AI Debt Is Spiralling Out Of Control

Tech bros want us to be profoundly scared of AI. They bang on about how it will render us all obsolete, how it will soon become dangerously superintelligent, and how these systems are necessary for global economic viability.. Of course, these are all demonstrably manipulative lies, as I, and many others, have pointed out time and time again. So, don't fall for the propaganda. But there is one aspect of AI we should be worried about: debt. You see, the AI bubble in and of itself isn't horrific; it is the fact that, through debt, it is tied to the very foundations of our economy. In other words, when this bubble inevitably explodes, it will almost certainly take everyone down with it. We have known about this ticking time bomb for a while now, but as AI debt racks up, it is increasingly looking like a nuclear bomb. Let me explain.

The Debt

J.P. Morgan has calculated that AI-related debt now accounts for 14.5% of its $10 trillion investment-grade bond index, meaning they alone currently hold $1.5 trillion in AI debt. Others believe AI debt makes up 15–20% of most corporate bond indices, suggesting that the current total AI investment-grade debt (bonds) could be much higher than $1.5 trillion. In fact, some have discovered that AI is linked to more debt than banks! That might sound like a lot, but don't worry — it is only going to get considerably worse. Morgan Stanley estimates AI hyperscalers could raise $400 billion in corporate bonds (debt) in 2026 alone for AI data centres, and Gartner estimates that AI hyperscalers are going to spend $690 billion on AI infrastructure in 2026, with most of that funded by debt.

This involves a lot of big numbers, so let's put it all into context.

The 2008 financial crisis is not a perfect analogy for the AI bubble by any means, but it can help offer some perspective. The crash occurred when roughly $1.4 trillion (roughly $2.1 trillion in today's money), or about 40–45% of the subprime mortgage-backed securities (find out what they are here), were wiped out as their value was either written down or debtors defaulted. This caused the US bond market to take a record loss of 4.94%. As such, this figure heavily suggests that the dodgy subprime mortgage-backed securities made up less than 10% of the bond market at the moment it all collapsed.

Because the financial landscape is so different, we can't draw any direction comparisons here. But we can say that AI debt might already be at the same level as subprime mortgage debt was in 2008, and that AI debt is currently taking up considerably more of the bond market than subprime mortgages did in 2008. So, in simpler terms, there is enough debt to cause a serious problem if these AI bonds default or are written down, and the entire financial system is far more exposed to this risk than it was to subprime mortgages back in 2008.

That should give you an appropriate sense of scale. But all of this debt is only a problem if those who take it on don't have the profits to pay it back. So, can the AI industry cover these costs? And what happens if they can't?

The Problem With AI Data Centres

Well, sadly, AI data centres aren't what you would call profitable.

According to Harris Kupperman, founder of the hedge fund Praetorian, these data centres make zero sense. He found that AI datacentres built in 2025 will incur $40 billion in annual depreciation while generating only $15 billion to $20 billion in annual revenue. In other words, they are losing money hand over fist in real-world terms.

Why is the depreciation so high? Well, the chips used in these data centres don't last very long at all.

Meta found that their AI data centre chips failed at a rate of 9% per year, meaning that after three years of operation, the data centre would lose a whopping 27% of its capacity if these insanely expensive chips weren't replaced. That is enough to make their operation unviable. Moreover, computer technology is still advancing rapidly and becoming more efficient. As a result, after just a few years of operation, it will be cheaper and more efficient to replace all the chips in an AI data centre. Therefore, the more realistic lifespan of an AI data centre is three years.

But, as Michael Burry has pointed out, AI data centre operators' books do not include this. They claim that these data centres have a lifespan of five or more years, which significantly underestimates depreciation and paints a much rosier picture than reality. And even then, most of these AI data centres are not profitable.

Ultimately, AI data centres pretty much function as cash-burning machines. This is surprising, as demand is hugely outstripping supply, so you would expect them to be the most profitable they could ever be right now.

The lack of profitability is a significant problem, as the vast majority of AI-related debt and bonds are used to build and operate data centres. These are not solid financial grounds on which to lend multiple trillions of dollars, let alone sell that debt as investment-grade bonds.

The Problem With AI

Okay, but what about the AI companies that pay to use the data centres? Surely they can step in and support the AI data centre operators?

Well, no, they are even more screwed than the data centre operators.

