What about the biotech industry? I would tend to believe that AI models, autonomous wet labs and improving clinical trial success rates would lead to biotech breakthroughs and improved FDA appoval cycle times...
Back when Berkshire-Hathaway was still a textile company, someone advised Buffett that that business was becoming more attractive because due to the greatly improved productivity of the new equipment that was available. His response was that the same equipment was available to competitors, thus the efficiency advantages would be competed away and would flow through to the customers.
If it had been possible to use the new equipment to creatively transform the business model...shorter runs to reduce inventory needs and respond to market demand more quickly, for example...then it could have been used to generate a sustainable competitive advantage...I don't know enough about the industry to know if this would have been feasible or not, but of course in the event Buffett did better turning his attention to other things.
The analogy with AI applications is prettyh straightforward
Great example, but I think it's the wrong comparison.
Traditional textile companies take commodities like cotton as input.
But what if you are a hypothetical textile company that produces its own custom yarn that no one can replicate? Then wouldn't the efficiency advantages only accrue to your business?
Now replace the custom yarn with proprietary data and see where we are going? If you are sitting in a mountain of data that no one else can easily built or access to, wouldn't AI make it much more valuable?
This is one of the more compelling second-order AI theses I have read. The key insight is that a technology can transform the economy without allowing the companies financing its infrastructure to capture most of the eventual value. Railroads lowered the cost of moving goods; the internet lowered the cost of distributing information; AI may lower the cost of producing intelligence. In each case, the more interesting question is not simply who builds the infrastructure, but who can turn the resulting abundance into pricing power, customer retention, lower operating costs, or entirely new demand. I would add one important qualification, though: cheaper intelligence does not automatically become durable margin. If every competitor gains access to the same models at roughly the same price, much of the productivity benefit may be competed away and ultimately passed on to customers. The real winners will therefore need more than “AI exposure.” They need a mechanism that prevents the value from leaking out—proprietary data, distribution, regulatory trust, workflow ownership, switching costs, or a system-of-record position that customers cannot easily replace. That is why your Chegg example is so useful: owning a large body of information was not enough because the customer relationship and workflow were not defensible. I also like your emphasis on measuring usage, pricing, and margins rather than accepting management’s AI vocabulary. A rising AI-related ARR figure sounds impressive, but the harder questions are whether customers are paying more, whether retention is improving, whether revenue per employee is rising, and whether those gains survive once competitors adopt similar tools. The historical comparison with railroads and telecom is powerful, although I would be careful not to make the infrastructure-versus-application distinction too absolute. Some infrastructure providers may retain extraordinary economics when capacity is scarce, technological leadership is difficult to reproduce, or switching costs are high. The likely outcome may not be a clean handoff from chips to software, but a gradual redistribution of profit pools as compute becomes less scarce and implementation becomes more valuable. To me, that is the strongest version of your argument: the next AI trade is not really “everything that isn’t AI.” It is companies for which AI becomes an invisible input rather than the product itself—businesses that buy increasingly cheap intelligence and sell something differentiated, trusted, and difficult to replace. The market has spent years asking who will provide the intelligence. The more important question now may be who can keep the economic value it creates.
“I think of the AI case for… an MDB as being meaningful, but I don't see the case for exponential growth. For that kind of growth, we need to find a buyer of AI who is able to do "1+1=5" kind of math.
Here are the characteristics I'm looking for:
1) existing set of customer relationships: distribution is usually the hardest part of the journey so let's identify companies that have that box checked;
2) access to tons of proprietary data that can be transformed into valuable information by AI: this is the heart of the transaction: I am looking for companies that can buy AI technology and use it in a transformational way, to spin straw into gold; [why "proprietary data"? Because if it's not proprietary, then anyone can do it, so there's no moat]
3) a track record of internal innovation: Building new products and new product lines is hard, and most large companies have forgotten how to do it. They can only do bolt-on acquisitions, which will really limit their ability to respond to this huge opportunity.”
Today I’d argue your list of qualities you’re looking for is necessary but insufficient. Yes, Autodesk has great market position — but can they make a big leap in revenue with AI?
I’m not arguing I nailed it: NOW, SNOW, AMZN, and CRWD are doing well with AI. But my other enterprise software plays, HUBS and INTU, are still proving out.
With HUBS, I thought they could capture the opportunity faster; they still may get there.
With INTU, I was so focused on the Quickbooks opportunity that I neglected to consider the impact on their TurboTax business. I’m hanging on but losing faith in management’s ability to capture the opportunity.
