Nothing pops a bubble faster than calling it one. At least, that is the argument from Deutsche Bank, which says the chatter about artificial intelligence froth has itself deflated: a “bubble in saying there’s a bubble”.
Searches for the term “AI bubble” hit a peak in August after an MIT report suggested companies were struggling to make money from their AI investments, and Sam Altman warned that investors might be “over-excited”.
Within days the Magnificent Seven tech stocks slipped nearly 4%. But interest has since collapsed: Google searches for “AI bubble” are now at just 15% of their August peak.
The report sets out why this matters. Market panics often burn hottest when sentiment turns, but the dotcom bubble offers a lesson in just how long these things can run.
The Nasdaq endured seven swings of more than 10% before finally peaking in 2000. Amazon and Yahoo were being written up as examples of excess in 1998, when the index was under 2,000.
It did not burst until it had passed 5,000 eighteen months later. In short, spotting a bubble is easier than timing its end.
For now, Deutsche highlights four forces shaping today’s AI cycle. First is a dose of realism. The much-trailed launch of GPT-5 underwhelmed in August, offering a smoother user experience rather than the leap towards general intelligence some had hinted at. Advances that would have dazzled a year ago were instead met with shrugs.
Second are infrastructure bottlenecks. Consumer adoption has been explosive — ChatGPT is on course for a billion weekly users by year-end, but enterprise adoption will only pay off once companies build the vast scaffolding of chips, data centres and energy systems required.
Third comes the hard graft of integration. History shows new technology only delivers returns when embedded into workplace systems that people actually use. That work is just beginning. Fourth is psychology. Every innovation passes through the familiar “hype cycle”: inflated expectations, disillusionment, adjustment, then real productivity. Rolls of the eyes are part of the process.
Sceptics have been quick to draw parallels with past manias. The cyclically adjusted price-earnings ratio for the S&P 500 is nudging 38, close to the dotcom peak of 44. Hedge fund managers from Ray Dalio to David Einhorn have warned of “extreme” spending on AI infrastructure, predicting that a great deal of value could yet be destroyed.
Yet others note that long-term investors who sat out the wobbles of the 1990s were handsomely rewarded. A $10,000 investment in US equities at the start of 1996 would have grown to $170,000 by mid-2025; miss the 20 best days in that period and the result would be a quarter of that. Notably, many of the best days followed immediately after the worst.
The wider lesson is that bubbles are rarely neat. Britain’s railway mania of the 1830s collapsed, then re-inflated into the bigger mania of the 1840s. The 1960s “tronics” boom fizzled, only for a second, harsher crash to follow. Radio stocks soared in the 1920s before enough sets were in homes to justify the enthusiasm. Some bubbles last months, others years.
For AI, the analysts conclude, the jury is out. A pullback in fear may even act as a safety valve, taking heat out of valuations and drawing in bargain hunters.
Investors should be wary of declaring the story over. As John Maynard Keynes is quoted in the report: markets can stay irrational longer than you can stay solvent.