Bombolla de la IA: què sosté la tesi d’un possible esclat i què no

Hello world. I'm an unemployed ex-big tech software engineer with 25 years of experience in the tech industry. So, is it just me or is the AI bubble looking extra bubbly these days? By extra bubbly, I mean these AI companies, they're still burning through truckloads of cash every month. But all of a sudden, the tone of their leadership seems to have changed. Now they're pushing for some of these insanely valued IPOs that's higher than anything that's ever been seen. While at the same time, guys like Sam Altman, they're now asking the government to help them financially prop up their companies. This is like peak bubble behavior right there. It's something that you would see right before a financial bubble bursts. Everybody is running around trying to find the next sucker. Oh, I mean investor to take the bag before the music stops. But why is this all happening now? Well, there has been a number of recent changes that exposes the fundamental unsustainability of the current AI bubble. So, let's take a step back and

walk through what's actually going on here. Now, I know I'm going to be upsetting some true believers who think that we're at or near artificial general intelligence or AGI. But the truth is, we don't have AGI. Sure, there's been some fancy agentic loops and nice tool integrations introduced, but the current state of AI is still just a probabilistic parrot predicting the most likely next word based on its past training data. Such a model is just not capable of true reasoning and true logic. And because of this derivative nature, AI is simply not capable of left field type innovations that fundamentally increases economic productivity. Like Claude, it's not just going to go and invent the warp drive anytime soon. However, what current AI technology can conceivably do is to replace repetitive cognitive human labor. And it is my belief that the AI boom is a gigantic leveraged bet on AI profitably replacing this human labor. Why is the replacement of human labor the most logical use case for AI you say? Well, if we look at the hard numbers, right? Somewhere between 3 to 4

trillion dollars have been invested in the American AI industry so far. Now, some of that money is in the form of investor cash, right? But the vast majority of that money is in the form of debt. Corporate [snorts] bond debt. And that debt has to be serviced. Suppose we use a normal interest rate for corporate bonds like 3 to 4%. Well, for 2 to 3 trillion dollars of debt, that works out to be about 100 billion dollars in interest that has to be paid every year. That means the AI industry has to make at least that much profit every year just to break even, just to service their debt. Now, suppose they got a good profit margin going, say 10%, which in reality they don't, but suppose they did. To make that kind of profit, the AI industry would need to replace a slice of the American economy that's equivalent to around a trillion dollars every year. And guess what? The only slice of the American economy that's big enough to sustain this kind of replacement is the 10 trillion dollar white-collar jobs economy. And that's

why AI must profitably replace white-collar jobs to keep this bubble going. Now, profitably replacing a trillion dollars worth of white-collar jobs, that's like saying we have to profitably replace 10 million American white-collar workers a year. And this plan, I think, is why all your AI leaders like Sam Altman and Dario Amodei, they've been going around for years now prophesizing that huge amounts of jobs will simply disappear in a kind of job apocalypse. But things are not going according to plan. In fact, the plan is actually turning into a kind of dumpster fire right now due to a couple of key reasons. Let's get into these reasons. Now, frontier American AI models like OpenAI's ChatGPT or Anthropic's Claude, for example, these are all closed models, meaning the tech companies behind these models, they control everything around the models. The algorithms, the data, the compute, everything. And they can sell their AI models to consumers for money in the form of subscriptions. But the problem

is that these closed AI frontier models, they are insanely expensive to train and to operate. The entire operation is grossly unprofitable. Companies like OpenAI and Anthropic, they're literally losing money on every single API call being made to their models. But here's the thing. Tech companies have long used a strategy where they would burn tons and tons of money to subsidize a service below operational cost. Then they would try to gain market share and become a monopoly. Once all the competitors are dead, they can then jack up the prices and profit. Now, I have spoken at length by another vlog on all of the algorithmic, software, and hardware innovations that Chinese AI companies have been making in this space. But, long story short, with just a fraction of America's compute resources, these Chinese tech companies have managed to create competitive open-source models. Models that, by most measures, are either slightly behind, on par, or even slightly ahead the best American frontier models. These Chinese models, being open-source, it means that

