The Inevitable AI Bubble
AI bubble is inevitable, losing money is not.
In the 1600s, there was the Tulip Bubble.
People were paying the equivalent of a house price to buy a bunch of Tulips.
Anybody who hears this today will ask the same question—how f.cking stupid were they?
But, if you went back in time and asked those people, they would say: No, no, the demand is off the roof.
You go fast forward 330 years, to the late 1990s, and people pouring insane amounts of money into no-revenue internet companies would tell you the same thing: The demand is off the roof.
Nowadays, you raise a question about AI, and you hear the same thing: You don’t understand, the demand is off the roof.
The demand was truly off the roof in the 1660s and 1990s, too, but this didn’t prevent a bubble, did it?
We are going through the same process again today, hearing the same arguments. It’s obvious to anybody who looks a bit closer that there are cracks in the AI trade. The pieces just don’t fit; they don’t make sense.
You just need to look a bit closer and think a bit deeper.
Today, we’ll look closer.
Understanding the AI Investment Landscape
People tend to understand AI investments as if they are something homogeneous, a unit of similar pieces.
It’s not.
Under the hood, we can divide AI investments into three groups:
Infrastructure: Investments in hardware and other necessary inputs like energy to generate the compute required to run AI models. It includes GPUs, servers, miscellaneous data center equipment, energy, and R&D by cloud providers to develop their platform.
Foundational model: Cost of training and running flagship AI models like ChatGPT, Claude, Gemini, Mistral, DeepSeek, etc. Payments to cloud providers are a big part, but investments in R&D and talent are also substantial.
Vertical AI applications: Generally, venture investments in customer-facing AI applications. They use foundational AI models under the hood, but wrap them into a software layer that enhances the value proposition of naked AI calls. This is why they are sometimes called wrappers.
Think of it like a pyramid. At the top of the pyramid, there are wrappers, below them are the foundational models, and infrastructure players are at the bottom.
As the infrastructure layer supports the whole pyramid, the amount of investment required is massive. This has two consequences:
It mandates an oligopolistic market structure.
It keeps the risk capital outside.
Incumbent hyperscalers already have the scale, and they see that the massive investment required creates huge barriers to entry, as it would be impossible for newcomers to reach profitable scale while burning hundreds of billions at the same time, leading to a naturally oligopolistic market. Even neo-clouds like Coreweave and Nebius are largely contractors of incumbent hyperscalers.
As the market structure is guaranteed to be oligopolistic and entry barriers are high, risk capital that needs to be fragmented to work stays largely out of this layer. Thus, this layer is funded almost fully by hyperscaler cash flows. Even neo-clouds themselves are dependent on hyperscalers. This further reinforces the oligopolistic market structure.
The middle layer, foundational models, are themselves applications on a basic level, and also support vertical AI applications. Training and running foundational models is also very costly, though not as much as creating and running the infrastructure. So, barriers to entry are still high, but much lower than the infrastructure layer.
As a result of the moderately high barriers to entry, there is an opening here for risk capital. However, given that the capital requirement is still very high, only the top firms that can afford to throw more than a few billion into a single deal can take part. The rest comes from hyperscalers and foundational model revenues.
The critical consequence of the high capital requirement in the first two layers is that the broader risk capital is pushed toward the top of the pyramid, where the capital requirement is pretty low, as it’s software+API calls to foundational models.
This is why we are seeing that seemingly ridiculous and not-so-special AI apps like AI-not takers can raise multi-million dollar seed investments. Broader risk capital wants to participate in the boom, but they are stuck in the application layer. So, they are investing in whatever they can find.
This is exactly the setup that gives rise to bubbles.
There is massive but fragmented capital that wants to move, but the available opportunities have no way to create long-term value. Yet, capital still flows to whatever it can find due to FOMO. Thus, valuations necessarily rise to the skies as massive capital flows into the smallest layer of the pyramid.
And just like that, we have too much money invested at valuations that are too high and in companies without any prospect for long-term value capture. Just like the dotcom bubble.
Now, let’s turn to the value question and try to explain why there will be no long-term value accrued at the top of the pyramid.
No Value Capture In The Application Layer
Pyramidical value chains work by upstream value addition.
Basically, the bottom of the pyramid supplies the raw input, the layer above it adds some value to it and marks up the price, and this is repeated up until the top of the pyramid. Almost every industry can be reduced to this pyramidical structure.
Think about something very simple, like silver.
Miner extracts silver from the ground and sell it to a refinery. Refinery purifies the silver and turns it into pure solid blocks, marks up the price, and sells it to a silverware maker. Silverware maker works with solid silver blocks, marks up the price, and sells it to end customers.
In this pyramidal value chain, added value accrues upstream:
The cost of extracting the silver is $20 per ounce.