As you say, the bulk of the value creation in a new technology wave belongs to the buyer. Which companies will win? TBD
I completely agree with your take. Distribution, in particular, seems to be underestimated by investors. But if we look at the expenses of any software firm, we will see Sales & Marketing as the largest expense line item.
The companies we stated above are for reference; we will do deep dives on these names to get a better sense of who is likely to win and whether they are worth a slot in our portfolio.
The inability to distinguish causality from correlation is widely recognized in computer science and cognitive science as one of the fundamental limitations of modern artificial intelligence.
This is the part of the AI trade I think gets really interesting.
Eventually the story has to move beyond who sells the chips and compute to what people actually build with all of this intelligence once it becomes cheap enough.
We’re seeing that firsthand at Trader’s Edge.
We’ve spent years developing the actual trading methodology and education around structure, levels, momentum, volatility and risk. Now we’re using AI and software to turn more of that knowledge into something traders can actually use in real time.
Apollo ONE is probably the best example. Instead of giving someone another indicator to interpret, we’re trying to put more of the decision-making process directly on the chart. Is a setup developing? Is momentum confirming it? Where is it invalidated? Where are the targets? When should you just stand down?
That’s where I think a lot of the real AI value eventually gets created. Not AI for the sake of saying you use AI, but taking actual domain expertise and making it dramatically more useful and scalable.
If anyone wants to try Apollo ONE, send me your TradingView username and I’ll add you.
yeah. its not like people are going to stop eating or getting dressed or learning saxophone. Life will go on at least for a little while even if our godchildren roam the galaxy without us.
It’s a great reminder that railroads captured only a sliver of the GDP they enabled, so todays AI infrastructure boom might reward the industries built on top of it more than AI itself.
So what? The whole point of an enabling technology is to allow the rest of the economy and it's GDP to grow which it did. BTW a growing pie rewards everyone, and the railways got their share. In the early 1900s they were a majority of the stock market. If you look at market capitalization instead of GDP, they did fine.
BTW railroad equipment companies (analogous to semiconductor companies now) did very well -
"... but the index was more or less in the same place after the hype was over."
Cherry picking just the telecom subindex in the Nasdaq doesn't prove the point. Who invested only in telecom during dotcom? Unfortunately, I don't have ready access to long enough Nasdaq data but even the oldest Nasdaq investable mutual fund proves this to be false -
Even at the bottom you would have made 2x the money if you invested at the beginning. What's more, you would have not been that much worse off than the S&P 500. I will bet it looks better with earlier data.
We do have better data if we look at Semiconductors (a much bigger portion of the dotcom boom/bubble) -
As long as you invested before mid 1997 you would not have lost money at the bottom. In fact, if you were a smart insider in the industry and you invested in 1990 when the combination of modern CPUs (32 bit 386 onwards), modern OS/GUI (Windows 3.0) became cheap enough, you would have made more than twice the money in semiconductors as the S&P 500 at the bottom -
The whole point is that you have to be early enough on a trend and not join when every Tom, Dick and Harry is running after the same thing.
Edit: Unfortunately, you will need a subscription to see the results of those links going back to the 90s, but if you have another way of seeing the same data, you will see that it's true.
It is difficult to tell today if we are early in the AI trade or not because every technology at the inflection point tends to accelerate more than the previous one.
What is worrysome is the amount of debt generated in a very short period of time.
Per Ed Zitron, the everything AI hater (I have a subscription and he is a good read):
"The Debt Markets Have Never Given This Much Money To The Tech Industry
Fun fact! Between 2008 and 2022, Microsoft, Google, Amazon, Meta, and Oracle had issued around $305.6 billion in bonds.
Between 2023 and 2026 so far, they’ve issued around $374.5 billion.
These numbers don’t include NVIDIA and SpaceX’s $25 billion bond sales, and are dominated by the $80.8 billion raised by Google and $98 billion raised by Amazon, with Meta and Oracle raising $25 billion apiece. These numbers also include bond sales in other currencies — for example, Google raised $5.5 billion Australian dollars, with the 20-year tranche pricing at a monstrous 6.98% rate, a full 180bps above the benchmark.
It also doesn’t include the $38 billion raised by Oracle to build data centers in Wisconsin and Texas for OpenAI, nor does it include any of the various data center SPVs like Meta’s $27.3 billion Beignet Investor vehicle for its Hyperion data center in Louisiana or its $12.3 billion Sopaipilla Investor LLC SPV.
