Hardly a day goes by without some new headline about the growing AI bubble. Goldman Sachs says that Microsoft, Amazon, Meta and Google alone will spend $350 billion on AI-related infrastructure this year.

Critics argue that these huge investments won't bring the returns firms need. For every dollar AI earns, firms currently spend six on capital. At the height of the dot com bubble 25 years ago, firms only spent four dollars on capital for every dollar of revenue. 

Supporters say that AI firms are already making money. In its first fiscal year the now defunct dot com darling pets.com made just $700,000, while OpenAI’s revenues are thought to be $13 billion. 

This debate might seem far from most people's lives. But more Americans own shares than before. And tech giants now dominate the stock market. According to the Economist a bust today of similar scale to that of the dot com era would wipe out 8% of American household wealth. 

Technology bubbles are not new. Before dot com there was a US electricity and radio bubble in the 1920s, a US railway bubble in the 1860s and two railway bubbles in Britain in the 19th century - to name just a few. 

Economist Carlota Perez argues that major technologies follow 50-60 year cycles. The pattern is one of growth, boom, bust then maturity. While there may be hysteria and pain along the way, in the long run we see big gains from technologies.

One of the benefits of some technological bubbles is that the frenzy phase involves roll out of infrastructure that will prove useful later. Railway mania left behind tracks and bridges that could be used by others even if the original investors went to the wall.

AI's infrastructure is mostly data centres, with chips making up 60% of the cost. But chips last just a few years due to obsolescence and wear and tear. Far less than bridges or optic fibres. This could spell trouble if we really are in the midst of an AI bubble. Will we be left with less to show for the billions invested than past technological cycles? 

Perhaps not. We just need to look more carefully. While the chips may have short shelf lives, other less obvious parts of the infrastructure build out could offer benefits after the frenzy. Eight opportunities are: 

Power

Data centres need power. Lots of it. In 2024 data centres used around 1.5% of global supply. This is set to more than double by 2030, mostly because of AI. Demand is so high that tech firms are even building their own power sources. Power (like AI) can be used for many things, including cutting carbon from the economy. The legacy of an AI bubble might not be AI or even data centres, but more power right when we need it to fight climate change. Of course, this will only work if the power is renewable and can be linked to and run with the rest of the grid.

Datasets

AI also needs huge amounts of data. The tech giants have hovered up much of the internet to train their machines. Other private data has also been marshalled to feed AI. Could these datasets be reused in the future for other data analytics even if the original plan involving AI fails? When Monarch Airlines went bust, its landing slots were worth more than its planes. Likewise, the data used to train AI might be more prized than the physical hardware. Governments might want to think about making such data easy to move. Public data trusts or commons might be created for some shared, anonymised, or open datasets to let them be used more easily later. Of course, in many cases there are questions of whether this data can be used at all since some tech companies did not seek permission from the data holders. 

Compute

Data centres predate AI and chips have uses beyond models. Tech giant Nvidia’s famous graphic processing units (GPUs) that have been so central to machine learning were originally created to boost the visuals of computer games, although the way they are used in data centres for AI is different. An AI bubble may leave us with too many data centres in the short run, whose chips might not be suitable for cutting-edge AI. But they might have other uses. In a world of “infinite compute,” processing power could be well used in fields like weather forecasts, science, or economic models. Maybe cash-strapped governments or non-profits could buy up compute for social good at fire-sale prices? To help this process, governments could proactively define a pathway to acquire decommissioned AI clusters for non-commercial use.

Open source models and skilled people

While many AI models such as Open AI and Claude are proprietary, others such as France’s Mistral AI or Llama from tech company Meta are open source. Even if the financial rug is pulled from under the private models by a bursting AI bubble, then the open source models will still in principle be usable by all. Although they are not always at the bleeding edge of technology, these open-source offerings might nevertheless prove useful. A similar case might be made for all the expertise that AI engineers have built up during the build out. Jobs may be lost in the short run if the bubble bursts, but in the long run these skills might be applied elsewhere. 

Robots

Advances in AI have been stimulating progress in robotics. For example, the UK Advanced Research and Invention Agency (ARIA) has recently announced a £57 million programme to help robot bodies keep pace with advances in robot brains. Even after an AI bust, some of the progress made in robotics will be useful, either when AI gets better or without AI at all - robots have been used in factories for decades without modern AI.

Heating and cooling

Data centres generate an awful lot of heat. And in some places this is being put to good use. Google is already planning to use its data centres in the historic port city of Hamina, Finland to provide district heating to homes and businesses. In the event of an AI bubble bursting and data centres being used for purposes beyond AI, this heating capability would likely remain. At the other end of the spectrum, the excess heat generated in data centres is stimulating advances in cooling technologies which could prove useful in a world made hotter by climate change. 

Fixing the technological plumbing

Perhaps the most overlooked opportunity offered by the hype around AI is the chance to smuggle in much needed digital transformation into organisations. We have been able to transcribe notes using digital technologies for years, but it has taken the inspiration of AI to make this happen in practice at scale. To implement shiny new AI, governments and companies are having to fix underlying technological infrastructure problems. For example, legacy systems storing data in a way that makes it hard to discover and even harder to turn into insight or efficiencies for the organisations that own it. Maybe the legacy of the AI bubble will be pushing us to get our technological house in order? 

Changing underlying systems rather than tacking on solutions

The hardest part of technological transformations are often the human elements: culture, habits, systems and institutions. Too often the full benefits of technology are not realised because it’s bolted on afterwards rather than fully integrated. When electricity was first used in factories, it was sometimes assumed to work like steam power with a central motor powering all the machines through belts and pulleys. Only later did industry realise that electric power could be distributed to small electric motors through plugs and wiring. The whole system of power distribution needed rethinking. One inadvertent benefit of an AI bubble bursting is that it might give people and organisations breathing space to work out how to change whole systems to best use the technology.

AI is not like tulips. In the 17th century, the Netherlands was gripped by tulip mania, where the value of the flower exploded. Eventually, the bubble burst, and many people lost money as the tulips had limited real utility. AI, like many technologies, does seem to have real practical applications, from writing to coding. Even if there is a short-term financial bubble around AI, the underlying technology seems likely to offer substantial benefits in the future. The question is, what is the legacy in the short-term?

Other technologies around which bubbles formed had supply chains and upstream infrastructure that were part of the initial build-out. Foundries prepared steel for rails, and someone made shovels for navies digging canals. What might make the current AI bubble different is that these sorts of enablers might be the most important part of its legacy. 

Former UK Prime Minister Gordon Brown learned 20 years ago: governments can't stop boom and bust. But what they can do is think about what happens when the inevitable bubbles burst. 

In the case of technologies, this should involve understanding how the accompanying infrastructure investment might be harnessed after the smoke of the crash clears. As AI shows, these opportunities might not be as obvious as one might think.

Acknowledgements

Thanks to Mark Byrne, Mallory Duran, Aidan Kelly, Shabeer Rauf, Karlis Kanders, Andrew Sissons and Paul Smith for their advice in preparing this blog. 

Please note: this blog is not financial advice. It was written with the support of AI, but the ideas are those of the author.