Miner sells to the refinery at $25 per ounce
Refinery sells to silverware maker at $32 per ounce
Silverware is sold at $40 per ounce
The total added value is $20, and the miner captures only 25% of it.
As the added value accrues upstream, there are significant incentives for vertical integration.
Miner can capture more of the value by operating a refinery itself.
These incentives exist in all pyramidical value chains. However, there are also two substantial disincentives that prevent bottom-layer companies from becoming vertically integrated up to the top of the pyramid:
Material difference in products
Added operational hardship
Unless the upstream product is materially different than the downstream and adds significant operational hardship, vertical integration is the obvious strategic choice; it allows firms to capture more value.
Most silver miners don’t have their own silverware brand. You need a refinery, then you should create different plants for silverware manufacturing, you need designers, a broad catalogue, additional certifications, etc. Silver fork is much different than silver ore, and the added operational hardship is significant.
On the other hand, most raw coffee traders also operate their own roasteries and sell their branded coffee. The difference between products is minimal, and the operation is simple, as you just need a roasting machine and packaging.
I call this convergence. In value pyramids, products and operations tend to converge under the same corporate umbrella in the absence of a material difference between products and added operational hardship.
This is exactly why most of the companies in the application layer of AI won’t be able to capture value in the long term, even if they create some.
Most products aren’t substantially different than what the foundational layer companies are doing. In the early days, we saw apps that were creating PowerPoint presentations or Excel files. Now, most foundational models do it internally, and they don’t charge extra money for these tasks.
We have had this argument for software companies literally since the day they emerged. Their products are also not that different from each other in essence; after all, they all bend the code. This is why investors following Buffett’s teachings have avoided software companies, as they saw no moat.
However, the shortage of developers and the high cost of a developer/hour acted as a limiting factor. Because of this, operational hardship was elevated as you needed different teams of developers for different products. If you assigned the same team to create both Stripe and Crowdstrike, you would get neither.
AI removed all those limiting factors from software.
This is why the natural strategic course for big AI labs is to expand to as many big verticals as possible once the gains from foundational model development flatline. There is no material product difference or additional operational hardship to obstruct them.
We are already seeing this.
One of the most profitable verticals now is AI-powered IDEs. Cursor and Windsurf dominated this initially, but it didn’t take two years for AI Labs to enter this. Claude Code is taking market share every day, ChatGPT Codex is gaining traction, and Google just launched Antigravity IDE.
This was, of course, forecasted by everybody, but the selling point for vertical AI was that they would lock in their customers by the time major AI labs launch competing products.
Well, there could be some stickiness, but no lock-in exists.
It’s structurally not possible because the API layer under the hood has no differentiation, and the software layer on top can be copied effortlessly, as I explained. Also, breaking the stickiness is easy for AI labs. They just need to remove the markup on foundational model usage in their own vertical product and keep it for external players using the API. Voila.
This is why most vertical AI companies won’t be able to capture value in the long term. Most valuable verticals will converge under the big AI labs, and profits will be competed away by the retail creators in less valuable verticals because no barriers to entry exist. Big AI labs won’t bother to create AI not takers, but everybody will be able to create their own, so profit margins will also converge to zero.
Result? Dozens of billions invested in the application layer will never see a return on investment.
I am certain that there’ll be some businesses that’ll become quite profitable by locking in users due to network effects, etc, but the number of them will be way fewer than we saw in software because of elevated pressures for convergence.
So, the bubble here is already formed. It’s not hard to see. Demis Hassabis said it openly on Hard Fork YouTube Channel that billion-dollar seed investments in AI are in a bubble.
The bubble is here, and it’ll pop when these companies can’t capture any value and turn to VCs again to raise more capital. By then, most FOMO will have subsided, and VCs will want to see returns on their investment. When they can’t see it, they won’t invest more. And just like that, we’ll have vertical AI companies dying like mayflies.
We hammered down the application layer first, because the bubble is obvious there. This doesn’t mean foundational and infrastructure layers are all good. They have their own problems.
Foundational Models & AI Infrastructure: Inevitable Commodity and Overcapacity
If you look at the techno magazine headlines in the late 1990s, you would see headlines like:
“The Internet Will Enter Every House”
“The Internet Will Change The World”
The Internet Will Change The Economy”
Well, the internet has really entered every house, changed the world, and the whole economy we live in.
Result? The internet has become a commodity. The internet had to become a commodity to reach every household and transform the world and entire economies.
Wholesale internet prices literally collapsed from the late 1990s to 2015:
It’s even cheaper now, with some sources reporting as low as $0.08 per Mbps in key American and European cities with fully developed infrastructure.
The case is obvious: The internet changed the world, but it had to become a commodity first to permeate every aspect of our lives in the first place to achieve this transformation.
