This year, Microsoft, Meta and Google simultaneously announced agreements to lease computing power worth tens of billions of USD, reflecting a new trend in the artificial intelligence (AI) race: expanding infrastructure quickly but minimizing financial risks.
Meta has raised nearly $30 billion to build a giant data center in Louisiana without having to shoulder the debt.
Google leased the computing capabilities to a small company, then sold a portion to OpenAI.
Microsoft has signed consecutive short-term data center lease contracts with the next generation of suppliers, neoclouds (a group of companies providing specialized cloud computing infrastructure).
The common point of these agreements is that technology corporations with huge quarterly profits avoid having to spend long-term on infrastructure projects that have lasted for decades, in the context of uncertain demand for AI.
Tories of billions of dollars are at risk as no one can accurately predict the amount of computing power AI will need in the future.
"The risk is like a toothpaste. You squeeze it one place, it will float to another, Professor Shivaram Rajgopal ( Columbia Business School) said.
According to him, the problem is not the risk of disappearance, but who is it being transferred to.
A typical example is Meta's Hyperion data center project in Louisiana. The group established a special-purpose company called Beignet Investor LLC, in partnership with Blue Owl Capital to borrow capital.
Meta builds a data center, but Blue Owl covers up to 80% of the financial costs. In return, Meta leased the data center through 4-year contracts, allowing this cost to be classified as operating cost rather than debt.
According to analysts, Meta is paying an "insurance fee" to avoid borrowing capital from itself. Instead of borrowing, Meta is hiring risks, said Solomon Feig, a private credit expert.
The majority of the capital was mobilized by Blue Owl through bonds maturing in 2049, sold to insurance funds, pension funds and large investors such as BlackRock.
If AI demand slows, Meta could withdraw from the deal from 2033, while the remaining risks will be on creditors and financial partners.
This approach has many experts concerned about the return of outside the accounting balance sheet items, which contributed to the dot-com bubble inflation in the early 2000s.
Not only Meta, Microsoft also pursues a similar flexible strategy. In 2024, the company signed a series of contracts worth tens of billions of USD with neocloud platforms such as Nebius, Nscale, Iren or Lambda.
Short-term contracts from 3 to 5 years help Microsoft quickly add computing capabilities, while avoiding the risk of "holding" redundant infrastructure if demand changes.
The rise of neocloud providers such as CoreWeave shows that risks are being pushed deeper into the financial system.
CoreWeave borrowed billions of dollars at high interest rates, tightly tying its future to OpenAI contracts.
As the cost of building AI is getting more and more expensive, experts say that technology giants are forced to spread the burden and the risk game is just getting started.
