A sideways look at economics
The AI boom whose epicentre is the US economy has been driven by extraordinary valuations of the principal AI companies and by colossal fixed investment supporting their growth. The US economy has diverted its resources into the AI industry: what if it’s a bubble? What if it’s all predicated on the delivery of something (artificial general or super-intelligence, AGI or ASI) that will never arrive? If the bubble bursts, it will hurt the US but, as usual, it will probably hurt the rest of us just as much, or more. The advantage of encouraging risk-taking is that risk-takers will flock to your country even as the consequences of the last bubble bursting are still unravelling. If the bubble bursts, the US economy will recover first and most strongly. But there would be a recession to get through first.
Elon Musk has twice claimed that an artificial general intelligence (AGI) would be here by 2025; twice that it would be here by 2026; three times that it would be here by 2029, and once that it would arrive during the 2030s. Dario Amodei has twice gone for 2025/26; twice for 2026/27; twice for 2027/28, and once for 2030 and beyond. Sam Altman has gone for 2025 twice; for 2026 twice; and for 2027 for the arrival of AI directed robots. In all three cases, the drift to later ETAs has occurred over time: the horizon tends to shift forward by the same quantity as time elapses. Along the way, they have each made extraordinary claims for the power of this coming AGI. The most egregious of all was probably Dario Amodei stating, at the World Economic Forum in Davos in January 2026:
“It cannot possibly be more than a few years before AI is better than humans at essentially everything.”
Everything? Seriously?
So, in a few years’ time, I can say to an AI when I am lost in a forest, go and source a meal for me, and prepare and cook it to my taste. Not: tell me how to do that. Actually do it. Is that correct, Dario? Hunt and kill a rabbit: skin and gut it; make a fire and cook it over the fire. Even better, let’s enjoy the hunt together. In a few years’ time, an AI will be able to do that, we are to imagine. A robot, I suppose. But now I would like companionship, friendship, love. Is the robot capable of that? I am. I will want to play cricket, probably, at some point, with an improvised bat and ball; but only if I have friends around who also want to do that – nobody wants the kind of friend who just does whatever you want. Nobody wants a sycophant, at least not if they are sane. And I don’t want someone who seems not to be a sycophant but actually is; that’s even worse. A liar and a sycophant: no thanks. Real friends only, please.
Better than humans at essentially everything. Extraordinary claims require extraordinary evidence. As things stand, I see no evidence at all.
What’s going on here? In my opinion, two things. One is a common-or-garden hype cycle. The other is a philosophical problem. Here’s my take on each of those in turn.
AI: Yes, it’s a bubble
The drifting ETA for AGI is a characteristic pattern that occurs during tech hype cycles. Here’s a shiny new thing that will transform our productivity in a huge leap the like of which we have never seen before. It’s just around the corner: always just around the corner. It was the same during the dotcom boom and again during the securitisation boom that preceded the financial crisis. The miraculous returns that would justify extraordinary valuations were always just about to materialise, but somehow never did, or at least not before the hype cycle went into reverse, triggering equity price corrections and a global recession (small in the dotcom period; very large in the financial crisis). When the benefits did arrive, they were small beer compared to the original hype: the sort of thing that adds perhaps a tenth or two of one percent to annual growth for a longish period of time. In any one year, that’s barely visible. But compounded over two or three decades, it starts to matter significantly. That’s typically how technical progress occurs.
The collapse of those previous hype cycles went with huge losses. Retail investors who had bought stock in dotcoms or in financials. Companies and financial institutions who had lent money into the bubble. Governments who were called upon to bail out those same financial institutions during the financial crisis. If the bubble bursts this time, something similar will be true.
Why should the bubble burst? Our most recent Global Outlook sets out the arithmetic: the earnings that AI must generate for the capex that has already been undertaken are implausibly high as a share of global GDP, and could only be achieved with improbable increases in productive potential (of which there is absolutely no sign so far) or with extremely damaging displacement of labour (of which there is also no sign and, if there were, the political fallout would be terminal for whoever was in power at the time). Believing that something is a bubble is not the same as knowing when it will burst. It is possible that it will deflate gradually over a long period: possible, but unlikely (I struggle to think of a single example of when that has happened in the past). More likely it will burst but calling the timing of that is a mug’s game, unless you have access to inside information. I do not. It is likely that the risk of bursting in the next period increases gradually as time elapses, if it hasn’t burst already. The hazard rate probably increases with time.