Quite simply, none of them has a viable route to profitability. Anthropic is arguably doing the best out of all of them, and it is estimated to have posted a $5.2 billion EBITDA loss in 2025. EBITDA stands for Earnings Before Interest, Taxes, Depreciation, and Amortisation, so this figure doesn't account for any losses incurred by Anthropic due to data centre depreciation, which it will incur as it transitions to owning and operating its own data centres. So its real-world losses might be far larger. OpenAI's costs are also growing exponentially faster than their revenue, with estimates suggesting they incurred a net loss exceeding $15.6 billion in 2025.

That is a sizeable gap to bridge before these companies can even begin to operate sustainably, let alone have the cash reserves to perpetually bail out the third-party data centres they use.

But both companies are effectively doing this anyway by borrowing tens of billions of dollars to build and operate their own AI data centres. They are justifying these insane expenditures by claiming their revenue will skyrocket as AI replaces white-collar workers across the globe, which will not only pay for these data centres but finally propel them to profitability!

Yeah, I call bullshit.

For one, even OpenAI has admitted that just shoving more data and more compute power into AI won't make it any better (read more here). This, combined with AI's diminishing returns (read more here), means that AI is about as good as it will ever be.

Yet the current models are still utterly useless. A recent study found that the very best AI models fail to complete more than 97% of the freelance work assigned to them! AI simply isn't good enough to automate jobs. This is why research has found that AI-driven job layoffs are a myth, given that AI fundamentally can't replace us. But it also can't successfully augment us either; AI is not making workers more productive and is, in fact, seemingly intensifying burnout. Also, it is not as popular as you might think. Only 5% of ChatGPT users pay for the service, which is an abysmal conversion rate. On top of that, as Ed Zitron has pointed out, OpenAI is likely gaming its user numbers, casting doubt on its 800 million active weekly users figure.

Quite simply, the models are not capable enough to replace jobs, and they need to be in order to bring in enough money to break even on the investment and associated debt these AI companies have taken on. But beyond that, these models aren't good enough to augment jobs or provide enough value to our daily lives to be wildly popular, and the science suggests that they will not improve. These companies will almost certainly never reach profitability and will instead be buried under their own debt and cash burn.

Indeed, even mainstream media is beginning to predict their downfall. The New York Times' Sebastian Mallaby has predicted that OpenAI will declare bankruptcy between now and sometime in 2027, and I agree.

So, no, these AI companies can't absorb the losses of AI data centres, as they themselves are drowning under their own losses and will likely go under sooner rather than later.

All of this suggests that the trillions of dollars' worth of AI-tied debt isn't especially secure and shouldn't be labelled as "investment grade", right? They are lent to risky businesses that are almost certain to fail. So, surely when these AI data centres and AI companies inevitably fold, and a colossal amount of these bonds fail, won't that cause a recession?

The Repurpose Argument

There is an argument against the idea of a recession, and that is to compare the AI bubble to the dot-com bubble. You see, much of the dot-com bubble actually involved building out internet infrastructure, like fibre-optic cables. Half of the reason the dot-com bubble was so brutal was because this infrastructure was location- and purpose-specific and so couldn't be used, rendering it a total write-off. Some have suggested that, even if the AI bubble bursts, there will still be demand for the data centres, as cloud computing and smaller AIs will still be interested in using them.

All I can say in response to that is that these people are not taking supply and demand into account. These data centres are experiencing peak demand right now, and if the AI bubble bursts, that demand will greatly diminish, creating a huge supply-side surplus and crashing the price of cloud computing across the board. Depending on how many AI data centres are built and how badly the AI bubble bursts, this could send the entire cloud computing industry into negative profit for years.

So, you can't simply argue that these trillions of dollars in AI bonds will survive the AI bubble crash because AI data centres will still be used. There is likely no backstop here. If the bubble pops, this colossal amount of debt is going to paralyse our financial system.

Summary

I'm sure I do not have to spell out why enduring another 2008, especially now, during a cost-of-living crisis and the current rise of neo-fascism, isn't exactly what we need right now and will cause a truly horrific amount of suffering. As I said, the AI industry hasn't so much as placed a ticking time bomb at the heart of our economy but a nuclear bomb. Yet we are seemingly sleepwalking towards this car crash whilst refusing to even acknowledge that the brakes exist. This entire thing is out-of-control.

Thanks for reading! Everything expressed in this article is my opinion, and should not be taken as financial advice or accusations. Don't forget to follow me here or support me over at Substack.

(Originally published on PlanetEarthAndBeyond.co)