In the same vain, trying to work out if hyperscales/neo-clouds will ever to get an end-state of competaitve advantages or not. Does scale actually make a difference here? With cloud, scale did have some advantages but the customer base was less concentrated (to put it mildly). With A.I inference, the leading labs are massively concentrated buyers. That could put pressure on cost per GPU (once we're not in a GPU rental shortage like the last 6+ months).
Could it lead to the overbuilding (analog) of railroads? Or is it like airlines, where scale and demand didn't create any margins overall for the players as everyone offered a commodity (in this case tokens). Maybe Nvidia wins along (think Boeing/Airbus for airlines).
I've been pegging OpenAI as the Yahoo of the AI era since mid-2025. That might change if they are able to win back more of the enterprise market, but Sam Altman and his team spent wayy too long confusing distribution with strategy.
There’s a similar trap in markets with these kinds of analogies.
Once a narrative proves wildly successful—Amazon, Uber, whatever it is—people start using it as a shortcut for thinking. Losses become “investment,” scale becomes “inevitable,” and time gets treated like a guarantee instead of a variable.
But price has a way of exposing whether the analogy actually fits. Two assets can look identical on the surface—same growth, same story, same excitement—and then behave completely differently when pressure shows up.
The difference is usually hidden in the structure, not the story. And by the time the market makes that distinction obvious, the easy part of the trade is already gone.
Seems like you put the banker / finance guy cap back on? Good for you! A few thoughts:
I think it’s interesting to focus on the AI lab companies and to an extent I do agree with most of what you’ve said. Faith has substituted economics. But, there’s at least some angle of a classic capitalistic defense. Whereas the companies you mentioned who burned in the past mainly lived off network effects, where at scale they can monetize the attention (FB) or find better unit economics at scale via improved logistics / pricing power (Uber, DD, Amazon). The AI lab companies have more a capitalistic approach. For example - you can’t start a competitor to Pilot pens because you’ll never be able to get the same economics. The labs have a similar barrier to entry but instead of being able to provide a lower price, you actually can’t even have a competitive product without enormous amounts of capital.
So the basic bet is - get reliance and raise prices. But this is really only going to work if you embed yourself within businesses and get people replaced by AI. If this happens, when the labs raise prices it’ll be insanely high-friction to re-staff. I think this point is why there’s so much pressure to replace workers - it lets the labs embed deeply into slow-turning enterprises, getting them more pricing power.
Still tbd whether the economics actually work out, I don’t think it looks great at the moment.
But I think what’s gong on the application layer rn is more insane. It’s possibly lower hanging fruit but from early to growth stage the concept of LTV and even what qualifies as ARR and why seems to have been completely thrown out of the window.
Like the amount of unicorns with dogshit LTV and ARR based on usage is copious. It’s like there’s not really any thought going into “where will they be in 5 years” and instead there’s a focus on grabbing land. But nobody is even actually grabbing land, just getting people to take tours. It’s like paying a realtor full lease commissions for every tour they do. Insane.
Now these companies are also raising small dollar values relative to the labs, but it still makes the entire ecosystem feel a little goofy. Of course, it’s always been goofy though as long as I’ve been around which is prob when I was in college early 2020s.
Speaking of my age, I’m one of the (seemingly) few people who was at the right age to watch internet grift culture up first. From hustle culture to dropshipping to course peddlers, silicon valley is using pretty much the same tactic to push AI. The main tactic is to generate insecurity and fear and then offer a solution. Every pinheaded graph I see, argument about ASI soon and BS stories about companies 10x’ing productivity follow this strategy. Which is obviously sketchy as shit and doesn’t give me faith.
These three points are basically future essay drafts I’ve been sitting on because they feel like low hanging fruit tbh.
Sidebar:
Do we have evidence of proprietary data flywheels working for application layer companies? I see this thesis a lot for verticalized companies but haven’t been able to figure out how it actually works in practice if they’re all just sitting on the same model.
"you actually can’t even have a competitive product without enormous amounts of capital", I would push back on this. Deepseek were able to get a state-of -the-art model just a few months later for barely anything (when compared to the U.S labs). If the demand for the leading edge models moves to good enough models with much lower costs, then many companies could offer these models. Anthropic themselves have mentioned how others (Chinese labs) can basically take the latest Anthropic model and make their own condensed version with very limited resources.