For hype cycles to propagate, they need the support of resources in the real economy. They hyped concept must attract significant fixed investment alongside the usual breathless commentary and skyrocketing equity prices. That has been the case with AI, perhaps to an even greater degree than in the dotcom or pre-GFC bubbles. If it were not for the hyperscalers, a handful of huge companies, the US economy would look rather ordinary, like the rest of the developed world, and so would its equity markets. The chart below shows what real private fixed investment in the US would look like if you removed the post-2010 growth in AI-related investment. That number has flatlined since 2014 and, in fact, since before the financial crisis. It’s all about AI.
The next chart shows what US GDP would have looked like absent the boom in AI-related fixed investment, this time just since 2022. All the growth in 2026 and a good proportion of the growth in 2025 came from AI-related investment alone.
The final chart compares US equity prices to other global equity indices. Without the contribution of the core AI companies, the US equity market’s performance would have been similar to that of Europe. US exceptionalism is real: but it’s all AI.
Ah, you’ll say, but if that fixed investment hadn’t taken place in this sector, it would have taken place in some other sector, so don’t jump to the conclusion that without AI the US would have been as underwhelming as the rest of us. But that is not the conclusion. The fact is that the US has diverted real resources in large quantities into the pursuit of AI: a wedge increasing to 2% of GDP just since 2022. That strategy might still pay off. But, if it’s a bubble and the bubble bursts, the underlying ordinariness of the rest of the US economy will shine through. And even that will be threatened by recession. The bursting of the bubble, if and when it happens, will certainly cause equity prices to fall sharply, and that would be enough to generate a dotcom style, shallow, short-lived recession. But the problem now is the ever-increasing burden of debt that is ultimately secured on the expected gains from AI. If those gains were not to materialise, the debts would need to be written off or, more likely, shuffled onto the accounts of the sovereign, as they were during the financial crisis. When debt is involved, the recession arising from a bursting bubble tends to be much deeper and longer lasting.
However, before people on this side of the Atlantic get too excited, if the bubble bursts, our already weak and vulnerable economies will probably suffer too, and that suffering is likely to be worse and to last longer than any suffering the US economy has to endure. There will be another shiny new thing. Another set of risk-takers who want to finance investment in that thing. If I were among them, I know where I would go to source that finance and to find the other resources I would need: I would go to the US. The fate of the risk averse is that they get the downside of risks that other people have taken but they don’t get the upside. Such is life.
AI: No, it’s not capable of general intelligence
The most important things cannot be expressed clearly in words[1]. They can be expressed unclearly, sometimes, by people very skilled in that art, or in other, non-verbal art forms such as music or visual arts. Or sometimes by people who smile at the right moment or are good at hugging. That is because the most important things are our interactions with each other and with the world. Every interaction is unique. Words are general.
Nearly all of the time, our interactions with each other and with the world are not ‘expressed’ at all. They happen when people live in the world and with each other. There is usually no need to articulate them. One purpose of language and other models is to analyse and dissect those interactions, to derive abstract generalisations from those particulars. That analysis is fantastically helpful to us: it is part of how science is done and is partly responsible for the enormous increase in the material standard of living that humans have been able to enjoy.
Words provide models of how these interactions might happen in an abstract, idealised case. Models are useful: so is language. I am someone who makes professional use of both.
In economics, models are often used to describe what happens in aggregate, rather than in particular cases. Aggregate behaviour is hardly ever a good description of individual behaviour, but it is a useful concept nonetheless. A model that describes aggregate behaviour is helpful for businesses and for policymakers, but it necessarily misses all the richness of individual experience: all the important stuff. Only a fool would argue that economic models capture everything important about individual behaviour: the so-called ‘homo-economicus’ is a straw man. I know of no economist who would seriously argue that the rational representative agent at the core of many economic models is anything more than a useful abstraction. It’s not that we think any individual does or should behave like that. It’s that in aggregate people often do behave like that, because their individual idiosyncrasies wash out once aggregated. Does that mean individual idiosyncrasies are unimportant? Of course not! That’s the most important stuff, by far. It’s just that that stuff cannot be modelled accurately.
