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.
Yes, I would say it is mostly contained in the application layer right now, so private markets are exposed the most. I don't think there is currently a bubble in the infrastructure layer, as we are still capacity-constrained, but we are going toward that point. I don't think there will be a crash in the top names. What I think will happen is that they won't likely see sufficient ROI, so they'll reduce investment over time. But there will be a ripple effect because of this, as many smaller AI plays in the public market are currently dependent on their spending.. In short, we'll likely see a market pullback as they reduce investment due to ripple effects, but I don't think it'll be a massive crash. Not at this spending level at least.
Commentators also compare it to the shale boom and its bust. A long period of disappointing returns, capex cuts, and grinding underperformance in the more marginal players (overall contained).
As long as there is more and more demand for AI use through the cloud why would the infrastructure level not have good ROI? If Amazon can serve a free model from Deep Seek or Baba how is that bad for them? It's just bad for companies like OpenAI and Anthropic.
How is serving AI in the cloud different from non-AI cloud from AWS, Azure...? Non-AI cloud only has 3 big players ( AWS, Azure, and GCP ). So the hyper scalers have a scale advantage and not anyone can start a cloud and match their prices.
Why wouldn't AI cloud profits and revenues keep growing just like non-AI cloud revenues kept growing at a high rate for many years.
In fact as foundation model prices per million token keep dropping the demand for AI through cloud would keep increasing. In some business models where you serve millions of customers the business model only works if the token prices are below a certain level.
If hyper scalers could keep their gross margins with non-AI cloud, they can keep them with AI cloud. Lower prices for foundation models through commoditization is even better for a cloud like AWS. Then AWS doesn't have to pay big fees to Open AI or Anthropic. They can keep a bigger share of the revenue.
Company use for AI just grows and grows. And even if the foundation model is open weight and free, there is still the price to pay to the AI cloud provider for serving the model.
Thoughtful case that LLM AI is a “Bubble”, essentially because the internet-trained LLMs are interchangeable commodities. Does LLM AI offer anything more than more efficient search of existing public or private data? If not, there’s nothing especially creative about LLMs, which might be why nobody had been able to articulate a return of, much less on, the expected $5 trillion being invested in compute infrastructure (excluding new power generation). Think about LLMs as a mechanism for translating code to conventional languages.
Real AI will involve cognitive applications that add real-world context, choice and cumulative memory in specific applications. @Yann LeCun “World AI”: systems that build internal simulations of how the world works using video, spatial, and other sensor data, enabling prediction, planning, and decision making in real environments, not just text. That’s what’s coming.
Definitely, current LLM technology is just next-toke-prediction.
You feel no light of real intelligence.
You know, children can’t do much, they don’t know much about the world but sometimes they make something that you “exactly, this is how you think about it.” LLM’s have never made me say this.
Agreed, well said. LLM's as they exist today have sucked the text world dry. There are no more meaningful data or advantages to be had. World AI here we come.
My take-away on this article is how Convergence is inevitable for AI industry. Perplexity is a wrapper, so it surely will cease to exits as its absolute reliance on ChatGPT, Gemini and other platforms nullifies any differentiation. It will exasperate further as differentiation between those platforms will dissipate, i.e. Anthropic will eventually broaden out of enterprise AI for writing codes to cover Retail market for creating images just like Gemini, and so forth. The catalyst for fundamental shift in the market trend will probably rise on the event of Open AI’s financial demise.
If the API layer is similar and the software layer can be replicated, then any attractive vertical gets arbitraged by the model providers. That flips the usual vertical SaaS playbook, because what made the software shine is exactly what AI eventually erodes.
If AI actually lives up to the “everything, everywhere” promise, then investors should worry less about an AI bubble and more about where can the play survive.
New recent sub here. Really enjoyed this one -- please keep them coming!
Thank you so much Raymond. Glad you liked it!
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.
Yes, I would say it is mostly contained in the application layer right now, so private markets are exposed the most. I don't think there is currently a bubble in the infrastructure layer, as we are still capacity-constrained, but we are going toward that point. I don't think there will be a crash in the top names. What I think will happen is that they won't likely see sufficient ROI, so they'll reduce investment over time. But there will be a ripple effect because of this, as many smaller AI plays in the public market are currently dependent on their spending.. In short, we'll likely see a market pullback as they reduce investment due to ripple effects, but I don't think it'll be a massive crash. Not at this spending level at least.
