AI is close to breaking. Here's why.
Matheus Cardoso
I'm 23, five years into tech, with a computer science degree and a thesis on generative AI. Is it just me, or is this bubble getting more inflated by the day? These companies keep burning truckloads of money every month — and all of a sudden, the tone from the executives has changed.
The tone changed
They're now pushing IPOs at valuations larger than anything on record, and at the same time people like Sam Altman are asking the government for help propping up their own companies. That's textbook late-stage bubble behaviour — exactly the sort of thing you see shortly before one bursts. Everyone racing to find the next investor willing to hold the bag before the music stops.
But why now? A series of recent shifts have exposed the structural fragility of this bubble. It's worth stepping back.
We don't have AGI
I know this bothers anyone convinced AGI is right around the corner. But we don't have AGI. Yes, some well-built agent loops and good tool integrations have shown up. Underneath, though, the current state of AI is still a probabilistic parrot predicting the most likely next word from past training data. A model like that isn't capable of real reasoning or logic.
And because of that derivative nature, AI isn't capable of the original innovation that would fundamentally raise economic productivity. It isn't inventing the warp drive any time soon. What today's technology can plausibly do is replace repetitive human cognitive work.
The debt math
My conviction is that the AI boom is a giant, heavily leveraged bet that AI will manage to replace that human work profitably. Why is labour replacement the most logical use? The numbers.
Somewhere between three and four trillion dollars has already gone into the American AI industry. Some of it is investor capital, but most of it is corporate debt. And debt has to be paid.
At a normal corporate bond rate, somewhere between 3% and 4%, two to three trillion in debt generates roughly a hundred billion dollars in interest every year. The industry has to produce at least that much profit just to break even.
Now assume a healthy 10% margin — which they don't have, but assume it. To produce that profit, the industry would have to displace a slice of the American economy worth roughly a trillion dollars a year. And the only slice big enough to absorb that is the white-collar labour market, valued at ten trillion.
A trillion dollars of white-collar jobs is another way of saying ten million American workers a year. That plan is what explains why Sam Altman and Dario Amodei have spent years prophesying a jobs apocalypse.
The Chinese open models
Except the plan isn't going as expected. American frontier models like ChatGPT and Claude are closed: the companies control the algorithms, the data and the infrastructure, and sell access by subscription. The problem is that training and running those models is absurdly expensive. The whole operation runs at a loss — they lose money on every API call.
On its own that wouldn't be news. Tech companies have long subsidised a service below cost, taken share, become a monopoly, and only then raised prices. The problem is that this time the competitors didn't die.
With a fraction of the compute available in the United States, Chinese companies built competitive open-source models — by most metrics slightly behind, level with, or even slightly ahead of the best American frontier models. Being open, they can be downloaded for free and run on the customer's own infrastructure, at a tiny fraction of the cost.
I've been using Moonshot's Kimi3 for a few days, and for my use cases the performance is comparable to the more limited tier of Claude Fable 5. I'm not the only one noticing: according to OpenRouter, Chinese open-source models already account for more than 60% of all tokens used by American companies.
So the idea that American AI companies could build a monopoly, raise prices and cash in — that idea is off the table.
And local models
You can take an open frontier model and, with quantisation and distillation, compress that enormous thing into a far smaller local one. Instead of a giant data centre, it runs on your home computer, a home server, or even a decent laptop.
There are two advantages. Running locally, you pay no subscription to any big tech company. And they work fully offline, which guarantees privacy over your own data.
Over the past few months these models have become surprisingly capable. A good local model like Qwen 3.5 is roughly equivalent to Claude Sonnet 4 on most tasks. As basic work becomes easy to solve that way, willingness to pay for premium AI drops sharply.
The productivity never arrived
And here is the most important reason: the productivity gains are arriving far more slowly than expected. We are nowhere near ten million white-collar workers being laid off by AI this year.
I've written about my own experience with agentic AI as a software engineer and about how much work it takes to get a quality result: managing context, working around gaps in the training data, building consistent workflows, avoiding hallucinations. Those challenges are real. Using AI productively takes a lot of effort and a lot of human intelligence.
And this isn't limited to engineering. Customer support was long considered an easy target for automation. A recent study I read surveyed thousands of companies and found that more than 70% of support agents that reached production had to be switched off or rolled back because of errors and miscommunication. The same study showed companies that jumped the gun, laid off their human agents, and had to hire them back in a hurry.
The summary is that we are far from replacing ten million workers a year. The most optimistic estimates point to fewer than a hundred thousand — which explains why both Altman and Amodei have been walking back the jobs-apocalypse prophecies.
What I'd do
The American AI industry borrowed and spent enormous sums on the premise that AI will replace cognitive work at scale and generate immense profits. Don't get me wrong: I do believe AI will have a transformative effect on the economy over the long run, the same way railways or the internet did. But like those earlier technologies, it simply isn't productive or reliable enough to do that today. And these companies need those profits now, to service a mountain of debt. The math doesn't work.
That's why I believe the AI bubble, like the railway and internet bubbles before it, is going to burst — and probably soon. So here's my only advice: these companies are going to get desperate to raise money and keep the show running, and I expect them to say or do almost anything to keep it going. One of those ways will be wildly overvalued IPOs, hoping retail investors pile in and buy the dream. Don't be that investor. When the music stops, the people at the top will already have sold — and someone is going to be left holding the bag. Don't let it be you.
That's what I had to say. If you'd like to keep the conversation going, find me on X.