It's certainly a lot of debt financing at a rather vast scale. Just like the impact of tariffs take a while to set in, there are deeper consequences for the AI infrastructure rush. I'm not always clear on who it is an opportunity for. Certainly people are making money off the FOMO in various ways including and mostly, the stock market.
Building AI infrastructure and high-performing models requires colossal investments and massive amounts of training data. Yet, it will still take years before many of these products, services, and business models achieve meaningful penetration and profitability.
That’s why the smartest U.S. companies are partnering strategically. It’s like when they built the American railway network before selling a single ticket — yet the wealth, progress, and opportunities that network created were phenomenal.
We’re witnessing something similar today. The real question is: who will compete with them once the network is built and their models are perfected, trained with our data and feedback (thanks to a $25/month subscription)?
At that point, they could charge us, and our companies, whatever they want, as we’ll be fully dependent on them.
I lived in the run up to and well beyond the dotcom crash in 2 markets - ISPs/Datacenters & cyber security. One massive factor was your #7 - mad as it seems a 2meg leased line in 1995 was 200k per year (🤯) & in 2001 it was under 2k. So competitive market price collapse & Moore's Law will be a future factor in the AI bubble. That and all the bad investments not playing out means and investors fleeing as a consequence. 2 major pins hitting the balloon or balloons 🎈🎈
It’s a powerful tool in two specific domains of ML: Computer Vision and NLP. I still think there is more potential with what can be done with deep learning in this fields as they are very fast, but I’d say the flurry of innovation of new tech has slowed. Convolutional Neural Networks and GANs are 8 and 6 years old respectively and no other completely novel and useful architecture has come out since that isn’t something we’ve seen before. At first, combination of CNNs and GANs lead to the Deep Convolutional GAN (DCGAN). We also have Convolutional LSTMS and other hybrid models, but I would argue we’ve found all the good ones and not many talk about that.
People are just applying these models in different ways. Eventually we will take deep learning as far as it can go, and there will be a sobering realization for the super enthusiasts that there are clear limitations given our CURRENT understanding of deep learning.
Outside of computer vision and NLP, deep learning is overhyped. Few realize that deep learning is almost certainly not sufficient for AI, and even fewer want to admit this is true. It’s not suitable for business because classic ML models, in particular Gradient Boosters, frequently outperform deep neural nets, and even if they perform slightly worse, they are MUCH faster to train. In most business contexts where you work with tabular data (often from a database), deep learning is not the correct algorithm for classification or regression.
Finally, in the business world, extracting insights from data and acting on it is most important. Data is the new oil, and businesses are in an arms race to refine it, develop refinement techniques for it, or just simply acquire high quality data and sell it. Deep learning is a poor refinement technique and is expensive. That’s why it’s not popular in the business world as much.
Insightful breakdown. Circular financing masks real demand and inflates valuations, classic bubble dynamics. The opportunity is real, but sustainability depends on genuine adoption, not just capital loops.
For more AI trends and practical insights, check out my Substack, where I break down the latest in AI.
It is not a closed system. Money in -> Money spent -> Revenue out.
Amazon trading at a loss for over 2 decades is the poster child for building value through investing, and turning on profit as growth objectives are achieved.
This feels like déjà vu — but with a balance sheet.
In my Leader’s Dispatch: The Robot Uprising Imploded — Here’s Why Humans Are Back on Top, I unpacked how companies like Klarna and IBM learned the hard way that you can’t automate empathy or outsource trust.
Now Guillermo shows us the same drama playing out with money. The AI market’s gone fully meta — funding itself to buy itself so it can sell to itself again.
Turns out, whether you’re replacing people or pricing sanity, circular logic eventually collapses under its own cleverness.
Most of the these companies also have customers outside of the circular structure. How do those revenues stack up and how fast are they growing?
It's certainly a lot of debt financing at a rather vast scale. Just like the impact of tariffs take a while to set in, there are deeper consequences for the AI infrastructure rush. I'm not always clear on who it is an opportunity for. Certainly people are making money off the FOMO in various ways including and mostly, the stock market.
