Apple, Alphabet, Microsoft, Oracle, Meta, and Nvidia spent $1.1 trillion on share buybacks in 5 years to pump up their shares. That’s at risk.
By Wolf Richter for WOLF STREET.
One of the big drivers of the surge of stock prices over the past many years has been the prodigious amount of corporate cash spent on share buybacks by Big Tech and other companies. Some of the share buybacks were funded from cash flow, some with borrowed money. Share buybacks are not like stock trading; they represent new money entering the stock market to remove shares from the stock market. Their purpose is to drive up the price of the remaining shares.
For example, from Q3 2020 through Q3 2025, six companies – Apple, Alphabet, Microsoft, Oracle, Meta, and Nvidia – spent $1.1 trillion on share buybacks (data via YCharts). These are amounts actually spent on share buybacks, not the announcements of future share buyback plans.
I’m leaving out Amazon for a reason: It already stopped share buybacks in 2022 to spend that cash on capital expenditures, such as AI infrastructure. And others may have to follow. More in a moment.
Apple topped the list with $437 billion in share buybacks over those five years, followed by Alphabet ($281 billion), Meta ($151 billion), Microsoft ($107 billion), and Nvidia ($87 billion). But Nvidia’s buyback program was ramped up in 2024; and over the past four quarters, Nvidia spent $43 billion on share buybacks.
Some of these share buybacks were funded with borrowed money by issuing bonds. And it shows on their balance sheets: Apple now has $112 billion in short-term and long-term debt, Microsoft $120 billion, Meta $50 billion, Alphabet $30 billion.
These companies are now engaged in an AI-spending war, on the premise that whoever spends the most wins?
The amounts of spending on capital expenditures, such as for data centers, are mindboggling. In the third quarter alone, just four companies – Microsoft, Amazon, Alphabet, and Meta – spent $114 billion; and they’re on track to exceed $400 billion in capital expenditures this year. And they all said they’d accelerate their spending further in 2026. And they’re not the only ones doing it.
Where does all this money come from?
Some of it comes from operating cash flow, some of it from short-term investments that they accumulated over the years, and a lot of it comes from massive amounts of new debt issuance.
Over the past three months, five companies alone – Alphabet, Amazon, Microsoft, Meta, and Oracle – have issued $88 billion in new investment-grade bonds, according to Dealogic: $18 billion in September, $30 billion in October, $40 billion in November.
The investment-grade bond market may have to absorb $1.5 trillion in AI data center bonds over the next five years, according to J.P. Morgan analysts, cited by Reuters.
Meta has come up with a way to keep a recent $27 billion AI bond sale off its own balance sheet in order to avoid having its ‘AA-‘ credit rating dented. This deal involves Meta and Blue Owl Capital. The issuing entity was a joint venture between Meta and Blue Owl Capital, which sold most of the bonds to Pimco. Meta is going to lease the computing power, but in four-year terms, to be renewed every four years, to avoid “finance lease” accounting, which would have put the liability on Meta’s balance sheet.
Ratings agencies are getting nervous about this type of off-balance-sheet debt. Fitch said today in a report on the circular hocus-pocus deals, as I’ve come to call them:
“Opacity is a rising risk factor, with limited public disclosures from AI model developers and increased use of off-balance-sheet debt from some public issuers [such as Meta]. Uneven disclosure on customers and contract terms obscures the extent of counterparty exposure.”
What’s left for the share buybacks…
But these are precisely the companies that have spent tens of billions of dollars every year to buy back their own shares – some of it with borrowed money.
Amazon has already spoken on this. Despite the super-hyped announcement in March 2022 that its board had authorized the company to buy back $10 billion in shares, the last share buybacks occurred in Q2 2022, and most of the authorization remains unused. Amazon has said that it has better things to do with the cash than blow it on share buybacks – it didn’t quite phrase it that way, it couched it in delicate corporate speak, but you get the idea.
Others will likely follow, maybe not this quarter or next quarter, but sooner or later, as the mindboggling spending amounts and the debt pile-up – even if it it’s off-balance-sheet – will make blowing huge amounts of cash on share buybacks an increasingly at-risk proposition.
Nvidia is now tangled up in massive circular deals with its own customers, such as OpenAI ($100 billion), and other companies. Those were just announcements at this point, but eventually the cash will start circulating. What will be left for share buybacks then?
There still will be the much-hyped announcements of share buyback “authorizations,” but actual share buybacks might get quietly reduced or put on ice entirely, as every cent that isn’t nailed down gets plowed into the AI-spending war.
Share buybacks incinerate corporate cash for the purpose of driving up share prices. Plowing that cash instead into the AI-spending war, in the worst-case scenario, would be no worse than burning it on share buybacks, but it would actually accomplish something real: Building AI data centers. Not everyone might be happy though: Stock prices might suffer when that prop gets pulled out from under them.
Enjoy reading WOLF STREET and want to support it? You can donate. I appreciate it immensely. Click on the mug to find out how:
The Verge recently had an interesting article on circular hocus-pocus, “I looked into CoreWeave and the abyss gazed back”
“Meta is going to lease the computing power, but in four-year terms”
In a recent interview the OpenAI CFO said that a new data center cost $45-50 billion. 15 Billion for the plant and 30 billion for “compute”. Compute being rows and rows of servers packed with NVIDIA Hyperscaler processors.
Meanwhile NVIDIA tells us that the current generation of processor package has a life of about 3-4 years – after that it is either obsolete or broken due to the temps the SC are necessarily exposed to (I’ve never been clear on what the cause of End-of-life is)
If this is remotely true, it explains why Meta would want to limit it’s lease to 4 years (besides the accounting benefit). Question is what happens after EOL ? Has the DC made enough money from it’s lease to invest in refitting it’s compute with a new generation of processors ? Has the price of compute declined really fast so that the next generation only costs 10 billion rather than 30 billion ?
Maybe the CFO gal was just being a bit undisciplined with her interview. It is after all, in that same interview that she suggested that the Federal government should backstop Open AI’s trillion dollars in buildout commitments….
Can it possibly be true that Open AI expects 2025 sales to be $13 billion while the company expects to lose $9 billion ???? Who is buying these bonds – PIMCO you say ???
“If this is remotely true, it explains why Meta would want to limit it’s lease to 4 years (besides the accounting benefit).”
The renewal is not optional. It’s required. For 20 years. Meta cannot walk away. That was a requirement to sell the bonds. Which is why this deal is so iffy from an accounting perspective.
I am reminded of animals with huge exotic antlers evolved for dueling. I also think of a “winner’s curse” where one overspends, to best the other player. Competition sometimes does interesting things with resource allocation. As equity holders, we are somewhere in the back of the line of capital structure, after bonuses and (now-growing) debts are paid. Suddenly today my GOOG stock has me thinking I’m an investing genius, but tomorrow, who knows?
S&P 500 is so concentrated, there are tens of trillions of dollars standing on a needle. Google is fighting the last war and is in precarious exponential ascend. Wealth divide is ocean wide. 2026 is going to be for the history books.