Izak Odendaal, Old Mutual Wealth Investment Strategist

Increasingly, people are talking about a bubble in artificial intelligence (AI). This sounds ominous, but it has not stopped the stock market from moving higher. There are many other commentators who are better placed to discuss the economic, social and geopolitical implications of AI, as well as the limitations and risks. Here we will simply explore the general characteristics of investment bubbles to see what applies to the current moment.

Chart 1: Google searches for “AI Bubble”

Source: Google

Each bubble is unique, but there are commonalities. There is usually a narrative of why some exciting new technological or social development will lead to ever rising prices. Economics Nobel laureate Robert Shiller described this as a process where price increases raise investor enthusiasm, which spreads by “psychological contagion from person to person, in the process amplifying stories that might justify the price increases and bringing in a larger and larger class of investors.”

There is the expectation of riches based on price movements, not underlying fundamentals. Economist Charles Kindleberger, author of a classic book on the theme, Manias, Panics and Crashes, describes a bubble as “a sharp rise in price of an asset or a range of assets in a continuous process, with the initial rise generating expectations of further rises and attracting speculators interested in profits from trading in the asset rather than its use of earning capacity.”

Inevitably, people believe there is a new paradigm that render old ways of analysing investments irrelevant. One will hear a variation of the phrase “this time is different”. This is also the title of another classic book on financial crises by Carmen Reinhart and Ken Rogoff, while Sir John Templeton called it the four most dangerous words in investing. Warren Buffet’s five most dangerous words in business, “everybody else is doing it” also applies.

A final characteristic of bubbles is that even with the above checklist – rapid price increases driven by expectation more than reality, euphoric narratives, and the drawing in of ever more participants – they are very hard to spot in real time. It is usually only after the fact that we can definitively say something was a bubble. When it is still blowing up, even hardened sceptics must admit that there is a possibility that things will work out.  

In the real world

Another general principle is that investment bubbles usually have a financial and a real economy dimension. The more pronounced the real-world dimension, the greater the fall-out after the crash.

A purely financial bubble will obviously hurt investors when it pops, but the broader impact is muted. Bitcoin is probably unique in having experienced several bubbles and crashes in its short existence, but because Bitcoin has limited connection with the real world, its price slumps did not have a big impact on the economy.

At the other end of the spectrum, however, are real estate bubbles, for instance Japan in the 1980s, the US and Spain in the early 2000s and China in the late 2010s. These bubbles distort the real economy, since an ever-greater share of activity becomes devoted to property. Other more productive sectors might lag behind as capital is redirected into the sector and people quit their jobs to become real estate agents or full-time landlords. It works for a while, but when the bubble bursts, the capital and labour attached to the real estate must be put to work elsewhere.

More damaging is the debt. Developers and homebuyers rely on loans which they must still repay even when property prices fall. Banks then end up with a sharp increase in non-performing loans. The 2008 sub-prime crisis was made worse by the fact that the debt was sliced and diced and spread throughout the global financial system, such that no one had a full picture of the massive vulnerabilities that had accumulated. When a property bubble bursts, in other words, balance sheets are impaired across the economy and painful deleveraging acts as a drag on economic growth for years.

The upside of down

Bubbles do have one positive side effect, however. They often leave behind cheap and plentiful infrastructure that would not have otherwise existed. The initial investors get wiped out, but others can buy the assets at cents on the dollar. One example is the 1800s railway bubbles. The railway was a truly transformative industry at the time, but people got overexcited. Investors bid up share prices of railway companies, who in turn laid tracks in all directions, without having to think too carefully whether they would earn a return. With all this overbuilding, many lines couldn’t be profitable and the bubbles burst.

However, afterwards, the tracks were still there, and since the new buyers paid little, they could generate a return in the decades ahead. In that sense, the railway bubble could be viewed as a social good as the first investors “subsidised” cheap railways for subsequent generations through their losses. The internet and telecoms bubble of the 1990s similarly resulted in thousands of miles of broadband fibre that could never generate a return. However, after the bubble burst and the expense of building out the fibre networks was written off, there was a platform for the internet and eventually, social media, to really take off.

It is the 1990s internet and telecom bubble that is the obvious parallel the current AI enthusiasm, but the lessons from the railway bubble of 1800s and the canal bubble of 1700s also apply. The so-called dotcom bubble was based on the belief that the internet would change everything. This belief was not wrong, just a bit early. More importantly, however, people were prepared to pay anything for exposure to internet companies. Considerations of cash flow, costs and profitability went out the window. The cash flows that were visible to investors were also exaggerated by the fact that the software companies bought from the hardware companies and vice versa. Money was circulating within the system, but the money from outside, from the end customer, was often in short supply.  