The fun part about the last two of those is that they don’t appear on Meta’s balance sheet as a result of some of the richest accounting shenanigans in history."
Even if AI is a commodity, it cannot run on rocks. You need chips, memory, networking, data centers and power to run it. It's but natural that semiconductor and related stocks are doing well. Even if the labs and hyperscalers design their own chips, you still need it to be fabricated which is only possible with 2-3 companies. And then memory, which is only possible with 3-4 companies. Commoditized AI or not, the other industries will benefit via productivity eventually. In the meantime, whoever enables the infrastructure to be built should benefit.
As for debt, even the Railway boom required a lot of debt. Sure, some of it was wasted but the railway infrastructure was an immense productivity boom. When it came to electricity, JP Morgan himself got into the game to raise the capital needed for what he thought was a very important enabling technology. I'm not saying I have proof that AI is like that, rather that the use of debt doesn't prove anything. It can be good or bad. It's just a tool in the financing tookit.
The distinction between owning the technology and capturing its economic value is important. We see something similar in systematic trading: cheaper, more capable AI models don’t automatically create a trading edge. Proprietary data, execution quality, and risk management ultimately determine whether that intelligence translates into sustainable returns.
Interesting view instead of the just reheated AI will kill humanity, and the gaslighting and manipulation by the usual AI scammers in charge.
It sounds more like AI will kill a lot of portfolios.
Fun timing, Jay Cooke & Co went under 153 years ago today financing the Northern Pacific. That kicked off the Panic of 1873.
What about the biotech industry? I would tend to believe that AI models, autonomous wet labs and improving clinical trial success rates would lead to biotech breakthroughs and improved FDA appoval cycle times...
Good point. But Biotech produces a lot of binary outcomes, which we are not comfortable with from a portfolio risk perspective.
Back when Berkshire-Hathaway was still a textile company, someone advised Buffett that that business was becoming more attractive because due to the greatly improved productivity of the new equipment that was available. His response was that the same equipment was available to competitors, thus the efficiency advantages would be competed away and would flow through to the customers.
If it had been possible to use the new equipment to creatively transform the business model...shorter runs to reduce inventory needs and respond to market demand more quickly, for example...then it could have been used to generate a sustainable competitive advantage...I don't know enough about the industry to know if this would have been feasible or not, but of course in the event Buffett did better turning his attention to other things.
The analogy with AI applications is prettyh straightforward
Great example, but I think it's the wrong comparison.
Traditional textile companies take commodities like cotton as input.
But what if you are a hypothetical textile company that produces its own custom yarn that no one can replicate? Then wouldn't the efficiency advantages only accrue to your business?
Now replace the custom yarn with proprietary data and see where we are going? If you are sitting in a mountain of data that no one else can easily built or access to, wouldn't AI make it much more valuable?
This is one of the more compelling second-order AI theses I have read. The key insight is that a technology can transform the economy without allowing the companies financing its infrastructure to capture most of the eventual value. Railroads lowered the cost of moving goods; the internet lowered the cost of distributing information; AI may lower the cost of producing intelligence. In each case, the more interesting question is not simply who builds the infrastructure, but who can turn the resulting abundance into pricing power, customer retention, lower operating costs, or entirely new demand. I would add one important qualification, though: cheaper intelligence does not automatically become durable margin. If every competitor gains access to the same models at roughly the same price, much of the productivity benefit may be competed away and ultimately passed on to customers. The real winners will therefore need more than “AI exposure.” They need a mechanism that prevents the value from leaking out—proprietary data, distribution, regulatory trust, workflow ownership, switching costs, or a system-of-record position that customers cannot easily replace. That is why your Chegg example is so useful: owning a large body of information was not enough because the customer relationship and workflow were not defensible. I also like your emphasis on measuring usage, pricing, and margins rather than accepting management’s AI vocabulary. A rising AI-related ARR figure sounds impressive, but the harder questions are whether customers are paying more, whether retention is improving, whether revenue per employee is rising, and whether those gains survive once competitors adopt similar tools. The historical comparison with railroads and telecom is powerful, although I would be careful not to make the infrastructure-versus-application distinction too absolute. Some infrastructure providers may retain extraordinary economics when capacity is scarce, technological leadership is difficult to reproduce, or switching costs are high. The likely outcome may not be a clean handoff from chips to software, but a gradual redistribution of profit pools as compute becomes less scarce and implementation becomes more valuable. To me, that is the strongest version of your argument: the next AI trade is not really “everything that isn’t AI.” It is companies for which AI becomes an invisible input rather than the product itself—businesses that buy increasingly cheap intelligence and sell something differentiated, trusted, and difficult to replace. The market has spent years asking who will provide the intelligence. The more important question now may be who can keep the economic value it creates.
