This fall, Microsoft announced a series of deals, totaling tens of billions, to lease computer power for its artificial intelligence ambitions. Meta secured almost $30 billion in financing to build a massive data center in Louisiana without taking on the debt itself. Google also committed to rent computing power from a small company and then sell some of it to OpenAI.
Those deals had one thing in common: They allowed companies that make massive quarterly profits to reduce their financial exposure to the frenetic, global buildup of data centers.
They also signaled new ways the biggest companies in tech are maneuvering to push some of the risk of the A.I. boom onto the shoulders of upstarts eager for a piece of the action. The moves let companies like Meta and Microsoft add computing power quickly and then wait to see how demand for A.I. shapes up before committing to projects that can last for decades.
Trillions of dollars are at stake as tech companies try to predict how much computing power A.I. will demand years down the line. If the big companies decide they don’t really need all that computing after the deals are over, the smaller companies and their lenders will be stuck with the consequences.
“Risk is like a tube of toothpaste,” said Shivaram Rajgopal, an accounting professor at Columbia Business School. “You press it here, it’s is going to come out somewhere else. It’s always in the system, it’s a matter of where.”
These deals also add a level of mystery to data center financing because many companies running the data centers for the tech giants are far from the household names of Silicon Valley. Some are privately held, do business with large start-ups, and borrow from private lenders, all of which offers less transparency about their stability.
Meta’s data center project in Louisiana mixes many of those creative financing elements into a multibillion-dollar plan taking shape among farmlands in the northeast corner of the state. Meta created a so-called special purpose vehicle named Beignet Investor LLC and worked with Blue Owl Capital, a private credit firm, to borrow money for the project.
Meta was responsible for constructing the data center, but Blue Owl was on the hook for 80 percent of the financing. As part of the arrangement, Meta agreed to “rent” the data center from Beignet with a series of four-year leases. That allows the tech giant to categorize the funding as operating cost, not debt, according to financial filings.
As part of the deal, Meta is paying a premium to Blue Owl so it doesn’t have to borrow the money itself, said Solomon Feig, a private credit lender at Pinnacle Private Credit. “Instead, Meta is renting risk,” he added.
Blue Owl primarily funded the project, called Hyperion, through a bond offering from Pimco, an asset management firm. Pimco, in turn, sold the so-called “Beignet bonds,” which mature in 2049, to its clients that include insurers, pension funds, endowments and financial advisers. BlackRock also bought some of the bonds.
“The key part of Meta’s strategy, in my view, is that they’re going to get as much of this built out with what the industry calls OPM: Other people’s money,” said Andrew Rocco, a stock analyst at Zacks Investment Research.
If the A.I. boom were to slow, Meta can walk away from the deal as soon as 2033. How much it may have to pay depends. Blue Owl could find a new customer, or sell the project, though the data center’s value could depreciate if demand for A.I. underwhelms. Meta promised to cough up enough cash to effectively repay the underlying debt without formally putting that debt on its books, said S&P Global, the ratings firm.
Other tech companies have expressed interest in similar financial structures. But since this was the first-of-its-kind, Meta provided more protections than future deals may require, according to two people involved with the deal who spoke on the condition of anonymity.
Mr. Rajgopal cautioned that the arrangement recalls other investment booms that relied on private credit and special purpose vehicles, which are less transparent ways of raising funds than through the traditional banking sector.
“I thought we had solved the off-balance-sheet problem,” he said, referring to the accounting methods banks used in the lead up to the dot-com bubble of the 2000s. “This is like Groundhog Day all over again.”
Meta declined to comment.
Big tech companies are also signing massive deals with a new generation of data center providers known as neoclouds. Typically contracts for three to five years, these deals give them more computing power quickly, without locking into decades long commitments.
By agreeing to shorter deals, big companies can get computing power that shows up in their financial reporting as a day-to-day operating expense, rather than a long-term capital investment, which can spook investors.
In September, Microsoft signed a $17 billion deal with Nebius, a neocloud from a founder of the Russian internet giant Yandex. In October, it committed $23 billion, according to Bloomberg, to get data centers from Nscale, a privately-held British neocloud. In November, it agreed to a $10 billion deal with Iren, a former bitcoin miner, and another multibillion-dollar agreement with Lambda, another neocloud.
Microsoft’s global infrastructure approach is built on flexibility “based on the near-term and long-term demand signals we see from customers,” said Alistair Speirs, a Microsoft executive, in a statement. The company has been racing to get enough computing power to keep up with customer demand, and its priority is to quickly meet needs around the globe, Mr. Speirs added.
But if demand shifts, “you don’t want to be upside down,” Satya Nadella, Microsoft’s chief executive, said in April.
That focus on flexibility became apparent when Microsoft quietly started pausing some of its building last fall. That pause coincided with the company renegotiating its relationship with its key A.I. partner, OpenAI, though a Microsoft spokesman denied it was related.
Microsoft, which had an exclusive arrangement with OpenAI, allowed its partner to begin getting some computing power from the software company Oracle and others. This fall, Microsoft began signing the slate of shorter-term neocloud deals, some of which will help supply the $250 billion in computing that OpenAI has promised to send Microsoft’s way.
(The New York Times has sued OpenAI and Microsoft, claiming copyright infringement of news content related to A.I. systems. The companies have denied those claims.)
Microsoft and Google have arrangements to provide computing power to OpenAI. They will supply some of that through deals with CoreWeave, the biggest neocloud provider. OpenAI has also committed to buy as much as $22.4 billion in computing power directly from CoreWeave.
To build its computing capacity, CoreWeave is taking on billions of dollars in debt, much of it at interest rates at 10 percent or higher. In effect, CoreWeave has tied its future to OpenAI.
When asked about the OpenAI relationship, a CoreWeave spokeswoman said the company is aggressively diversifying its customer base. She said no single customer represents more than 35 percent of its future revenue under contract, but declined to say how much of its computing power would ultimately be used by OpenAI.
Industry experts say the costs of the A.I. build out have gotten so large that it is no longer possible to mitigate most of the risk. So the tech giants are spreading it around.
“It is very savvy of them,” said Alex Platt, an analyst at the investment bank D.A. Davidson. “There are only a handful of companies that can even do it.”
Cade Metz contributed reporting.
Karen Weise writes about technology for The Times and is based in Seattle. Her coverage focuses on Amazon and Microsoft, two of the most powerful companies in America.
The post How Tech’s Biggest Companies Are Offloading the Risks of the A.I. Boom appeared first on New York Times.
