In that equation that the world is trying to solve with AI, there is a half that not many people have noticed: debt. Behind every AI-generated chat and video is a gigantic network of data centers, and those data centers are being financed with a mountain of borrowed money. And therein lies the problem. In what is borrowed.
Debt and more debt. According to recent datathe issuance of secured debt linked to data centers in the United States is estimated to be $25.4 billion by 2025. It is 112% more than the previous year. If we add up all the complex financial instruments (known as asset-backed securities (ABS) and commercial mortgage-backed securities (CMBSS)), the snowball is already huge: there are almost $49 billion tied to these securities.
Bonuses for everyone. Here there are not only startups asking for loans, no. The technology giants that are setting up these infrastructures – the so-called hyperscalers – are also taking advantage of this mechanism. Companies such as Microsoft, Google, Oracle or Meta have rediscovered the bond market as a source of financing.
Better to spend what is not mine. They all have huge amounts of money, but instead of spending their own cash, They have raised 100,000 million dollars in debt issues so far this year. The goal: buy thousands of GPUs and build data centers before the competition.
What are you doing, Oracle? If there is a company that embodies the vertigo of this excessive bet, it is Oracle. The company created by Larry Ellison has committed to meeting a Pharaonic $300 billion deal with OpenAI. That has forced it to become the largest issuer of corporate debt (outside the financial sector). The numbers are scary: your total debt has grown to 111.6 billion dollarswhile its cash has dropped by 10,000 million. Citi estimates they’ll need to borrow another $20 billion to $30 billion every year (every year!) for the next three years just to keep building.
excessive ambition. There are also examples of startups that are exploiting this facet. One of the clearest is the one from CoreWeavea company famous for renting computing capacity for AI. The company has secured credit lines of $2.5 billion backed by leading investment banks such as JPMorgan. The market message seems clear: “if you’re going to build for AI, here’s the money.”
How to get a 30-year mortgage. Analysts of all kinds have been keeping the fly behind their ears for some time, and one of the latest Moody’s reports is a good example. Concrete buildings are usually financed with terms of 20 or 30 years, but the technology inside (such as AI chips) changes radically every 3 or 4 years. Does it make sense to go into debt three decades from now for a technology that evolves so quickly?
cheap money. Investors are also agreeing to charge minimal interest, just 1% above what the safe US public debt pays, when they assume that risk. It’s a worrying classic sign of euphoria. There is so much money wanting to enter the sector that those who lend it have lowered their guard and demand very little return for their risk. They firmly believe in the promises of AI while increasingly more analysts warnhorrified, that we are facing an “irrational exuberance.”
Having money is no longer enough. All this is already scary, but the real bottleneck for expansion is not even capital or chips, but the electrical grid. As Satya Nadella, CEO of Microsoft, pointed out, there is no power for so many chips. The situation is so worrying that a Deloitte study indicated in a study that there are a seven-year waiting line to connect some data center projects to the electrical grid. And if companies want to obtain financing, they need have guaranteed electricity supply for your data centers. If there is no plug, there is no loan.
Big Tech looks for electrons. At OpenAI they already warned of the problem months ago when talking about the “electron gap” describing electrons (energy) as the new oil. Almost all the major companies in the industry are making a move. Google has signed an agreement with TotalEnergies to be delivered 1.5 TWh of electricity over the next 15 years, and Meta did something similar with Treaty Oak Clean Energy to get 385 MW of its solar plants in Louisiana.
The bubble before the big question. All of this further increases the fear that the AI bubble will end up bursting in a big way. Meanwhile, the big unknown is whether the demand for artificial intelligence will be capable of paying the immense electrical and financial bill that it is signing today in 5 or 10 years. The credit party continues.
In Xataka | While Silicon Valley seeks electricity, China subsidizes it: this is how it wants to win the AI war
GIPHY App Key not set. Please check settings