On a different angle, other small labs are looking at whole new ways to make models which could, if solved, be far far cheaper to create than current methods. There is no telling that a new method completely will be found which doesn't need billions in compute.
I think the Deepseek model would be illegal in the US / EU given copyright.
I do agree that smaller specialized models could emerge and see teams testing. They could be viable too because it seems like intelligence is proving to be a commodity, so if an external vendor can give a better experience at a lower price on a specialized model, that could work.
But note that even in this case, the larger models will have more pricing power. They could pull a Rockefeller and just burn at the gross margin level until the smaller specialized models run out of capital.
Could be. My view is that A.I companies can (currently) seemingly easily distill state-of-the-art A.I models at a much lower cost, and they often then open-source them. Even Apple is sort of doing this, paying Gemenio a relatveily small amount and then getting permission to make their own models off of Gemeni’s for Apple products.
If 3-4 players stay at the cutting edge, they will likely have pricing competion between them to get/keep share (potentially, like Uber/Lyft or airlines dynamics). Smaller model companies which come afterwards and open-source, as they don’t have the high depreciation cost of creating ever better leading edge models which then only remain somewhat relevant for like 6-12 months, can charge a lot less and be able to run at lower margins maybe.
I don’t see the leading edge models getting to a position where they can take price (pricing power) when most people would likely be satisfied with a following model nearly as good but much much cheaper.
Although it’s very hard to know ahead of time. Why even top VC funds are spreading bets for either scenerio.
Fantastic read.
And a good John D Rockefeller analogy is always appreciated
Super helpful read. Feel like Bolon-Yokte is missing from the god comparators, but doesn’t take away from the point.
An awesome read, thanks.
In the same vain, trying to work out if hyperscales/neo-clouds will ever to get an end-state of competaitve advantages or not. Does scale actually make a difference here? With cloud, scale did have some advantages but the customer base was less concentrated (to put it mildly). With A.I inference, the leading labs are massively concentrated buyers. That could put pressure on cost per GPU (once we're not in a GPU rental shortage like the last 6+ months).
Could it lead to the overbuilding (analog) of railroads? Or is it like airlines, where scale and demand didn't create any margins overall for the players as everyone offered a commodity (in this case tokens). Maybe Nvidia wins along (think Boeing/Airbus for airlines).
such a great ending! so spot on whether they
succeed or fail, there will be value in the infrastructure they leave behind
Thanks Tan! Agreed
I've been pegging OpenAI as the Yahoo of the AI era since mid-2025. That might change if they are able to win back more of the enterprise market, but Sam Altman and his team spent wayy too long confusing distribution with strategy.
There’s a similar trap in markets with these kinds of analogies.
Once a narrative proves wildly successful—Amazon, Uber, whatever it is—people start using it as a shortcut for thinking. Losses become “investment,” scale becomes “inevitable,” and time gets treated like a guarantee instead of a variable.
But price has a way of exposing whether the analogy actually fits. Two assets can look identical on the surface—same growth, same story, same excitement—and then behave completely differently when pressure shows up.
The difference is usually hidden in the structure, not the story. And by the time the market makes that distinction obvious, the easy part of the trade is already gone.
Terrific post! As an economics grad, I love the use of good theory and practical examples. Thanks a lot!
I love when you introduce and consider economics...
I've been in too many organizations that are better than rationalization than economics.... keep at it!
Seems like you put the banker / finance guy cap back on? Good for you! A few thoughts:
I think it’s interesting to focus on the AI lab companies and to an extent I do agree with most of what you’ve said. Faith has substituted economics. But, there’s at least some angle of a classic capitalistic defense. Whereas the companies you mentioned who burned in the past mainly lived off network effects, where at scale they can monetize the attention (FB) or find better unit economics at scale via improved logistics / pricing power (Uber, DD, Amazon). The AI lab companies have more a capitalistic approach. For example - you can’t start a competitor to Pilot pens because you’ll never be able to get the same economics. The labs have a similar barrier to entry but instead of being able to provide a lower price, you actually can’t even have a competitive product without enormous amounts of capital.
