Artificial intelligence models can produce apparently brilliant responses to human queries which on closer examination reveal bits of possibly regurgitated nonsense. A similar dynamic may be at work in the ballooning investments and valuations of technology companies seeking riches from self-teaching computers. From railways to telecom networks, previous booms in capital-intensive industries show that new technologies can change the world, yet leave investors short-changed.

Investors cannot get enough of AI. OpenAI, the startup behind the ChatGPT chatbot, raised funds in February at an $80 billion valuation, or nearly three times as much as a year ago. Smaller rivals Anthropic and Mistral have seen similar gains. Yet established tech giants have enjoyed the biggest windfall. The collective market capitalization of Microsoft (MSFT.O), (AMZN.O), Alphabet (GOOGL.O) and Meta Platforms (META.O) has risen by $5 trillion since the release of ChatGPT in November 2022.

For firms designing AI models and building the networks that train and run them, the response has been a spending arms race to grab new markets and prevent rivals from disrupting their existing businesses. Microsoft, Amazon, Alphabet and Meta collectively splurged $200 billion on capital expenditure last year, with about half that sum going towards technological infrastructure, according to Bernstein. The research firm estimates this will increase by over 50 per cent this year, mostly because of AI.

Compared to these huge sums, the income generated by AI remains small. Sequoia estimates generative AI now brings in about $3 billion a year in revenue, up from zero a year ago. However, the venture capital firm notes that it took companies offering software as a service a decade to reach a similar threshold. The promise of future riches has unleashed some gigantic forecasts for the spending needed to make the advanced semiconductors, build the data centers, and generate the power to train and run AI models. OpenAI boss Sam Altman earlier this year put the figure at up to $7 trillion, a person familiar with the matter told the Wall Street Journal.

Such a wild binge might even be justified. Smart systems which automate more labor, manipulate bigger piles of data, and complement or surpass human intelligence could be worth many multiples of that amount. Moreover, the world has experienced similar-sized booms before.

Take railways. Cumulative investment in building the United Kingdom’s train network between 1845 and 1850 was equal to about 30 per cent of Britain’s gross domestic product in 1850, according to economic historian Andrew Odlyzko. At the end of the 20th Century, telecom firms borrowed $1 trillion to build out fiber-optic cable and mobile phone networks, according to the chairman of the US Federal Communications Commission. That was equal to about 3 per cent of world GDP at the time. A $7 trillion investment program today would equal about 30 per cent of US GDP for one year, or about 7 per cent of the world’s annual economic output.

The investments currently being made will only make sense if they generate huge profit, however. To see what kind of income tech firms might require, start with Nvidia (NVDA.O), whose chips are the default for training AI systems. Analysts expect the company run by Jensen Huang to bring in almost $100 billion of revenue from its data center division this year, according to LSEG data. Though this unit’s sales aren’t all AI-related, most of it probably is.

Assume AI servers using these chips have a useful life of four years. Other costs such as networking gear, space in data centers, and the power required to run AI models add about 75 per cent to the bill, according to research outfit SemiAnalysis, lifting the total outlay to $175 billion. Now assume the tech companies making these investments expect a 40 per cent operating profit margin. This implies they collectively will need to earn $292 billion of AI-related revenue over four years, or about $73 billion a year. That’s certainly possible, but still a long way from the $3 billion a year generative AI is bringing in today.

It’s a feature of new tech markets that early estimates can wildly undershoot the mark. In 1980 McKinsey tried to guess the size of the mobile phone market 20 years hence, the real figure was 125 times the consultant’s estimate. Today, launches of AI products are routinely described as “magical”, while companies marvel at supposed 30 per cent productivity increases from deploying the technology. As in the telecom boom, there’s an obvious benefit to size. The biggest gains in AI over the past few years stem from models becoming bigger, using more data and more computing power.

However, investment booms more often disappoint their financial supporters. Investors in British railways overestimated demand for the new form of transport and had lost about a third of their money by 1850. The country entered a depression. The telecom frenzy of the late 1990s also produced gargantuan losses. Huge investments in laying fiber-optic cable and overpayment for European licenses to operate 3G mobile networks doomed many startups to failure, and big companies to years of sub-par returns.

The AI bonanza comes with these pitfalls, and more. Railway track and fiber-optic cable, once installed, lasted for decades. The lifespan of an AI model and the servers on which it runs is much shorter. Changes in human tastes or behaviors, new information, or training on flawed data can quickly make an expensive large-language model obsolete.

Multiple giant companies and many nations want to develop AI, so advances may be commoditized. And if AI does prove superior to human intelligence, it seems probable governments will impose strict limits on its uses – reducing the resulting profit.

Perhaps the end result will be much like both the railway and telecom booms. The technological visionaries were right: people can now read this article on their smartphone while riding on a train. That’s some consolation for the many investors who lost their shirts in the process.