they can be downloaded for free and then run on a customer's own compute infrastructure. And this could be done for a tiny fraction of the cost of using American closed AI models. A concrete example of this is Moonshot's Kimi 3 model, right? I've been using this model for a couple of days now, and to me, this model's performance is comparable to the nerfed version of Claude 5 Fable, at least for my use cases. And I'm not the only person recognizing this, right? According to OpenRouter, Chinese open-source models now account for more than 60% of all tokens used by American firms. So, the idea that American AI companies can somehow create a monopoly, jack up the prices, and rake in the profits, this idea is now off the table. Now, speaking of open-source models, you can take a frontier open-source model, and through techniques like quantization and distillation, you can compress this massive model down into a much smaller local AI model. And instead of being run on a big data center somewhere, these local AI models can be run on your home desktop, or a home server, or even a good laptop. Now,

there are two benefits to this approach, right? One is that because these models are running locally on your own computer, you don't have to pay any subscription costs to the big tech AI companies. The second benefit here is that these local models can be run entirely offline, disconnected from the internet, giving people total privacy over their own data. And in the last couple of months, these local AI models have suddenly become very capable. A good local model like the Gwen 3.5, for example, it's roughly comparable with Claude Sonnet 4 on most tasks. And as basic tasks are easily handled by these local open-source models, people's propensity to pay for premium AI from these big tech companies, well, it diminishes drastically. And all of this leads to the most important reason, which is that productivity gains from AI is happening way slower than expected. We ain't seeing anywhere near 10 million white-collar workers being laid off by AI this year. Now, I have spoken at length about my own experiences with agentic AI as a software engineer, and

all the challenges around getting good quality output from AI. Fundamentally, the challenges around managing context, tiptoeing around training data gaps, creating consistent workflows, and avoiding hallucinations. These challenges are real. It just takes a lot of human effort and human intelligence to use AI productively. And these challenges are not just happening in software engineering. It's happening in other fields like customer service, which was long considered to be this low-hanging fruit for AI to automate. There was this recent study that I read where they surveyed thousands of companies and it turned out that over 70% of the customer service agents that went live, well, had to be either rolled back or shut down because of various mistakes and miscommunications that it was making. The same study actually showed a number of companies had prematurely jumped the gun and laid off their customer service reps only to then frantically having to hire these people back. So, the bottom line here is that we're nowhere near being able to replace 10 million white-collar American workers

a year with AI. The most optimistic estimates show something less than 100,000 workers being replaced a year by AI, which I think is probably why both Sam Altman and Dario Amodei are now kind of walking back their AI job apocalypse prophecies. So, net net, the American AI industry has borrowed and spent huge sums of money. They did it based on the assumption that AI will replace human cognitive labor at scale and for immense profits. Now, don't get me wrong. I strongly believe that AI will have a transformative effect on the economy over the long term in the same way that railroads or the internet did. But AI, just like these past technologies, is simply not productive enough or reliable enough to do this today. And these American AI companies, to service their mountain of debt and not go bankrupt, they need to realize these profits today. And the basic economics of it all, it just doesn't work out. Thus, I believe that the AI bubble, very much like the railroad and internet bubbles, is going to pop and likely very soon. So, here's my one piece of advice to

you. These AI companies are going to be desperate to raise money to keep this whole show going and I expect that they'll say or do just about anything to keep the machine running. One way they will try to do that is through wildly overvalued IPOs. They're hoping that retail investors are going to pile in and buy into this dream. Don't be that investor because when the music stops, the people at the top they would have all cashed out and someone will be left holding the bag. Don't let that someone be you and that's all I have to say about that. Hope it helps. Anyways, if you have a morbid curiosity to join me on this life journey, please feel free to subscribe to my channel and subscribe to my Substack newsletter. If you want to support me in my V log creation efforts, please consider becoming a paid member of this channel, a paid member of my Substack, or just buy me a coffee. If you would like a one-on-one coaching session with me, please feel free to schedule it. Anyways, thanks so much for watching. Talk soon. Bye.