Double, double toil and trouble

There are very few conversations about the market these days that do not include a reflection over whether the US equity market is in a bubble. Recent market gyrations reflect the unease investors have as performance continues to be strong. This is despite mixed signals from economic activity and a level of policy uncertainty that would normally be a deterrent to long term decision making. We think it is naïve to be complacent about valuations, but it is equally problematic to panic.  

Are we in a bubble?

Alan Greenspan memorably argued that you can only recognise a bubble after it bursts. He was exactly right. Markets trade on expectations of earnings long into the future. It’s only later that we can be certain if these expectations were widely misplaced. This isn’t exactly helpful to investors.

The important question is how investors can formulate their expectations about the future growth of earnings. How can one anchor expectations around AI? This raises several interrelated issues. First, AI is a promising technology that could have seismic implications for the profitability of all industries. Second, by increasing productivity it can raise overall economic growth. Third, the scale of the benefits of AI remains unknown. Fourth, because the scale and benefits remain unknown, the amount that AI service providers are likely to be able to charge users on a sustainable basis remain unknown. Fifth, because the likely profitability of AI companies is unknown, it is unclear if the large sums being invested in AI research and data centres will ultimately earn a return.

This level of uncertainty has all the necessary ingredients to propel a bubble. Popular narratives will capture the imagination of optimists over the growth potential. Scepticism over long term earnings can yield to the instincts of participating in nearer term rallies. As more people participate out of a fear of missing out on a seismic change in the economic landscape, valuations become increasingly divorced from plausible outcomes. Are we there yet? Maybe. Maybe not. Let’s examine the issues.

AI, productivity and economic growth

There is a dramatic divergence in opinion around the impact of AI on productivity and growth. At one end of the spectrum are economists like the Nobel laureate Daron Acemoglu, who focus on the impact on AI replacing certain tasks in existing jobs. He argues that the current trajectory of AI developments mainly replaces tasks done by humans, rather than complementing them. This will reduce labour demand and will ultimately only add a paltry 1% to global GDP over the next decade. More optimistic economists such as Tyler Cowen would argue that regardless of the figures for GDP, AI is giving people access to information and skills they do not otherwise have. As an example, he imagines someone living in a rural part of the developing world gaining access to medical advice of a quality they could never otherwise imagine. Closer to home, I imagine someone in the UK getting good advice faster than the NHS can ever plausibly deliver.

At the other end of the spectrum are investment managers trying to sell investments based on a dream. A high-profile American ETF manager thinks that AI will push GDP growth to 7% per annum, and, somehow, robotaxis will generate $10 trillion dollars in revenue globally by 2030 (equivalent to a third of US GDP today).  Beware investment managers bearing AI themes.

There is tangible evidence that AI is making a difference in the real economy, even if the numbers are not yet reflected in overall productivity. Adoption rose rapidly but has slowed considerably in the past few months, as many projects underdelivered. This is unsurprising. Firms need time to figure out how best to deploy technologies. More importantly, the biggest immediate benefits of AI seem to lie in industries that rely on knowledge workers, such as finance and law. But in such areas, firms also have a fiduciary duty to their clients. Where does liability lie? A constraint on the roll-out of robotaxis was the question of where liability would lie if there were an accident.  Laws need to adapt to allow for greater use of the technology. In turn, governments need time to observe the technologies in action before changing laws. All this to say, after an initial burst of excitement, adoption is likely to happen slowly, then suddenly.

A more seismic impact of AI, however, is not likely to happen before new products and services built upon AI are created. The first reliable source of electricity was developed in 1800. The telegraph system was patented in 1837. Street lighting came in the late 1870s. Reliable electric motors were not widely available until the late 1800s. That’s almost a hundred years. While technological development today happens at a more accelerated pace, it’s not clear whether some investors in AI today are thinking that the “killer app” could take as long as 5 years to develop. 

Zero long-run profitability

Every student of microeconomics learns that, in theory, when a market is in perfect competition, economic profitability will tend to zero in the long term. In other words, profits cannot exceed a level that is just sufficient to compensate shareholders’ risk-adjusted opportunity cost (risk premium + risk-free rate).  

