Why the AI business model doesn’t compute
Bubbles are a good thing. When they burst, they help us recalibrate what we are doing and why. New players and ideas get a chance to emerge. The key factor is to ensure you’re not hurt too badly if/when it pops.
As an economist, I’m reasonably certain we’re in a financial bubble around AI. (Side note: this doesn’t mean the technology itself isn’t transformative, just that the current economics around it might not be sustainable.)
Let’s have a look at the numbers and how technology behaves, and you can make up your own mind.
Trillions Going Out, Billions Coming In
Major AI companies have invested $2.4 trillion in AI infrastructure.
Yet the frontier models are generating only around $78 billion a year in revenue. (Note: this doesn’t include AI revenue from Google/Microsoft, but that is largely existential revenue where AI has been embedded.)
That is an extraordinary gap between what is being spent and what is being earned. For the investment to make sense, the gap has to be closed. The problem is that the investment in AI assets won’t remain productive for very long.
Data Centres Have a Short Shelf Life
Let’s start with this fact: the GPUs inside AI data centres may need to be replaced every 2 or 3 years. Worse, upgrading the GPUs can mean replacing much of the infrastructure around them. New GPUs can’t be plugged into the old ones as an ‘upgrade’, as they are not usually backwards compatible. On top of that, their power consumption, heat dissipation and networking requirements can be fundamentally different. That means replacing the chips may also require rebuilding or substantially retrofitting the data centre itself.
These are immensely expensive assets with unusually short economic lives.
This is quite unique: previous bubbles left behind durable infrastructure. By 1900, much of America’s railway network was already in place, and parts of those transport corridors remain in use to this day. The average large US power transformer lasts around 40 years. The dot-com crash literally left billions of kilometres of excess fibre-optic cable we still use today.
Data centres will be different. Here’s a clue: how many old phones and laptops do you have sitting in a drawer or cupboard? It’s the same thing, just way more expensive.
AI data centres are quasi-temporary infrastructure—yet investors are ‘going long’ on the spend.
Video of the Week – Do you trust Meta?
The Unit Economics Are Backwards
This leaves the industry facing three simultaneous, unsustainable pressures: poor unit economics, rapid infrastructure obsolescence (2–3 years), and capital expenditures vastly outstripping revenue.
That cannot continue indefinitely. Something has to change—here are some things that might…
The Stock Market Is Concentrated Around the Same Bet
There’s the mirage of scale—seven companies now represent roughly 35 per cent of the US stock market.
Welcome to the AI industry’s weird, circular economy. Tech giants aren’t just selling hardware; they are actively financing their own customer base by pouring enormous sums into data centres, cloud networks and the AI models that require their chips. In this closed loop, one giant’s massive capital expenditure instantly becomes another’s record-breaking revenue. If it looks like an inside game of hot potato where everyone is buying from each other, trust your instincts… And then don’t forget that most AI activity is still losing money.
The companies making most of the money from the boom are those making the ‘centres’ to hoover up and store the data—those laying the concrete, putting up the walls and providing the energy. The risk is not that AI disappears. The risk is that the financial assumptions attached to it do.
Intelligence Is Becoming a Commodity
Technology is a game where smaller is better—it always has been, and AI will be no different.
More and bigger data centres won’t necessarily lead to more or bigger profits, and may not even lead to better AI models.
If you’ve got grey hair, you may remember when computers were huge mainframes: highly centralised pieces of infrastructure. In many ways, it feels like that all over again.
And just like last time, we’ll find a way to create AI models that can operate on smaller, cheaper and more decentralised technology. We’ve seen this movie before.
Already, the price of intelligence is falling extraordinarily quickly, putting pressure on an already precarious business model.
Chamath Palihapitiya recently compared one million AI tokens—the unit of currency most AI models operate on—to a “barrel of intelligence”, likening it to a commodity such as oil.
Using his figures, that barrel costs approximately:
• $56 from Anthropic.
• $26 from OpenAI.
• $1.50 from Meta.
• $1 from xAI or Google.
• $0.50 from Chinese models.
By this comparison, some Chinese models are now approximately 112 times cheaper than Anthropic per million tokens.
If AI is analogous to oil, imagine one producer selling a barrel for $56 while another sells a comparable barrel for 50 cents—and the oil is almost a perfect substitute.
Such a price gap cannot survive forever.
The Future Is Small… LLMs → SLMs
AI is already better than most people at most things. Sure, they still need guidance, but we don’t need smarter models. We need more specific ones.
Comedian Steven Wright once said:
“I went to a general store, but they wouldn’t let me buy anything specific.”
It is a great joke. It is also a classic business insight. People and businesses rarely have general problems. They have specific ones.
A law firm does not need an AI that knows everything about astrophysics, poetry, zoology and medieval history. It needs an AI that understands contracts, legislation, precedents, discovery and legal workflows. An engineering company needs engineering AI.
A hospital needs medical AI. Yet today, we use giant general-purpose models to solve narrow, specific problems. We carry huge costs for capabilities the end user doesn’t need.
It is like hiring every academic at a university to answer one accounting question.
What we need are Small Language Models trained on narrow domains while being less resource-hungry.
Mind the Gap
The great investor Charlie Munger said, “The market can remain irrational longer than we can remain solvent.” And here we are again—albeit with the most powerful technology we’ve ever invented.
The question is whether two or three companies can each spend $200 billion, $300 billion or $400 billion a year, with the revenue not looking as though it will make up the gap.
They say AI will change everything. It’ll probably change the corporate landscape currently in its orbit too.
Keep Thinking,
Steve.