Commentators also compare it to the shale boom and its bust. A long period of disappointing returns, capex cuts, and grinding underperformance in the more marginal players (overall contained).
Yeah I expect something like that rather than a dotcom style rapid collapse
But then also big tech will bleed and lose ground (FCF, margins, ROI)
As long as there is more and more demand for AI use through the cloud why would the infrastructure level not have good ROI? If Amazon can serve a free model from Deep Seek or Baba how is that bad for them? It's just bad for companies like OpenAI and Anthropic.
How is serving AI in the cloud different from non-AI cloud from AWS, Azure...? Non-AI cloud only has 3 big players ( AWS, Azure, and GCP ). So the hyper scalers have a scale advantage and not anyone can start a cloud and match their prices.
Why wouldn't AI cloud profits and revenues keep growing just like non-AI cloud revenues kept growing at a high rate for many years.
In fact as foundation model prices per million token keep dropping the demand for AI through cloud would keep increasing. In some business models where you serve millions of customers the business model only works if the token prices are below a certain level.
If hyper scalers could keep their gross margins with non-AI cloud, they can keep them with AI cloud. Lower prices for foundation models through commoditization is even better for a cloud like AWS. Then AWS doesn't have to pay big fees to Open AI or Anthropic. They can keep a bigger share of the revenue.
Company use for AI just grows and grows. And even if the foundation model is open weight and free, there is still the price to pay to the AI cloud provider for serving the model.
Very good work. Thanks. Some of the top tier companies will be acquired by the bottom tier companies, but that won't stop the bubble from bursting.
Exactly, convergence will take place through acquisitions too. Very nice point. Thanks for the contribution!
It's always a pleasure to reas you
Thank you for the kind words David!
Thoughtful case that LLM AI is a “Bubble”, essentially because the internet-trained LLMs are interchangeable commodities. Does LLM AI offer anything more than more efficient search of existing public or private data? If not, there’s nothing especially creative about LLMs, which might be why nobody had been able to articulate a return of, much less on, the expected $5 trillion being invested in compute infrastructure (excluding new power generation). Think about LLMs as a mechanism for translating code to conventional languages.
Real AI will involve cognitive applications that add real-world context, choice and cumulative memory in specific applications. @Yann LeCun “World AI”: systems that build internal simulations of how the world works using video, spatial, and other sensor data, enabling prediction, planning, and decision making in real environments, not just text. That’s what’s coming.
Definitely, current LLM technology is just next-toke-prediction.
You feel no light of real intelligence.
You know, children can’t do much, they don’t know much about the world but sometimes they make something that you “exactly, this is how you think about it.” LLM’s have never made me say this.
Agreed, well said. LLM's as they exist today have sucked the text world dry. There are no more meaningful data or advantages to be had. World AI here we come.
Great text as usual!
Thank you Matthew! Glad you liked it.
Really encapsulated much of my thoughts, kudos. It’s really down to a matter of when, not if, for the AI bubble to pop
Glad you liked it!
My take-away on this article is how Convergence is inevitable for AI industry. Perplexity is a wrapper, so it surely will cease to exits as its absolute reliance on ChatGPT, Gemini and other platforms nullifies any differentiation. It will exasperate further as differentiation between those platforms will dissipate, i.e. Anthropic will eventually broaden out of enterprise AI for writing codes to cover Retail market for creating images just like Gemini, and so forth. The catalyst for fundamental shift in the market trend will probably rise on the event of Open AI’s financial demise.
Totally, if I were to provide one key idea from this, it would be the inevitability of convergence.
If the API layer is similar and the software layer can be replicated, then any attractive vertical gets arbitraged by the model providers. That flips the usual vertical SaaS playbook, because what made the software shine is exactly what AI eventually erodes.
If AI actually lives up to the “everything, everywhere” promise, then investors should worry less about an AI bubble and more about where can the play survive.
But demand was not through the roof with the internet. Not in any way shape or form.
This doesn't make a good case for a bubble. There's zero exposure to the biggest companies in the public market.
I've still yet to see a well-reasoned analysis supporting a bubble.
Where would you place chip manufacturers? At the very bottom like the prices for Mbit? I’m missing the shovel makers – power/energy providers.