Perfect competition happens when companies are selling identical products that can be delivered at identical costs. This does not happen often because firms differentiate their products. But many products are awfully close.  Is it any wonder that Alphabet’s CEO Pichai told the BBC that there are “elements of irrationality” in the current AI boom?

Competition between developers of AI models is now taking place in multiple sectors. There’s considerable competition around chips. While Nvidia is a clear leader, it is now starting to face competition from its own clients – such as Alphabet – who have started testing their own chips. There’s competition in building data centres and in the use of energy to supply them. There’s competition over the data used to train the models. And there’s some serious competition emerging from China, where the government provides active support to strategically important sectors.

It is hard to foresee, at this point in the cycle, that businesses built upon providing a commodified product – such as computing power or energy – will have attractive profitability beyond the near term. Value will accrue to businesses holding the assets that cannot be commodified, such as important patents, a distinctive service, or a trusted brand behind a service.

The valuation conundrum

Benjamin Graham famously said that the combination of precise formulas with imprecise assumptions can justify any value one wishes. A new industry is ripe for assumptions that are hard to challenge, because there is no precedent upon which to anchor expectations.

Like the Railway Mania of the 1840s or the telecom/3G bubble of the 1990s, AI is a general-purpose technology that requires very large upfront investments for revenues from a service with an unknown value to consumers. Unlike the “dot com” bubble of the late 1990s, however, AI investment is being led by companies that are already generating very high cash flows. This matters because it means that some key investors are not too exposed to market conditions to finance their investments. In other words, the investment boom by the large companies can go on for some time.

This only complicates matters because the manner with which we value these companies needs to change. This isn’t simply because their future revenue streams are likely to change in composition; it’s because their current assets are also changing in nature. Over the past two decades, for example, investors have started to largely ignore a well-tested metric, the price to book ratio, because technology companies had very few tangible assets. They always looked expensive compared to any historical average. But how should we be looking at these companies now, as they accumulate ever more expensive physical infrastructure? And at what rate should that infrastructure be depreciated?  When should we be looking at the price to book again?

One reasonable analysis I have read argues that current valuations for the market are justified if we get to a point in a few years where we have, globally, 300 million paid users of AI services at $100 a month. This is not out of the realm of possibility. But to get there, companies will need time to test the reliability of the services and adjust their processes accordingly. It also means many companies would not find their needs well served by a Chinese competitor at $25 a month.

An exceptionally difficult market to invest in

Rising valuations leave less room for error: a rising market means that risks are increasing. More worryingly for the economy, as the hyperscalers become a larger part of the market, they also become a larger share of the investments of every household. Consequently, a substantial adjustment to the market could have serious wealth effects on consumers. This wouldn’t simply be bad news for AI-focused companies.

Investors are unlikely to be as well rewarded as they hope, or as badly as they fear

Our argument so far casts doubt on the starry-eyed expectations we come across. But it would be naïve to suggest that if half the promises of AI are fulfilled, earnings growth would not justify a significant share of what we can perceive as froth in the market today. It would be equally naïve to overstate the levels of overvaluation in the market today. Standard multiples such as Price to Earnings may look absurd. They look less absurd when we compare them to the rate of growth of earnings that has been delivered of late.

As the market rises, it is important to reduce risk reasonably. Investors attempting to participate in the AI rally through “lower valuation” equities, such as those of electric utilities that supply AI data centres, are likely to have a harder time recovering from a market correction than investors in companies that own significant intellectual property. This may sound counterintuitive because utilities are traditionally considered defensive, rather than cyclical.

Like past investment booms over a new technology, from the Railway Mania to the telecom bubble, many investors could lose a lot of money. But these booms also ended up leaving behind infrastructure that proved invaluable to consumers and producers alike. The consequence was a long-run improvement in productivity and growth. In the long run, Keynes famously said, we are all dead. He also wrote in Economic Possibilities for our Grandchildren that technological progress means that by 2030, people in rich countries may only need to work 15 hours a week to fulfil their basic needs. Could he be right on both counts?