This is why telecommunication companies across the world heavily invested in internet infrastructure, laying fiber-optic cables and renting transmission towers literally everywhere. They achieved it in the end. Internet prices collapsed. It became a commodity.
What they skipped was that collapsing internet prices also meant collapsing profits for them. A penny charged above cost meant losing market share. Bubble popped. Most of them went bankrupt. Those that remained resorted to pitiful strategies to make money, like selling phones in connection with data plans with 12+ months commitment, so they could lock in customers and make a bit of money without price competition.
Everybody remembers Cisco, but it has reclaimed its dotcom bubble highs; Vodafone, on the other hand, is still far away. The TMT bubble was immense.
The promise of AI today is similar to that of the Internet. Maximalists say it’s going to permeate every aspect of our lives, powering everything we interact with. For that to happen, foundational model prices should collapse, like the internet.
If you are a normal daily user, you may not feel it, but they are still very expensive. Enterprises feel it. This is despite AI Labs losing money on their API pricing. If they wanted to turn profitable, the prices would be much higher.
We have to deploy massive amounts of infrastructure so prices can drop substantially, and AI will really permeate our lives. Yet, once this happens, AI models will be commoditized, and the opportunity to generate massive returns will be gone.
Can’t they just pace the investment so models can be very profitable? They can’t.
The genie is out of the bottle. If anybody tries to slow it down, other profit-seeking actors will enter the market to address supply constraints. The market mechanism will work. Capitalism will do its thing. There is no way back. They have to build until there is overcapacity, until the prices collapse.
Even the fact that there can always be better chips and so better models is irrelevant. At some point, all of them will be good enough, smart enough (AGI?). Just because something can be better, it doesn’t mean there will be additional economic value.
Think about sidewalks. They can be way better, they can be heated, for instance. But there is no economic value in it. Nobody is okay with paying double the city tax to walk on heated sidewalks.
Same for phone carriers. Once they are good enough, there is no economic value in doing better. A carrier can invest to provide service on top of Mount Everest, but it’s not asked for. If it works in the countryside, it’s good enough, and nobody will pay a premium, so the carrier can maintain service at Everest.
This applies to every service. Don’t think it’ll be different for AI. It applied to intelligence itself. Nobody paid a premium to employ 200 IQ geniuses in call centers; 90 IQs working on minimum wage was good enough. It’ll apply to artificial intelligence, too.
In short, if AI is to fulfill its promise, foundational models are doomed to be commoditized and infrastructure to be overbuilt. Once this happens, ROI for AI labs and infrastructure providers will be minimal because of commoditization and overcapacity.
We are rapidly going toward this point; it’s obvious in the numbers.
Hyperscalers are projected to hold $2.5 trillion in AI assets by the end of 2030.
Assuming 10% WACC and 20% annual depreciation, they need at least 30% cash return on each $ invested in infrastructure just to break even.
This means they will need to add $750 billion in aggregate operating profit by the end of 2030. Their total operating income now is around $450 billion, so they have to nearly triple their operating income in the next 5-6 years.
It’s not impossible, but they’ll only break even then. This is a clear signal that we are going toward a bubble in the infrastructure layer, too.
Foundational models are also getting rapidly commoditized.
We got ChatGPT 5 and Gemini 3 this year; nobody felt a leap like ChatGPT 3.5 to ChatGPT 4. I have been trying Gemini 3 for three weeks now, and it’s good, but I can’t say it is much better than ChatGPT 5.
If you look at the benchmarks, you’ll see the improvement, always. But benchmarks don’t determine anything. Imagine we had a sidewalk benchmark. We could always do better sidewalks, scoring higher in the benchmark. But it wouldn’t be important. We wouldn’t feel the difference after they are good enough. And wouldn’t want to pay extra for sidewalks scoring better in the benchmark.
Maybe a few more generations of models, and we won’t feel the difference anymore. Diminishing returns from model development will intersect with overcapacity, and models will be commoditized; the bubble will pop. AI labs will be pushed to verticals to generate profits, wrappers will go bust, and infrastructure providers will operate like utilities.
It’s not an if question; it’s a matter of time.
Bubbles Are Inevitable, Losing Money Is Not
The fact that there will be a bubble is now clear to almost everybody. I exclude those who are willingly blind to it.
Sundar Pichai himself recently admitted that there are elements of irrationality, euphemism for a bubble:
Nobody can blame them, I explained above; it’s almost like there is nothing that they can do to prevent it. This is what he means by saying “nobody is immune.” The genie is out of the bottle. When high demand meets lightning-fast diffusion, a bubble is almost a structural necessity.
It’s basic economics, supply grows until marginal cost is equal to marginal revenue:
When the demand is excessive and diffusion is so fast, suppliers build capacity as fast as they can to exploit high prices and generate massive profits. But the market participants aren’t in perfect coordination with each other. Most firms underestimate the buildout by their competitors. They are selfish. They want to meet all the demand.
