I’ve been in the market long enough to know that the word “bubble” gets thrown around with the same enthusiasm as red cards against Springbok locks in November. Everything is a bubble until it isn’t and nothing is a bubble until it bursts. Even the Nobel laureates can’t agree.
Robert Shiller, whose timing during both the dot-com craze and the housing boom remains spooky to this day, describes a bubble as a kind of social contagion passed from one person to the next through stories of sudden fortune and rising prices.
Cliff Asness offers a more clinical test when he says a bubble is when no “reasonable future outcome” can justify present prices. And Eugene Fama, ever the efficient-markets purist, essentially throws up his hands and says bubbles are things we identify only after the wreckage is visible.
Taken together, these definitions tell us two things. First, bubbles are part mathematics, part mass psychology. And second, you almost never know you’re standing inside one until it’s too late. Which may explain why the question I keep hearing is: “Are we in an AI bubble?”
It’s a fair question. Valuations at the top end of the US tech market are brushing levels last seen during peak dot-com. The cyclically adjusted price:earnings ratio has climbed above 40. Nvidia’s market capitalisation has overtaken the GDP of entire Group of Twenty countries.
If Shiller is right that narratives fuel manias, there has never been a more fulsome narrative than AI being this omnipotent technology that promises to rewrite every industry, conquer inefficiency and generate profits on a civilisational scale.
But hype alone isn’t the giveaway. The giveaway is the maths. And the increasingly strained explanations now required to justify the capital flowing into this space.
A full R400bn. That is the amount the major tech platforms— Microsoft, Google, Amazon, Meta and their fellow travellers — are expected to spend on AI data centres and associated infrastructure in 2025. SA’s entire GDP.
And yet when you ask what all this capex will produce, the answers are as clear as the black boxes that large language models have become. “Transformational AI agents.” “Full-stack inference optimisation.” “Compute-as-destiny.” Maybe I’m just a Luddite, but I can’t see where the earnings are going to come from to justify this kind of capex.
Charlie Bilello, a market commentator with the gift for deflating financial nonsense, likes to say that if someone can’t explain a technology in one simple sentence, it’s probably a bubble. Try explaining generative AI’s business model in a sentence.
Analysts estimate that for AI to justify this spending it will have to deliver $60bn-$80bn in fresh earnings before interest, taxes, depreciation and amortisation within a few short years. That’s an industrial revolution’s worth of earnings in one cycle. Not a chatbot that drafts your emails in the tone of a tired management consultant.
The clearest sign that we’re near the peak is Sam Altman’s recent comment, delivered with his breezy all-low-caps confidence, that AI data centre buildout would require “trillions” and that governments may have to provide a backstop. Not after the crash or during the crisis, but now. In advance. Before anything has gone wrong. He later tried to walk it back, but the slip was revealing because this is where the 2008 parallel becomes uncomfortable.
Before the financial crisis banks behaved as if they had a government guarantee, courtesy of the famous “Greenspan put”. But they never walked into Washington and requested one explicitly. Their bailout pleas arrived after the fire started.
As Peter Armitage of Anchor Capital told me last week, “this AI spend boom, which is the biggest capex boom in the history of mankind, has happened at the same time as falling inflation and lower US rates. It’s been fortuitous. But it’s also forced.” His point is that the hyperscalers can’t afford not to spend.
And with that comes overbuilding, the oldest risk in economics. “The history of capacity is [that] people tend to overbuild,” Armitage warned. “And that’s what’s concerning people.”
Which brings us to Nvidia’s results on Wednesday. Up 1,000% since ChatGPT launched and 40% year to date, Nvidia has become the pick-and-shovel trade of the modern gold rush, the company everyone buys because it’s selling the hardware everyone else needs.
Armitage put it starkly: “This is all going, bizarrely, to one company, Nvidia. They make a 75% gross margin. They make a 65% operating margin. They’re making wildly excessive profits because of the scarcity of their product.”
But he also pointed to the danger of extrapolation: “Nvidia is now a $4.5-trillion company. It was number 60 in the world a few years ago. It’s raced to number one. But you’ve hit the law of diminishing returns. If everything is phenomenal, can it become a $6-trillion company? Possibly. But a lot of the upside has happened already.”
And Nvidia’s outlook depends almost entirely on one thing: unsustainably high hyperscaler capex. If Microsoft, Amazon, Meta or Google flinch, even slightly, Nvidia’s story buckles.
Morgan Stanley has upgraded Nvidia’s target price once again, citing strong demand for its Blackwell chips and early enthusiasm for the next-generation “Vera Rubin” line.
The risks heading into Wednesday fall into a few clear categories. First, capex discipline. If any hyperscaler hints at slowing investment, the music stops. Second, margin pressure. Nvidia’s chips do engineer miracles but are expensive to make. Competition from ASICs — chips that do one job faster, cheaper and more efficiently than a general GPU — could pressure margins.
Third, inventory build-up. If chips are stacking up in warehouses instead of racks, that’s a red flag. Finally, the biggest risk of all is the mismatch between hype and revenue. For all the talk of AI transforming the global economy, tangible revenue at the hyperscaler level has been — let’s be kind here — modest.
This doesn’t mean AI isn’t real. AI may well reshape entire sectors. Nvidia may beat earnings by a wide margin, again. But when a booming industry starts asking governments to absorb the downside before the upside has arrived, let’s just say caveat emptor.
• Avery, a financial journalist and broadcaster, produces BDTV’s ‘Business Watch’. Contact him at michael@fmr.co.za.
Would you like to comment on this article?
Sign up (it's quick and free) or sign in now.
Please read our Comment Policy before commenting.