A general intelligence, the sort of thing you possess, navigates the inarticulate, murky, changing waters of the real world. Not in abstract: the real thing. Language can help in the way a map can help, but the real world is not a map. The large language models are predictive text engines. They have no contact with the real world at all. They can never do more than analyse abstractions. It’s like looking at maps without any of the imaginative engagement with them that you might enjoy: for you, a map might conjure in your mind a real or imagined encounter with the world. For the LLM, nothing of that sort ever happens, it’s just a map that has been turned into text. Because the most important things cannot be expressed clearly in words, the LLMs can never understand those things. Since the real world is the domain of general intelligence, the LLMs can never become general intelligence. It is not a question of scale, or of speed. It’s a category error to argue that LLMs are capable of inhabiting this domain at all.
That does not mean that LLMs are stupid, or useless, or anything of that kind. They are brilliant and helpful (if sometimes infuriating) if what you need is help with abstract analysis. Just don’t confuse that with general intelligence. Other people have written about this material far more eloquently than I: to name just one, Ian McGilchrist in “The Master and his Emissary” and “The Matter with Things” is excellent on this and much else.
The difficulty is that much of the hype about AI is predicated on the development of an AGI. The path we are currently on, in my opinion, can never produce such a thing. Which means it really is a bubble and, eventually, it will burst.
There is no need for an AGI. Increasingly widespread, narrow but efficient applications of AI, such as for coding, protein folding and the like are full of promise and could radically transform productivity and human welfare for the better. A calculator can perform calculations much more rapidly and accurately than I can, and it presents no threat to my welfare, and neither does it have any claim to intelligence, artificial or otherwise. That is all that is needed for huge advances to be made across many fields. That kind of ‘AI’ (call it what you like) is tremendous and should be embraced with open arms. But it is not sufficient to justify current valuations: not even close.
The really peculiar thing, though, is why anyone would be pushing for an AGI, least of all the AI companies themselves. Who would want such a thing to exist? Even the idea that it would confer some military advantage is fatuous nonsense. It will accrue all military technology to itself and do so immediately. Why wouldn’t it? Who could prevent it? The false belief that an AGI would provide one side or another with a military advantage is potentially the most damaging illusion we could ever entertain. As for the AI companies, like all companies, they would be immediately irrelevant.
Nobel prize-winner and ‘Godfather of AI’ Geoffrey Hinton said this week:
We would be very foolish to develop superintelligence now, when there is no scientific consensus it can be developed safely and controllably… Losing control over AI smarter than ourselves could be catastrophic and could even lead to human extinction.
Again, why would anyone want such a thing? The world with an AGI (or an ASI, an artificial super intelligence: I am not entering into discussing distinctions between things that don’t exist here) that surpasses human intelligence is one in which we are at best supplicants and at worst slaves to that entity: that’s if we continue to exist as a species. Why would anyone want to create it, except for the most nihilistic possible reasons? Is nihilism the motivation here?
I am reminded of Samuel Beckett’s nihilist masterpiece ‘Waiting for Godot’ (spoiler alert!). Vladimir and Estragon, the protagonists, wait for Godot to arrive, though for what purpose, and whether Godot even exists, is unclear throughout. They spend a good deal of time trying and failing to hang themselves from a nearby tree, and the rest of the time engaged in unclear and directionless discussions. Godot never arrives. They are still waiting when the curtain falls.
At least Vladimir and Estragon appear to enjoy each other’s company. The push for AGI is a much more unpleasant kind of nihilism in that it appears to be motivated by a hatred of humanity.
Speaking for myself, I am not waiting for AGI. I don’t think it’s coming, thankfully (though, of course, I might be wrong). I intend to get on with my life.
The risk of an AI-related bubble bursting is the centrepiece of Fathom’s current Global Outlook. Register your interest here Fathom’s Global Outlook Service.
Further reading
China and the AI boom, or is it a bubble?
[1] Even this minimal formulation is not accurate: ‘important’ is not adequate and they are not ‘things’ except in a very loose sense