Chart 2: Nasdaq Composite Index price: earnings ratio

Source: LSEG Datastream

The internet euphoria also coincided with an era of general optimism in the years following the collapse of the Berlin Wall and leading up to the millenium. This contrasts with today, where AI stands alone as a source of excitement in an otherwise gloomy world.

There are two very simple, practical lessons from this era. The first is that companies must ultimately be profitable. They can spend a lot to grow, but revenues must eventually catch up. The mere association with the exciting new technology of the day is not a source of sustainable returns. Secondly, even solid and profitable companies can deliver poor returns when the price paid for them is too high. Microsoft, for instance, grew its earnings solidly through the 1990s and up to today. But in late 1999 it traded at an exorbitant 61 times forward earnings. After the bubble burst, it would take 16 years for the price to recover. The overall Nasdaq Index would only surpass its 2000 peak 14 years later.

Hype and hyperscalers

This brings us to AI. There are elements of the above that seem to apply, notably the excitement, the rapid price increases, the stories of revolutionary change and vast amounts of capital being poured into building datacentres and related infrastructure. There are important differences too. The overall stock market isn’t nearly as expensive today and the frothiness not nearly as widespread across sectors and geographies. The listed companies leading the AI charge are extremely profitable. This includes the so-called hyperscalers, Alphabet (Google), Amazon, Meta and Microsoft, as well as the likes of Nvidia and Oracle. Secondly, AI-related revenues are already rising rapidly off a low base, and there is a path to profitability for the privately held AI companies that are currently making a loss.

The problem is that the sheer scale of the investment that is being poured into datacentres and the associated infrastructure. The four listed hyperscalers alone plan to spend $300 billion this year and $400 billion next year. For now, this is sustaining economic growth in the US and elsewhere despite tariff headwinds and there is no sign of it slowing. Estimates towards the end of the decade are expressed in trillions, not billions.

Chart 3: US real fixed investment in information processing equipment

Source: LSEG Datastream

Whether all this spending can earn a decent return remains to be seen, however, especially when considering that specialised computer chips make up a large part of the costs of these datacentres and they become obsolete within a few years. While the railway bubbles bequeathed infrastructure that was useful many decades later, that is not the case here.  

Also important is the funding of all this capex. Up to now, the hyperscalers have largely paid for it out of their ample cashflows. But increasingly, it might be funded by debt, especially since the big lenders will also be eager to get a slice of the action. If there was a downturn, leverage would make things much worse.

Ultimately though, it comes down to the still unanswerable question of whether the technology can deliver on its obvious potential and can be applied profitably throughout the economy. Experts in the field disagree. The optimists argue that AI will change everything and will keep improving, and therefore all this spending will ultimately be worth it. The pessimists will argue that there are fundamental limitations to AI and we should temper our enthusiasm. Economic historians will point out that it took businesses many years to figure out how to use previous game-changing technologies and that the economic cycle has to date not been abolished. But perhaps this time is truly different.

Chart 4: US S&P 500 various valuation metrics

Source: LSEG Datastream

Richly valued

What we can say unambiguously is that the US equity market, led by its AI champions, is richly valued by just about any traditional metric as summarised in chart 4. This does not a predict a crash, but points to below average long-term returns at an asset class level (within which some stocks will perform better than others). Chart 5 is a simple plot of the 10-year real annual return earned from investing at different price: earnings ratios over the past 40 years. Valuation is a useful guide to long-term expected real returns and currently points to a low single-digit return over the next decade.

Chart 5: US equity valuations and subsequent returns in dollars

Source: LSEG Datastream

However, to completely bet against US equities would be risky, since momentum might carry it higher still. To go all-in and bet on the AI story would also be risky. Diversification remains the sensible approach, since the future is unpredictable and timing the market is a mug’s game. Identifying a bubble is one thing, calling the top is another, so a better approach is rather to reduce exposure as the asset class becomes more overvalued.

Three final thoughts. Firstly, and perhaps counterintuitively, is that it is good that people worried about an AI bubble. It would be more worrying if everyone was super bullish and the last sceptic capitulated. Secondly, overvaluation can lead to losses, but it is leverage that is truly destructive. Lastly, humans have developed incredible technologies over time, including now AI, but our psychological make-up is still basically as it was hundreds of thousands of years ago. We are still beholden to our emotions and biased in our thinking. The greed, jealousy and cognitive blind spots that drove previous speculative manias are still around. The same goes for the fear and loathing experienced in the downturns. We can’t avoid these feelings, but as investors we must try our best to manage them.