I agree this isn’t great but I think it’s a bit different than the dotcom bubble. There is real usage today vs hypothetical future usage back then.
Great insights, @Guillermo, as always!
Building AI infrastructure and high-performing models requires colossal investments and massive amounts of training data. Yet, it will still take years before many of these products, services, and business models achieve meaningful penetration and profitability.
That’s why the smartest U.S. companies are partnering strategically. It’s like when they built the American railway network before selling a single ticket — yet the wealth, progress, and opportunities that network created were phenomenal.
We’re witnessing something similar today. The real question is: who will compete with them once the network is built and their models are perfected, trained with our data and feedback (thanks to a $25/month subscription)?
At that point, they could charge us, and our companies, whatever they want, as we’ll be fully dependent on them.
I lived in the run up to and well beyond the dotcom crash in 2 markets - ISPs/Datacenters & cyber security. One massive factor was your #7 - mad as it seems a 2meg leased line in 1995 was 200k per year (🤯) & in 2001 it was under 2k. So competitive market price collapse & Moore's Law will be a future factor in the AI bubble. That and all the bad investments not playing out means and investors fleeing as a consequence. 2 major pins hitting the balloon or balloons 🎈🎈
I didnt think about Moore’s law! Super interesting consideration that can change everything for OpenAI
It’s a powerful tool in two specific domains of ML: Computer Vision and NLP. I still think there is more potential with what can be done with deep learning in this fields as they are very fast, but I’d say the flurry of innovation of new tech has slowed. Convolutional Neural Networks and GANs are 8 and 6 years old respectively and no other completely novel and useful architecture has come out since that isn’t something we’ve seen before. At first, combination of CNNs and GANs lead to the Deep Convolutional GAN (DCGAN). We also have Convolutional LSTMS and other hybrid models, but I would argue we’ve found all the good ones and not many talk about that.
People are just applying these models in different ways. Eventually we will take deep learning as far as it can go, and there will be a sobering realization for the super enthusiasts that there are clear limitations given our CURRENT understanding of deep learning.
Outside of computer vision and NLP, deep learning is overhyped. Few realize that deep learning is almost certainly not sufficient for AI, and even fewer want to admit this is true. It’s not suitable for business because classic ML models, in particular Gradient Boosters, frequently outperform deep neural nets, and even if they perform slightly worse, they are MUCH faster to train. In most business contexts where you work with tabular data (often from a database), deep learning is not the correct algorithm for classification or regression.
Finally, in the business world, extracting insights from data and acting on it is most important. Data is the new oil, and businesses are in an arms race to refine it, develop refinement techniques for it, or just simply acquire high quality data and sell it. Deep learning is a poor refinement technique and is expensive. That’s why it’s not popular in the business world as much.
Great one!
Insightful breakdown. Circular financing masks real demand and inflates valuations, classic bubble dynamics. The opportunity is real, but sustainability depends on genuine adoption, not just capital loops.
For more AI trends and practical insights, check out my Substack, where I break down the latest in AI.
This is missing 2 sources of money flow.
1. Investor capital
2. Revenue and expected growth of that
It is not a closed system. Money in -> Money spent -> Revenue out.
Amazon trading at a loss for over 2 decades is the poster child for building value through investing, and turning on profit as growth objectives are achieved.
Deep puncturing of
Delirious bubble
This feels like déjà vu — but with a balance sheet.
In my Leader’s Dispatch: The Robot Uprising Imploded — Here’s Why Humans Are Back on Top, I unpacked how companies like Klarna and IBM learned the hard way that you can’t automate empathy or outsource trust.
https://substack.mark-carroll.com/p/leaders-dispatch-humans-algorithms
Now Guillermo shows us the same drama playing out with money. The AI market’s gone fully meta — funding itself to buy itself so it can sell to itself again.
Turns out, whether you’re replacing people or pricing sanity, circular logic eventually collapses under its own cleverness.