Whether end demand can accelerate fast enough to generate a reasonable return on investment right across the value chain is ultimately the million/billion/trillion-dollar question.
Artificial intelligence (AI) bubble talk is back. Megacap tech stocks have had another strong year, but with focus intensifying on the circular nature of recent high-profile deals and unprecedented concentration in US equity benchmarks, investors are once again concerned around the potential for a major correction in AI stocks.
The AI ecosystem
To address these concerns, first we must get a grip of the increasingly sprawling AI ecosystem (see Exhibit 7).
The hardware manufacturers sit at the bottom of the tech stack. This group includes companies such as Nvidia (which designs highly sophisticated chips), TSMC (which manufactures chips for Nvidia, among others) and ASML (which creates precision tools that are used in chip manufacturing).
Next come the ‘hyperscalers’. This group includes industry titans such as Alphabet, Amazon, Meta, Microsoft and Oracle . These companies build and operate data centres that are the backbone of AI and cloud computing more broadly. With an estimated half of AI-related capex typically being spent on chips, the hyperscalers are critical customers for the hardware manufacturers.
We’ll call our third group the AI architects. These companies develop large language models (LLMs) and other AI applications that can be used for everything from video creation, to chatbots, to complex medical research. OpenAI’s ChatGPT is arguably the most famous LLM to date, but there are many other popular examples, such as Anthropic’s Claude, Google’s Gemini, xAI’s Grok, and Mistral’s Le Chat.
While these buckets neatly encompass the supply side of AI infrastructure, it is far harder to apply a blanket definition to the demand side. Users range from individuals (e.g., using ChatGPT), to enterprises (e.g., a call centre using AI chatbots), to companies that place AI at the heart of their service offering, such as software companies who sell AI-driven tools.
In a well-functioning, mature ecosystem, the money spent by end users must be sufficient to generate profits throughout the system. Take the example of a weekly supermarket shop. The money you spend at the till should be enough for the supermarket, the real estate operator, the food manufacturers and the farmers to all take their cut.
In today’s AI ecosystem, the demand from end users, both individuals and enterprises, is currently insufficient for all the other players to turn a profit. For several years, the hyperscalers have been largely footing the bill, using revenues from other parts of their businesses to fund the build out of AI capacity, which in turn has created profits for the hardware players. Whether end demand can accelerate fast enough to generate a reasonable return on investment right across the value chain is the million/billion/trillion-dollar question.
Bubble watch
Given the incredible run-up in tech share prices since the pandemic, it is no surprise that comparisons are being made to prior investment bubbles.
Valuations: High, but not prohibitively so if earnings deliver
While today’s valuations are not as extreme as the dotcom bubble, the S&P 500’s 12-month forward P/E ratio of 23x is not far off the peak valuation in 2000, while the tech sector trades on 31x forward earnings (see Exhibit 8).
Hardware names currently sit on higher valuations (35x 12-month forward earnings) than the hyperscalers (30x). For the AI architects, however, valuations are harder to compare given that major players such as Anthropic, Mistral and OpenAI are still private today. This lack of visibility makes accurately assessing the fundamentals of a group integral to the entire ecosystem much more challenging.
Two key lessons can be drawn from prior run-ups in tech valuations. The first is that high valuations do not automatically prohibit future returns. An investor who purchased Microsoft at a peak valuation of 68x in December 1999 would be up over 1,000% over the following 25 years (provided that they held the stock through a 65% drawdown in 2000). The second lesson, however, is that stock markets generally do a very poor job of predicting future winners during technology revolutions. Looking at the most popular tech companies in the S&P 500 at the dotcom peak in 2000, only Microsoft retained its place in the largest 10 companies a decade later (see Exhibit 9). Identifying long-term winners in real time is an incredibly complex task.
Financial health: Much improved versus 2000, but with cracks beginning to emerge
The financial strength of the AI megacaps is arguably the biggest reason to downplay comparisons with the dotcom bubble. Unlike the flimsy nature of corporate balance sheets 25 years ago, Amazon, Apple, Alphabet, Meta, Microsoft and Nvidia are today sat on a cash pile to the tune of $450 billion. Given that tightening credit conditions have often been a major trigger for prior bubbles to burst, this suggests that today’s tech titans are in much better shape.
