Is AI a bubble?

Something curious has happened in Silicon Valley. The debate about whether artificial intelligence represents a speculative bubble has effectively ended. Not because investors have concluded it isn’t a bubble, but because they’ve decided bubbles don’t matter. This intellectual pivot, from denial to acceptance to enthusiastic embrace, represents either sophisticated long-term thinking or a spectacular failure of collective rationality. The stakes are high enough that getting the answer wrong could reshape the global financial system.

Consider the remarkable candour now commonplace amongst those with the most to lose. “Of course there’s a bubble,” says Hemant Taneja, chief executive of General Catalyst, a venture capital firm that raised an eight billion dollar fund and backed AI companies including Anthropic and Mistral [1]. Sam Altman, whose OpenAI sits at the epicentre of the frenzy, offers investors words they rarely hear from chief executives: “I do think some investors are likely to lose a lot of money” [2]. Even Jeff Bezos, who built Amazon amidst the dotcom carnage, now draws careful distinctions between financial bubbles (bad) and industrial bubbles (possibly good) [3].

This shift from “there is no AI bubble” to “AI is a bubble and bubbles are great” suggests we’ve moved through Elisabeth Kübler-Ross’s stages of grief at remarkable speed [4]. What it doesn’t suggest is any lessening of the mania itself.

The mathematics of madness

The scale of capital deployment defies easy comprehension. US venture capitalists have committed 161 billion dollars to AI companies over the year to date, two-thirds of their total spending [5]. Ten loss-making AI start-ups have gained close to one trillion dollars in valuation over the past twelve months alone [6]. Google, Amazon, Microsoft and Meta will spend 750 billion dollars on data centres this year and next, with Morgan Stanley projecting total global spending to reach three trillion dollars by 2029 [7]. That final figure equals roughly fifteen per cent of the European Union’s entire GDP [8].

Historical context makes these numbers even more startling. When venture capitalists funded the original dotcom boom, they invested 10.5 billion dollars into internet companies in 2000, approximately twenty billion dollars in today’s money. During 2021’s software frenzy, they deployed 135 billion dollars into software-as-a-service start-ups. This year, they’re on course to exceed 200 billion dollars on AI alone [9].

Yet valuations have become untethered even by the generous standards of venture capital. Start-ups generating merely five million dollars in annual recurring revenue now seek valuations exceeding 500 million dollars, according to a senior Silicon Valley venture capitalist. “Even during peak Zirp (zero-interest rate policies), these would have been 250 million to 300 million dollar valuations,” he notes [10]. The market, as he puts it more bluntly, “is investing as if all these companies are outliers. That’s generally not the way it works out” [11].

Goldman Sachs chief global equity strategist Peter Oppenheimer maintains that we’re witnessing “not a bubble… yet” [12]. His analysis comparing the Magnificent Seven tech companies to definitive historical bubbles finds them relatively reasonable on forward price-to-earnings ratios compared to 2000, 1989 or 1973 [13]. Yet other metrics flash warnings. The technology, media and telecoms sector trades richer on a price-to-book basis than at the 2000 peak. Palantir, the data and AI company, commands a forward price-to-earnings multiple of 225, the highest valuation of any S&P 500 company [14].

By October, a Bank of America survey found that an AI bubble had become perceived as the number one downside risk to global growth, overtaking even concerns about Trump administration policies that had dominated for most of the year [15]. More than half of fund managers in an earlier survey believed AI stocks were already in bubble territory [16].

Productive mania

The intellectual defence of AI excess rests on historical precedent and economic theory. Eric Schmidt, Google’s former boss, makes the case enthusiastically: “Bubbles are great. May the bubbles continue” [17]. Their function, he argues, is to redirect vast pools of capital into frontier technology and infrastructure, which ultimately benefits society regardless of what happens to investors.

Schmidt poses a thought experiment that clarifies the logic. What if a technology company achieved artificial general intelligence, then superintelligence? Such technology would exceed the sum of human knowledge and solve humanity’s hardest problems. “What’s the value of that company?” he asks. “It’s a very, very large number. Much larger than any other company in history, forever, probably” [18].

