Nvidia CEO Jensen Huang has become the poster child for the AI boom as his company’s value has surged roughly 300% in two years. On a recent earnings call he sought to calm investors, saying talk of an AI bubble is misplaced. But many of the loudest voices insisting that AI investment will keep accelerating are also those with the most to gain from continued spending — and a growing chorus of analysts and economists sees warning signs beneath the hype.
Prominent proponents call this a super-cycle, not a bubble. White House AI adviser and venture capitalist David Sacks said he sees a boom, not the start of a bust. Andreessen Horowitz cofounder Ben Horowitz and JPMorgan executive Mary Callahan Erdoes have dismissed bubble talk, arguing that demand and growth metrics justify high valuations and investment.
Skeptics, however, point to an enormous and risky flow of capital into largely speculative bets. Paul Kedrosky, a venture capitalist and MIT research fellow, says a startling amount of money is chasing a “revolution” whose pace of improvement has slowed. Nobel laureate economist Daron Acemoglu warns much of the industry’s rhetoric is exaggerated: useful technologies will emerge over the next decade, he says, but current expectations are overblown.
The scale of spending is striking. OpenAI’s CEO Sam Altman has claimed the company makes about $20 billion a year and plans to spend as much as $1.4 trillion on data centers over eight years — growth that depends on ever-broader customer adoption. Yet research from sources including McKinsey and other analyses suggests most firms aren’t seeing chatbots materially change their bottom lines, and only a small share of consumers currently pay for AI services.
Big Tech’s investment plans are vast. Amazon, Google, Meta and Microsoft are projected to spend roughly $400 billion on AI this year, mostly on data centers. Some companies may allocate as much as half their cash flow to building that infrastructure. Morgan Stanley estimates tech firms could spend about $3 trillion on AI infrastructure through 2028, with only half covered by internal cash flows. If demand slows or capacity outstrips need, analysts warn, much of that spending could go to waste.
To fund the rush, firms are increasingly using debt and complex financing structures that obscure liabilities. Goldman Sachs analysts say hyperscalers have taken on about $121 billion in debt over the past year — a more than 300% jump from typical levels. Companies are also using special purpose vehicles (SPVs) and joint ventures to finance data centers so the debt doesn’t appear on their balance sheets. An example: a joint entity funded by Blue Owl Capital and Meta for a Louisiana data center involved a $27 billion loan backed by Meta’s lease payments; Meta owns a minority stake but gets full computing capacity while the loan doesn’t show up as Meta debt. If the facility becomes uneconomic, Meta could still be obligated for large payments to the financier.
Critics say such off-balance-sheet financing echoes risky past practices. “Special purpose vehicle” gained infamy with Enron two decades ago; while today’s deals are more transparent, some analysts warn the strategy shouldn’t form the backbone of long-term corporate strategy.
Circular and subsidized deals amplify concerns. Nvidia and OpenAI announced a multibillion-dollar strategic arrangement in which Nvidia provides funding for data centers and OpenAI fills them with Nvidia chips. That structure — where a vendor supplies capital that helps a customer buy the vendor’s products — can artificially inflate demand metrics. Paul Kedrosky notes it’s common at small scale but unusual at the tens-to-hundreds-of-billions level; he sees echoes of dot-com-era practices.
Smaller players have also pivoted to capture AI demand. CoreWeave, once a crypto-mining startup, built out data centers to serve AI firms. OpenAI struck deals with CoreWeave worth tens of billions, exchanging usage rights or stock for compute capacity; Nvidia also owns part of CoreWeave and has guaranteed to buy unused capacity through 2032. Such interlinked arrangements — stock, capacity guarantees, cross-ownership — worry some economists who fear they could mask true end-user demand.
Market participants have begun to act on these concerns. Investors including Peter Thiel and SoftBank have reduced or sold large Nvidia positions. Michael Burry, famed for betting against subprime housing in 2008, has publicly bet against Nvidia and accused parts of the AI industry of relying on circular financing and “fancy accounting.” He argues true end demand is small and many customers are effectively funded by their dealers.
Even executives in the field acknowledge excesses. Sam Altman said investors might be “overexcited” about AI even as he called it one of the most important developments in a long time. Google CEO Sundar Pichai recently admitted elements of irrationality in the market and warned no company would be immune if a downturn arrived.
Analysts point to historic parallels. The dot-com bubble featured massive debt-financed buildouts — notably of fiber-optic capacity — for demand that lagged expectations, producing significant losses. If AI infrastructure is overbuilt and demand underwhelms, some warn the result could be worthless assets, loan losses for financial institutions, and broader market pain.
The central risk is mismatched assumptions: companies and investors are placing huge bets that AI-driven revenue growth will materialize quickly and at scale. If the technology’s performance gains slow, enterprise adoption stalls, or monetization proves harder than promised, the cascade of debt, off-balance financing, and interdependent deals could reveal fragility beneath today’s boom.