This is why “demand is off the roof” is a bullshit argument to assert there is no bubble. To the exact opposite, skyrocketing demand is a “must” condition for a bubble. Must conditions, high demand & fast diffusion, are present, a bubble is inevitable.
Trump’s Crypto Czar openly admits its inevitability:
If you enter a recession when you slow down any particular investment, that investment is in a bubble. Given that it’ll necessarily slow down, the bubble will necessarily pop. Again, genie out of the bottle.
Firms still investing massive amounts despite the inevitability of a bubble isn’t surprising to me. We understand why they can’t stop. What’s interesting to me is that people are still investing in AI stocks at high valuations despite the main market actors accepting the high possibility of a bubble.
Bubbles are inevitable when the must conditions exist, but losing money is not.
Those who bought Amazon at any point in 1997 were in the money even after the dotcom crash, but those who bought at the top in 1999 didn’t recover until 2009.
If you bought Nebius back at $30, you have nothing to worry about; it was a no-brainer. You’ll likely sit on massive unrealized gains when the bubble peaks; hopefully, you can cash out there. Even if you can’t, you won’t likely lose money. But if it goes to $200 in a week and you buy there, you will likely have some problems.
I see people rely on two things while buying despite elevated valuations:
We are aware of the risk of a bubble, so the risk is under control.
There can’t be a bubble when everybody is afraid of a bubble.
Not true. The same concerns existed at the time of the dotcom bubble. They did it anyway.
Paul Cohn is the founder of Agility Equity Partners. He was a VC in Silicon Valley during the dotcom bubble. Here is what he says about the VC psychology at the time:
People were aware, but they did it anyway. Don’t assume it’s under control. The CEO of a $3 trillion company is openly saying that they may be participating in a bubble. There is no risk control. You can’t rely on any assumption of safeguards. You can only refuse to participate.
It’s not bubbles that kill, it’s chasing bubbles.
🏁 Final Words
Application-layer AI investments are already in a bubble, as a huge amount of capital is stuck in companies that don’t have any way to capture value and generate returns in the long term.
The world we live in has changed, but investors are still stuck in the software era, thinking. Most VCs of today are VCs of the software era. They know one thing: verticalization works.
It worked in the software era because there were limitations preventing convergence. Endless product differentiation, developer shortage, and high cost of development put a lid on the firm size and mandated verticalization. AI removed those limitations. There will be more convergence than they expect. Most companies will go bust.
Tell me why Perplexity needs to exist?
It was a creative idea, but all the foundational models internalized what it does. All the flagship models now have better web search and deep research features than Perpexity. How will Perplexity live? Convergence is inevitable. The bubble will pop.
See it for yourself below. Its app downloads dropped by 80% in a month as they stopped paid marketing.
Why would anybody use the Perplexity app when we have Claude, Gemini, and ChatGPT? What makes it worth $20 billion?
When it comes to the foundational model layer, commoditization is inevitable. That’s the only way AI fulfills its promise. Imagine it. There’ll be smart robots in every house and every business. All the products will internalize AI features. For such a future to happen, AI should be a commodity; it should be as cheap as the internet.
We all know what happened to the largest internet service providers as wholesale prices collapsed. Don’t expect it to be different for AI.
A sufficient level of commoditization requires massive infrastructure buildout. So, infrastructure providers will keep building until that point. The problem is that we’ll be able to see that point only after we pass it, just like it happened with the internet. Meaning they’ll inevitably get into a bubble. No one is immune, as Sundar Pichai said.
The funny thing is that whether there is or will be a bubble is inconsequential if you:
Don’t chase companies that may be in a bubble.
Don’t ever overpay for your positions.
Don’t try timing the market.
If you bought UnitedHealth in March 200, just before the dotcom crash, you would still double your money by December 2001. Someone looking at your portfolio wouldn’t be able to tell there was a crash. But if you had chased Cisco… You know the story.
So, it’s intellectually stimulating to discuss a bubble, but if you are acting like an investor, and not a speculator, it shouldn’t be consequential for you.
Stop here, turn to yourself, and think for a moment: If there is a bubble, will it be consequential for you?
That’ll tell you what to do next.
New recent sub here. Really enjoyed this one -- please keep them coming!
Can we assume this bubble implosion will be mostly contained to the private market (e.g. VC world) and the narrow set of application layer public companies? I can’t imagine a sudden brutal crash for Microsoft, Amazon, Google, etc. At worst a slow burn? They will most likely continue to invest in the next 2/3 years, supporting an infrastructure ecosystem (eg, energy, grid players). It is interesting that OpenAi is planning its IPO next year, it could already be too late for its investors to monetise.