So the basic bet is - get reliance and raise prices. But this is really only going to work if you embed yourself within businesses and get people replaced by AI. If this happens, when the labs raise prices it’ll be insanely high-friction to re-staff. I think this point is why there’s so much pressure to replace workers - it lets the labs embed deeply into slow-turning enterprises, getting them more pricing power.
Still tbd whether the economics actually work out, I don’t think it looks great at the moment.
But I think what’s gong on the application layer rn is more insane. It’s possibly lower hanging fruit but from early to growth stage the concept of LTV and even what qualifies as ARR and why seems to have been completely thrown out of the window.
Like the amount of unicorns with dogshit LTV and ARR based on usage is copious. It’s like there’s not really any thought going into “where will they be in 5 years” and instead there’s a focus on grabbing land. But nobody is even actually grabbing land, just getting people to take tours. It’s like paying a realtor full lease commissions for every tour they do. Insane.
Now these companies are also raising small dollar values relative to the labs, but it still makes the entire ecosystem feel a little goofy. Of course, it’s always been goofy though as long as I’ve been around which is prob when I was in college early 2020s.
Speaking of my age, I’m one of the (seemingly) few people who was at the right age to watch internet grift culture up first. From hustle culture to dropshipping to course peddlers, silicon valley is using pretty much the same tactic to push AI. The main tactic is to generate insecurity and fear and then offer a solution. Every pinheaded graph I see, argument about ASI soon and BS stories about companies 10x’ing productivity follow this strategy. Which is obviously sketchy as shit and doesn’t give me faith.
These three points are basically future essay drafts I’ve been sitting on because they feel like low hanging fruit tbh.
Sidebar:
Do we have evidence of proprietary data flywheels working for application layer companies? I see this thesis a lot for verticalized companies but haven’t been able to figure out how it actually works in practice if they’re all just sitting on the same model.
"you actually can’t even have a competitive product without enormous amounts of capital", I would push back on this. Deepseek were able to get a state-of -the-art model just a few months later for barely anything (when compared to the U.S labs). If the demand for the leading edge models moves to good enough models with much lower costs, then many companies could offer these models. Anthropic themselves have mentioned how others (Chinese labs) can basically take the latest Anthropic model and make their own condensed version with very limited resources.
On a different angle, other small labs are looking at whole new ways to make models which could, if solved, be far far cheaper to create than current methods. There is no telling that a new method completely will be found which doesn't need billions in compute.
I think the Deepseek model would be illegal in the US / EU given copyright.
I do agree that smaller specialized models could emerge and see teams testing. They could be viable too because it seems like intelligence is proving to be a commodity, so if an external vendor can give a better experience at a lower price on a specialized model, that could work.
But note that even in this case, the larger models will have more pricing power. They could pull a Rockefeller and just burn at the gross margin level until the smaller specialized models run out of capital.
Could be. My view is that A.I companies can (currently) seemingly easily distill state-of-the-art A.I models at a much lower cost, and they often then open-source them. Even Apple is sort of doing this, paying Gemenio a relatveily small amount and then getting permission to make their own models off of Gemeni’s for Apple products.
If 3-4 players stay at the cutting edge, they will likely have pricing competion between them to get/keep share (potentially, like Uber/Lyft or airlines dynamics). Smaller model companies which come afterwards and open-source, as they don’t have the high depreciation cost of creating ever better leading edge models which then only remain somewhat relevant for like 6-12 months, can charge a lot less and be able to run at lower margins maybe.
I don’t see the leading edge models getting to a position where they can take price (pricing power) when most people would likely be satisfied with a following model nearly as good but much much cheaper.
Although it’s very hard to know ahead of time. Why even top VC funds are spreading bets for either scenerio.
Dude, I think you just schooled yourself.
You wrote an entire column on bad analogies and then used a WeWork analogy?
What is the emoji for milk coming out my nose.
WeWork was a dumpster fire of an idea that I don't think even the founders believed.
Analogies can help to a "point". In tech we went past that point awhile ago!
Amazon isn’t a universal reference point, especially across fundamentally different models.