Why the AI boom could become a 2026 stockmarket nightmare
This fund manager says investors need to prepare for lower returns as the AI boom faces an inevitable reckoning over unsustainable costs.
At the outset, let me state unequivocally that no one knows whether the equity market will crash. The fact is, we can’t even definitely identify a bubble until after its demise, which therefore means we cannot know for certain if we are in one.
With that caveat out of the way, I am reasonably confident we should expect greater volatility and lower returns in 2026. Let me explain why I think that’s a reasonable assessment.
Since 2022, I have suggested that investors maintain a bullish disposition.
In January I said that I had hitched my wagon to the bullish camp since 2022 and that as 2025 progressed – provided the prospects for deflation, economic growth, earnings growth and supportive liquidity were positive – the underlying narrative would be one of a durable bull market.
It’s panned out well, so far, but I fear changes are afoot.
Support from positive economic growth, disinflation and expanding liquidity can no longer be assured, and that’s because inflation is not declining much, if at all, economic growth is slowing (the majority of US growth can be attributed to the construction of data centres), and global liquidity growth is also slowing.
Consequently, the volatility we saw a little of this year could be more pronounced in 2026.
Moreover, with three years of very solid returns under our belt, valuations have risen to the point that future returns are expected to be lower over the next few years. If they are positive, American economist Robert Shiller expects they will amount to the low single digits for the S&P 500.
We’re living and investing during one of those infrequent episodes called a thematic boom. This time, the theme is AI. What is less well known is that the AI boom is merely a manifestation of abundant liquidity. When there’s plenty of money floating around – due to central bank largesse – that money needs a home. If the period of abundant money coincides with the emergence of a new general-purpose technology, money will converge on the companies that benefit from the scale and sale of that technology. As they rise in market value, they begin to dominate not just the narrative but also the market indices.
And that’s what we have seen for the past few years. Excessive liquidity found a home in AI stocks, driving prices higher and culminating in Nvidia’s market capitalisation reaching $US5 trillion ($7.6 trillion), despite the company’s forecast 2026 revenue amounting to $US65bn, or just 1/77th of its market cap.
We could experience heightened volatility and lower returns in 2026 because a little bit of mania has crept into markets.
But investors need to recognise that the business of AI is fundamentally different from the high unit profitability of the software businesses we’ve invested in since 2010. Software as a Service (SaaS) businesses enjoy very high margins – up to 85 per cent because the software is written once and sold infinitely.
Why AI is different
AI, however, is different. Delivering AI-generated outputs introduces a physical cost to every digital interaction.
For most of us, apps such as ChatGPT, Grok, Claude, Gemini and Replit deliver a polished experience that makes us feel that AI is an elegant, easily accessible tool that improves productivity and generates new opportunities.
Under the hood, however, what’s required are multibillion-dollar models, trained on thousands of GPUs (each built on $US250m Extreme Ultraviolet lithography scanners) housed in multibillion-dollar data centres that require city-sized electricity supplies and, locally at least, an estimated 20 per cent of Sydney’s water supply.
And after all the investment, providing the output also costs the AI company money.
It’s estimated that a single ChatGPT query uses 10-15 times more energy than a traditional Google search. And while a Google search costs Alphabet roughly US0.003c, a generative AI response can cost upwards of US1c-US2c. It might not seem much, but that’s a 300 to 400-fold increase in marginal operating costs.
On the subject of energy, and by way of example, OpenAI’s Sam Altman has said his “audacious long-term goal is to build 250 gigawatts of capacity by 2033”. To put that in perspective, that’s more than India’s entire peak power demand in 2025. India has almost 300 million households and is the world’s fourth-largest economy.
To manage this, Truthdig estimates OpenAI would need to purchase 30 million GPUs a year, and run them 24/7, 365 days a year. That would cause faster burnout, requiring more frequent replacements and upgrades.
To meet those requirements, OpenAI would need the world’s 10 most advanced fabricators to operate around the clock all year, every year, demanding levels of energy that would squeeze the industry and drive up prices by reducing availability. And that says nothing of the price and reduced access to clean supplies of water, which would have adverse long-term societal and health implications.
Stockmarkets tend to cast their shadow before them, meaning investors will exit before the rubber hits the road, and 2025’s AI dream could quickly turn into a 2026 nightmare.
For all of those reasons, and a few more that will require covering in a future column, I suspect 2026 will look a little different to 2023, 2024 and 2025: Lower returns with more volatility.
Roger Montgomery is founder and CIO at Montgomery Investment Management.
More Coverage
To join the conversation, please log in. Don't have an account? Register
Join the conversation, you are commenting as Logout