Here’s what I wrote in mid-2023:
“I think of the AI case for… an MDB as being meaningful, but I don't see the case for exponential growth. For that kind of growth, we need to find a buyer of AI who is able to do "1+1=5" kind of math.
Here are the characteristics I'm looking for:
1) existing set of customer relationships: distribution is usually the hardest part of the journey so let's identify companies that have that box checked;
2) access to tons of proprietary data that can be transformed into valuable information by AI: this is the heart of the transaction: I am looking for companies that can buy AI technology and use it in a transformational way, to spin straw into gold; [why "proprietary data"? Because if it's not proprietary, then anyone can do it, so there's no moat]
3) a track record of internal innovation: Building new products and new product lines is hard, and most large companies have forgotten how to do it. They can only do bolt-on acquisitions, which will really limit their ability to respond to this huge opportunity.”
Today I’d argue your list of qualities you’re looking for is necessary but insufficient. Yes, Autodesk has great market position — but can they make a big leap in revenue with AI?
I’m not arguing I nailed it: NOW, SNOW, AMZN, and CRWD are doing well with AI. But my other enterprise software plays, HUBS and INTU, are still proving out.
With HUBS, I thought they could capture the opportunity faster; they still may get there.
With INTU, I was so focused on the Quickbooks opportunity that I neglected to consider the impact on their TurboTax business. I’m hanging on but losing faith in management’s ability to capture the opportunity.
As you say, the bulk of the value creation in a new technology wave belongs to the buyer. Which companies will win? TBD
Thanks for the comment.
I completely agree with your take. Distribution, in particular, seems to be underestimated by investors. But if we look at the expenses of any software firm, we will see Sales & Marketing as the largest expense line item.
The companies we stated above are for reference; we will do deep dives on these names to get a better sense of who is likely to win and whether they are worth a slot in our portfolio.
Maybe pharma.. LLY is jumping into the AI game, using their proprietary data, with vigor.
The inability to distinguish causality from correlation is widely recognized in computer science and cognitive science as one of the fundamental limitations of modern artificial intelligence.
This is the part of the AI trade I think gets really interesting.
Eventually the story has to move beyond who sells the chips and compute to what people actually build with all of this intelligence once it becomes cheap enough.
We’re seeing that firsthand at Trader’s Edge.
We’ve spent years developing the actual trading methodology and education around structure, levels, momentum, volatility and risk. Now we’re using AI and software to turn more of that knowledge into something traders can actually use in real time.
Apollo ONE is probably the best example. Instead of giving someone another indicator to interpret, we’re trying to put more of the decision-making process directly on the chart. Is a setup developing? Is momentum confirming it? Where is it invalidated? Where are the targets? When should you just stand down?
That’s where I think a lot of the real AI value eventually gets created. Not AI for the sake of saying you use AI, but taking actual domain expertise and making it dramatically more useful and scalable.
If anyone wants to try Apollo ONE, send me your TradingView username and I’ll add you.
https://www.tradingview.com/script/bVJ0QJAl-APOLLO-ONE-By-Trader-s-Edge/
Keep AI off my hard drive. I don’t want any part of it.
😂
When AI infrastructure is consensus, the next margin story is often what cheap intelligence enables elsewhere.
Keep AI out of my hard drive.
Wait for on device inference!
I won’t want that either.
yeah. its not like people are going to stop eating or getting dressed or learning saxophone. Life will go on at least for a little while even if our godchildren roam the galaxy without us.
Picking future winners? Hmmm. I would rather short the present winners ;)
It’s a great reminder that railroads captured only a sliver of the GDP they enabled, so todays AI infrastructure boom might reward the industries built on top of it more than AI itself.
So what? The whole point of an enabling technology is to allow the rest of the economy and it's GDP to grow which it did. BTW a growing pie rewards everyone, and the railways got their share. In the early 1900s they were a majority of the stock market. If you look at market capitalization instead of GDP, they did fine.
