively modest capital investments with virtually no debt. The AI trade has fundamentally changed the economics of these businesses, transforming them into highly indebted, capital-intensive quasi-utilities.

The point of this article isn’t to predict a specific outcome so much as to highlight a very real risk facing the world’s biggest companies and major benchmark indices, as well as the myriad of ETFs which tracks them. Whichever way this AI trade ultimately goes, the current risk-return profile appears to favour benchmark-agnostic active investors willing to look for value beyond US Big Tech.