This isn’t merely Silicon Valley self-justification. Nobel laureate William Nordhaus estimated that between 1948 and 2001, innovating companies captured only 3.7 per cent of the value their innovations created, with 96.3 per cent flowing to society at large, mostly through consumer benefits. Put differently, spillover benefits were 26 times larger than private profits [19]. If AI follows this pattern, investments could generate enormous social value whilst destroying investor capital, a painful irony that nevertheless represents economic progress.

The canonical example remains Britain’s railway mania of the 1840s. Investors lost fortunes. Share prices collapsed. Yet Britain ended up with railways that transformed the economy for generations. As Victorian historian John Francis wrote: “It is not the promoters, but the opponents of railways, who are the madmen” [20].

Marc Benioff, Salesforce’s co-founder and chief executive, estimates that perhaps one trillion dollars of AI investment might be wasted, but that the technology will ultimately yield ten times that in new value. “The only way we know how to build great technology,” he argues, “is to throw as much against the wall as possible, see what sticks, and then focus on the winners” [21].

Inconvenient complications

This sanguine narrative glosses over substantial problems. William Quinn, co-author of Boom and Bust: A Global History of Financial Bubbles, notes that funding railways through a bubble rather than central planning “left Britain with a very inefficiently designed rail network. That’s caused problems right up to the present day” [22].

George Hudson, the “railway king” who controlled four of Britain’s largest railway companies whilst simultaneously serving as mayor of York and an MP, kept his empire afloat through distinctly Ponzi-like operations. He funded dividends for existing shareholders from freshly raised capital and defrauded investors by having companies he controlled buy his personal shares at above-market prices. Only parliamentary immunity from arrest for unpaid debts kept him from ruin before his eventual exile to France [23].

Historian William J. Bernstein notes that “the closest modern equivalent would be the chairman of Goldman Sachs simultaneously serving in the US Senate” [24]. One need not work hard to imagine contemporary parallels.

Profitability vacuum

The gap between investment and return is already uncomfortably wide. Research from the Massachusetts Institute of Technology found that 95 per cent of companies surveyed were getting zero return from their investments in generative AI [25]. OpenAI, three years after launching ChatGPT, has reached thirteen billion dollars in annualised revenue, unprecedented growth for a start-up. Yet the company is on course to burn 8.5 billion dollars in cash this year [26].

Is OpenAI worth 500 billion dollars? The question seems absurd until one considers the alternative. If the company achieves AGI, perhaps any finite valuation is too low. If it doesn’t, the current valuation is fantastical.

This binary logic pervades the entire sector. OpenAI and competitors are racing against Meta, Google and others in a capital-intensive contest to train ever-better models, meaning the path to profitability extends further than for previous start-up generations [27]. The deals with chipmakers and cloud providers represent bets that AI demand will continue its stratospheric growth, enabled by research breakthroughs and new products that remain hypothetical.

About one-third of AI-related capital expenditure is sinking into short-lived assets like Nvidia’s graphics processing units, which have a useful life for frontier applications of roughly three years [28]. Tech analysts Azeem Azhar and Nathan Warren note that GPUs “age in dog years” [29]. Unlike railways or power grids that serve for generations, this infrastructure may become obsolete before generating returns, though the depreciation pressure might impose discipline that was absent in earlier bubbles.

Geopolitical wildcard

China’s rapid advancement at dramatically lower costs introduces another dimension of uncertainty. Beijing-based Moonshot AI unveiled its Kimi K2 Thinking model for less than five million dollars in training costs [30]. When Chinese firm DeepSeek launched a low-cost ChatGPT competitor earlier this year, Nvidia lost nearly 600 billion dollars in market value in a single day [31].

Jensen Huang, Nvidia’s chief executive, recently warned that China will “win the AI war” as it becomes a tech superpower [32]. This wasn’t defeatism but frustration; Trump’s administration won’t allow Nvidia to sell advanced chips in China, potentially ceding ground in the world’s second-largest economy. If Chinese companies can match or exceed Western AI capabilities at a fraction of the cost, what does that imply for the trillion-dollar bets being placed in Silicon Valley?

