3 min read
In short
The free cash flow of the largest tech companies has turned negative: they’re borrowing money to build AI infrastructure. I’ll explain why this isn’t just another reason to panic, but a real shift in risk for the entire industry.
Two years ago, Amazon, Alphabet, Meta, Microsoft, and Oracle generated a combined free cash flow of over $200 billion. Now it has fallen below zero. Capital expenditures on AI have eaten up all their cash—and the companies have started borrowing.
About 30% of the hyperscalers’ capex is now covered by new debt. This share has tripled in recent years. The four largest borrowers in the tech sector have issued more than $170 billion in corporate bonds over the past year—roughly the size of the UK’s annual budget deficit. S&P has downgraded Oracle’s credit rating to one notch above junk status.
While “shovel” buyers are drowning in capex, “shovel” sellers are sleeping on mattresses of cash. The free cash flow of Nvidia, Micron, Broadcom, and AMD has exceeded $400 billion. JPMorgan analyst Jason Hunter directly compares the current trend to 1999–2000, when equipment suppliers soared while buyers sank under the weight of their own capital expenditures.
But here’s where it gets really interesting—and it’s not that “the bubble will burst tomorrow.”
Consensus forecasts analyzed by the Wall Street Journal suggest that the five major investors—Microsoft, Alphabet, Meta, Amazon, and Oracle—have pledged to double revenue over three years while simultaneously cutting ~$80 billion in operating expenses. Purdue accounting professor Kevin Koharki put it simply: “I can’t think of a scenario where that has ever happened.” Typically, growth requires more salespeople and support, not fewer.
This means that the entire financial model of the AI boom rests on one assumption: AI will automate business processes to such an extent that revenue will double while headcount is reduced. If you work with agents and LLMs, you know how optimistic this promise sounds given the current level of technological maturity.
That said, the counterarguments against an “inevitable crash” are also compelling. Big Tech’s debt-to-profit ratio is still lower than that of a typical S&P 500 company—and may remain so through the end of the decade. Forward P/E ratios for chipmakers like Nvidia, Broadcom, and Micron are in the moderate range of 10x–25x. U.S. corporate profits are at record levels, and there are currently no signs of a 1990s-style bubble—such as falling margins and a growing current account deficit.
Investor Gavin Baker draws attention to a more mundane but real problem: not financial leverage, but physical infrastructure. Power plants need to be built and data centers connected quickly enough so that demand for agents does not outpace the supply of computing power.
For practitioners, the conclusion is twofold. On the one hand, AI infrastructure is becoming cheaper—open-source models like Kimi K3 from China’s Moonshot AI are approaching the performance of top American models at a fraction of the cost. On the other hand, the economics of hyperscalers are becoming increasingly fragile: they’ve borrowed money to build out their infrastructure, and they need to prove that AI agents will generate revenue—not just fame. If you’re building products on third-party APIs, it’s worth remembering: the cost of inference today is subsidized by companies that aren’t yet confident they’ll see a return on their investment.