Critically, this financial strength has so far allowed many of the hyperscalers to fund their AI-related capex from free cash flow. We note, however, that capex commitments planned for the coming years will eat further into available cash piles (see Exhibit 10). Recent high profile moves from Meta, Oracle and xAI to tap both public and private debt markets bear watching in this regard.
The increasingly circular nature of AI deals over recent months does, however, have some echoes of the late 90s. September’s announcement that Nvidia plans to invest up to $100 billion in OpenAI was arguably the highest profile of these investment deals. OpenAI in turn will funnel much of this investment into securing new compute capacity, which will then drive demand for Nvidia’s chips.
Optimists would argue that Nvidia is using the cash on its balance sheet to simply bring forward future demand. Yet the more that the fate of individual companies becomes intertwined, the greater the risk that a single failure could lead the broader system to unravel.
Market structure: A greater skew towards private markets requires a different approach
Initial public offering (IPO) activity is another major point of difference today. Exuberant IPO activity was a key feature of the dotcom bubble: in 1999, the median first-day return for the 476 US companies that went public was a whopping 57%, compared to just 7% on average between 1980-2024. These bumper returns enticed companies to rush into the market far quicker than normal (see Exhibit 11).
While IPO activity has picked up in the second half of 2025, the situation remains far more muted. In part, this reflects a growing preference from companies to stay private, given the combination of a lower regulatory burden and ample availability of private capital, as we discuss in Don’t be fearful of private markets.
The good news from companies waiting much longer to IPO is that this has strengthened the overall fundamental picture of the listed tech sector. The bad news is that for investors accessing the AI theme only via public markets, they may well miss out on exposure to some of the long-term AI winners that are only accessible via private equity strategies currently.
Earnings: Supernormal profit delivery has become the norm, but could be tough to maintain
Unlike in the late 1990s, when valuations surged while earnings lagged, today’s run-up in share prices has been driven far more by expanding earnings. This is not only a revenue story, but also a margin one: US tech sector margins are currently more than double that of the broad S&P 500. Investors now assume that tech companies will deliver supernormal profits compared to the rest of the index.
Admittedly, the profitability of the AI names in the private markets is generally far weaker, but their revenue growth has still been impressive. OpenAI is reported to now be generating annual revenues north of $13 billion, although given the company’s huge R&D spend, it is not expected to reach profitability until much closer to 2030.
Whether this earnings growth can continue depends on both end demand, as well as whether the tech firms can maintain their supernormal margins. Investment booms do have an unfortunate habit of leading to overcapacity, which in turn damages margins. The shale boom of the 2010s is a prime example: breakthroughs in drilling techniques led to a surge in oil supply, before a sharp collapse in oil prices led to a wave of bankruptcies (see Exhibit 12).
The biggest challenge for the hyperscalers in accurately estimating future demand is the many variables involved. By 2030, will the resource intensity of current AI have reduced materially? Will we have made major breakthroughs in artificial general intelligence (AGI)? And how quickly will AI hardware be made obsolete by new technological developments? None of these questions have concrete answers today, leading to wide error bars around estimates for both future capex needs and earnings overall.
End demand: The most important question, and the most difficult
If AI ultimately allows companies to create major new revenue streams and significantly cut back on their labour costs, the tech giants will continue to generate impressive earnings growth. Yet if consumers and corporates prove more resistant to spending big sums on new AI tech, achieving a return on today’s surging investment will be far more challenging.
The AI architects have been keen to stress the astronomical growth in their user base, although paid users have been much harder to come by. In OpenAI’s case, ChatGPT has eclipsed all records for the pace of technology adoption, yet only 5% of the 800m users of ChatGPT are currently paying customers, according to the Financial Times.
