AI债务问题比想象更严重：科技巨头隐藏1.65万亿美元表外债务

As you probably already know, in the past couple of years, AI-adjacent companies, some of which were once flush with cash, have had to borrow absurd amounts of money to finance their AI investments. In just the first 7 months of 2026, US tech companies have already borrowed more than $300 billion, and JP Morgan expects them to issue an additional $200 billion before this year is done. This will take the total issuance to about $500 billion, which would account for 20% of all debt issued in US markets this year.

To put this in context, in the build-up to the dot-com bubble, tech-related debt peaked at 14% of total issuance. So, this implies that AI spending is ramping up faster than in any other comparable historical episode, including the railway boom. However, it turns out this is only the tip of the iceberg, and the AI companies have been using a whole load of financial trickery to borrow trillions more off book. So, in this video, we're going to explain how these companies are able to effectively hide trillions of dollars of debt, and what risks this might present for the sector as a whole.

Which flag is better, Canada or Iceland? How about Japan or Mexico? Albania or Barbados? How about a battle of the dragons, Wales or Bhutan? Cast your vote in our official flag ranking survey by clicking the link in the description. So, this story sort of began last week when Nikkei published a report which claimed that five of the biggest AI-adjacent companies in the US, namely Alphabet, Microsoft, Amazon, and Oracle, were hiding an additional $1.65 trillion in off-the-book debts, more than the roughly $1.35 trillion dollars of debt that they've explicitly admitted to on their balance sheets.

This was all confirmed by another investigation by Bloomberg, published at about the same time, which found that, like its fellow tech giants, Nvidia also had more off than on-book debt. Now, you might assume that shady financing is the norm among big US corporations, but while there always has been a bit of it going on, the volume of hidden debt has ballooned over the past couple of years. And unsurprisingly, most of this increase can be attributed to AI-related costs.

There are two main methods by which these companies have been able to, quote-unquote, hide this debt. The first is via what are known as long-term purchase agreements, which are basically when, instead of borrowing money in order to buy something, you commit to buying something in the future. This still creates a debt-like liability. After all, you now owe someone money in the future, but you don't have to chalk it down as a loan on your balance sheet.

In the context of the AI boom/bubble, these long-term purchase agreements are mainly used for chips, with lots of AI companies committing to buying some very expensive high-end chips in the future. Nvidia, for instance, has $119 billion in binding, non-cancelable future purchase obligations, primarily with manufacturing partners like TSMC, for bespoke pr