BTW railroad equipment companies (analogous to semiconductor companies now) did very well -
https://x.com/Jesse_Livermore/status/2065516938156249307
I'm sorry but the below is not quite true -
"... but the index was more or less in the same place after the hype was over."
Cherry picking just the telecom subindex in the Nasdaq doesn't prove the point. Who invested only in telecom during dotcom? Unfortunately, I don't have ready access to long enough Nasdaq data but even the oldest Nasdaq investable mutual fund proves this to be false -
https://legacy.portfoliovisualizer.com/fund-performance?s=y&symbol=RYOCX
Even at the bottom you would have made 2x the money if you invested at the beginning. What's more, you would have not been that much worse off than the S&P 500. I will bet it looks better with earlier data.
We do have better data if we look at Semiconductors (a much bigger portion of the dotcom boom/bubble) -
https://legacy.portfoliovisualizer.com/fund-performance?s=y&symbol=FSELX
As long as you invested before mid 1997 you would not have lost money at the bottom. In fact, if you were a smart insider in the industry and you invested in 1990 when the combination of modern CPUs (32 bit 386 onwards), modern OS/GUI (Windows 3.0) became cheap enough, you would have made more than twice the money in semiconductors as the S&P 500 at the bottom -
https://www.portfoliovisualizer.com/fund-performance?s=y&sl=31ZnLMm14UTHC1qAlPz2EI
The whole point is that you have to be early enough on a trend and not join when every Tom, Dick and Harry is running after the same thing.
Edit: Unfortunately, you will need a subscription to see the results of those links going back to the 90s, but if you have another way of seeing the same data, you will see that it's true.
It is difficult to tell today if we are early in the AI trade or not because every technology at the inflection point tends to accelerate more than the previous one.
What is worrysome is the amount of debt generated in a very short period of time.
Per Ed Zitron, the everything AI hater (I have a subscription and he is a good read):
"The Debt Markets Have Never Given This Much Money To The Tech Industry
Fun fact! Between 2008 and 2022, Microsoft, Google, Amazon, Meta, and Oracle had issued around $305.6 billion in bonds.
Between 2023 and 2026 so far, they’ve issued around $374.5 billion.
These numbers don’t include NVIDIA and SpaceX’s $25 billion bond sales, and are dominated by the $80.8 billion raised by Google and $98 billion raised by Amazon, with Meta and Oracle raising $25 billion apiece. These numbers also include bond sales in other currencies — for example, Google raised $5.5 billion Australian dollars, with the 20-year tranche pricing at a monstrous 6.98% rate, a full 180bps above the benchmark.
It also doesn’t include the $38 billion raised by Oracle to build data centers in Wisconsin and Texas for OpenAI, nor does it include any of the various data center SPVs like Meta’s $27.3 billion Beignet Investor vehicle for its Hyperion data center in Louisiana or its $12.3 billion Sopaipilla Investor LLC SPV.
The fun part about the last two of those is that they don’t appear on Meta’s balance sheet as a result of some of the richest accounting shenanigans in history."
If AI will eventually become a commodity......
Even if AI is a commodity, it cannot run on rocks. You need chips, memory, networking, data centers and power to run it. It's but natural that semiconductor and related stocks are doing well. Even if the labs and hyperscalers design their own chips, you still need it to be fabricated which is only possible with 2-3 companies. And then memory, which is only possible with 3-4 companies. Commoditized AI or not, the other industries will benefit via productivity eventually. In the meantime, whoever enables the infrastructure to be built should benefit.
As for debt, even the Railway boom required a lot of debt. Sure, some of it was wasted but the railway infrastructure was an immense productivity boom. When it came to electricity, JP Morgan himself got into the game to raise the capital needed for what he thought was a very important enabling technology. I'm not saying I have proof that AI is like that, rather that the use of debt doesn't prove anything. It can be good or bad. It's just a tool in the financing tookit.
To quote The Joker:
"I am not a monster, I am just ahead of the curve!"
Another great article. Quick question. Would it be possible for you to update the great study about crypto DCA from November 20th, 2024?
The distinction between owning the technology and capturing its economic value is important. We see something similar in systematic trading: cheaper, more capable AI models don’t automatically create a trading edge. Proprietary data, execution quality, and risk management ultimately determine whether that intelligence translates into sustainable returns.
How to Position for Copper Multi-Decade Demand Silver delivered 144% gains to subscribers who positioned early.
https://paretoinvestor.substack.com/p/copper-crisis-2026-supply-shortage