Survey data presents a similarly mixed picture. The US Census Bureau’s AI survey suggests that only 9% of US businesses are using AI to produce goods and services, with large variation across sectors. Somewhat concerningly, when breaking out this survey by business size, recent responses suggest that demand from the largest companies may be starting to slow (see Exhibit 13).
Other surveys point to a much more positive outlook, such as the Ramp AI Index that shows 44% of surveyed businesses pay for an AI subscription (see Exhibit 14), or KPMG’s Quarterly Pulse Survey that evidences a steady increase in the deployment of AI agents designed to boost employee productivity.
Triggers for a turn in sentiment
We now focus our attention on identifying potential triggers that could lead market positivity around the AI theme to reverse.
A high-profile misstep
Given the increasing number of interlinkages within the AI ecosystem, one high-profile ‘flop’ could be very damaging for the broader universe. Earnings reports are the most obvious place to look. Nvidia, for example, beat both earnings and revenue expectations in every quarter from Q1 2023 to Q2 2025. The size of the beats has been diminishing as expectations ramp up, and we should also consider that the pace of capex growth from the hyperscalers is set to slow (see Exhibits 15 and 16). A big miss from any of the AI megacaps would likely create significant concern about firms across the industry. Another example would be if OpenAI pressed ahead with IPO plans in 2026 and subsequently faced difficulties in achieving its desired valuation.
Capacity constraints
Whether the provision of energy and raw materials can keep up with the demand for future compute capacity is a question that has so far received far less market attention, but is another crucial one. News reports of electricity blackouts linked to data centre expansion, or a shortage of critical minerals that are essential to chip manufacturing, are both examples of issues that could potentially shift sentiment.
A liquidity event
We must also consider that an external shock, independent of technology, could be the catalyst for a correction. With valuations elevated, technology is increasingly trading as a ‘high beta’ theme, outperforming in rising markets and underperforming when markets fall. If our tail risk scenario of an inflation shock that triggers higher bond yields plays out, we would expect AI-related stocks to be hit particularly hard.
How to position
With AI sentiment such a dominant driver of markets today, clearly investors do not have the luxury of waiting to see how demand develops before taking a view. Typically, when faced with unpredictable outcomes, diversification is a shrewd approach. For AI unpredictability, we see three specific ways to boost diversification.
1. Diversify across the ecosystem: The risks and opportunities facing AI-related companies can vary substantially. If, for example, hyperscalers overestimate the future demand for compute, this overinvestment could lead to downward pressure on prices, hurting hyperscaler margins while benefitting the consumers of this compute capacity. We may also find that there are winners and losers within each AI bucket. In the hardware bucket, for example, time will tell whether the AI ecosystem can sustain multiple chip manufacturers, or whether Nvidia, Broadcom and AMD (among others) are ultimately facing a ‘winner takes all’ scenario.
2. Diversify regionally: The particular exposure of US indices to the AI theme makes a strong case for regional diversification, as we lay out in Diversify selectively across global equities. First, we may find that the ultimate winners of the AI race are situated in other parts of the world, most notably Asia. Second, if the market’s attention shifts from AI producers to AI users, Europe’s sector mix could benefit. Alternatively, if AI sentiment turns altogether, the low exposure to technology of markets like the UK and Switzerland suddenly looks more appealing.
3. Diversify across public and private markets: History provides many examples of those responsible for the build out of a new technology later being ousted by more dynamic, younger start-ups. If value creation is ultimately more focused in the developers of the tools that can harness AI, rather than the infrastructure providers, the long-term winners could well still be lurking in the private markets today. This is one of the reasons we are structurally positive on private equity – see Don’t be fearful of private markets.
Conclusion
The earnings of the tech giants have been incredibly impressive, but the outlook for future AI demand remains highly uncertain and, in turn, it is highly uncertain whether high expectations will be fulfilled. Moreover, navigating the timing of when we could encounter such disappointments is even harder. For these reasons we would warn against positioning portfolios strongly in either direction. Diversification across the AI ecosystem, across regions, and across public and private markets should provide the best risk/reward, helping investors manage whichever twists and turns this latest technology revolution throws at us.