Crashes

Carlota Perez, author of Technological Revolutions and Financial Capital, sees AI as an extension of the information technology revolution beginning in the 1970s, the fifth great technological revolution she identifies. Her framework describes a predictable pattern: an installation phase featuring creative destruction, social disruption, over-investment, financial mania and bubbles. Those bubbles fund vital infrastructure enabling subsequent mass rollout and broader economic benefits through what she terms a “golden age” [33].

“I have not seen a golden age happening without a crash,” Perez states [34]. She warns that capital markets are currently misfiring, focusing more on speculative games like crypto than productive investments, with global debt exceeding three times GDP. “This could also be a trigger for gigantic instability,” she adds [35].

Recent market action offers a preview. Tech stocks experienced their worst week in seven months during early November, with Nvidia losing 500 billion dollars in market value over five days [36]. The Nasdaq fell more than 4.5 per cent over the week, with firms investing heavily in AI losing nearly one trillion euros in market value amid investor anxiety that billions deployed may not generate hoped-for returns [37].

Patrick Honohan, former governor of the Central Bank of Ireland, warns starkly that the AI bubble now ranks amongst the biggest threats to the global financial system. Whilst equity fluctuations might seem inconsequential, he notes, “they have an effect on what’s happening — what’s being bought using these valuations — and raising capital” [38]. His nightmare scenario envisions investors gradually realising that AGI is “not five or ten years away, it’s 50 or 100 years away,” triggering substantial falls that ripple through interconnected markets [39].

Asked whether we’re in an AI bubble, scientist and entrepreneur Stephen Wolfram says: “The answer is obviously yes.” As for talk about AGI? “It’s a meaningless thing” [40]. Andrew Odlyzko is similarly unimpressed by analogies between railways and AI. People at least understood how railways worked and what they were supposed to do. This is not the case with generative AI. “We are losing contact with reality,” he says [41].

Strategic response

For executives and investors, the question is not whether to engage with AI — the technology’s potential is too significant to ignore — but how to do so with appropriate scepticism. Simon Edelsten, fund manager at Goshawk Asset Management, counsels against trying to time the market perfectly. “Watch individual stocks and sell if they are too aggressively valued,” he advises [42]. His firm has reduced exposure to Magnificent Seven stocks from twenty per cent to roughly nine per cent by selling Tesla, Nvidia and Meta whilst retaining positions in Microsoft and Amazon where valuations “look merely stretched, not ridiculous” [43].

William Quinn offers comfort in noting that when banks stay away from bubbles, their bursting has limited effects, true in the 1840s and potentially true today [44]. Unlike the dotcom era, leading AI companies have genuine cash flow and profits. The critical question is whether they’re deploying that capital wisely, which recent earnings reactions suggest investors are beginning to scrutinise more carefully.

The lesson from 2000, Edelsten argues, was that “if the valuations of the shares you are buying seem reasonable then you need not worry too much about the stretched valuations of the shares others own” [45]. Long-neglected sectors like healthcare and consumer staples offer alternative homes for capital. His firm recently bought Nestlé — about as far from AI excitement as one can get — whose nearly four per cent yield in Swiss francs “looks pretty sweet” [46].

A bit of both

Perhaps both camps are correct. AI likely represents genuinely transformative technology whilst simultaneously existing in a bubble characterised by irrational exuberance and catastrophic capital misallocation. Bret Taylor, OpenAI’s chair, articulates this duality clearly: AI will “transform the economy”, he says, and “create huge amounts of economic value in the future. But I think we’re also in a bubble, and a lot of people will lose a lot of money” [47].

The challenge for business leaders is distinguishing between revolutionary technology and revolutionary valuations. History suggests bubbles often fund important infrastructure whilst leaving wreckage in their wake. The winners will be those who can maintain that distinction even as markets lose contact with reality.