My AI investment thesis is that every AI application startup is...
Yishan@yishan
17 views
Nov 10, 2025
~4 min read
Advertisement
1
My AI investment thesis is that every AI application startup is likely to be crushed by rapid expansion of the foundational model providers.
App functionality will be added to the foundational models' offerings, because the big players aren't slow incumbents (it is wrong to apply the analogy of "fast startup, slow incumbent" here), they are just big. Far more so than with any other prior new technology, there is a massive and fast-moving wave that obsoletes every new app almost as fast as it can be invented. There is almost no time to build a company and scale it.
There are two ways AI application startup founders can make money:
- Make a flash-in-the-pan app that generates a ton of cash and bank the cash (my estimate is that you have about 12-18 months cashflow generation)
- Make a good enough app that you get acquired by one of the big players for sufficient equity
The situation is highly unstable - we don't know if it's going to crash or go to the moon but both scenarios make it very unlikely that any AI application startup will independently become a generational supercompany (baseline odds are low to begin with).
The best odds are finding an application niche in a highly specialized field with extremely unique and specific data barriers, ideally ones relating to real atoms (hardware or world-related) data and not software/finance.
App functionality will be added to the foundational models' offerings, because the big players aren't slow incumbents (it is wrong to apply the analogy of "fast startup, slow incumbent" here), they are just big. Far more so than with any other prior new technology, there is a massive and fast-moving wave that obsoletes every new app almost as fast as it can be invented. There is almost no time to build a company and scale it.
There are two ways AI application startup founders can make money:
- Make a flash-in-the-pan app that generates a ton of cash and bank the cash (my estimate is that you have about 12-18 months cashflow generation)
- Make a good enough app that you get acquired by one of the big players for sufficient equity
The situation is highly unstable - we don't know if it's going to crash or go to the moon but both scenarios make it very unlikely that any AI application startup will independently become a generational supercompany (baseline odds are low to begin with).
The best odds are finding an application niche in a highly specialized field with extremely unique and specific data barriers, ideally ones relating to real atoms (hardware or world-related) data and not software/finance.
2
Great, this is blowing up so I will offer some additional follow-up:
This is NOT your typical prediction of "the incumbents are agile" or the old "what if Google clones your startup" midwit investor question.
The entire novelty of this thesis is that unlike in the past, specific elements of the AI industry are likely to make it so that application companies cannot outrun the wave of obsolescence, which will rush along far, far more quickly than prior technology waves.
The foundational technology has not stabilized in any way whatsoever, and applications require a sufficiently stable foundation for some extended period of time in order to create value and then a system for monetizing that value (i.e. "a business"). The wholesale rate of change in the nature of the foundation is the reason why I think almost all application startups will not survive to achieve any significant scale, not because the current large players are special.
Most companies don't survive sea changes in the business-technological environment. But these sea changes happen slowly enough that one can build businesses in between. PC, desktop internet, mobile internet, etc, all took many years to play out, and were spaced out enough for application companies to grow, mature, and become incumbents themselves. As a baseline, most startups don't survive during a rapid period of change either. The small minority of incumbents who survive need extreme agility and enough of a stable footing in the last epoch (i.e. a revenue base that doesn't dissolve too quickly) to fund their evolution.
Moreover, it's usually new startups that drive the disruption that challenges incumbents. This is not the case with AI. In this case, the largest players are the ones continually causing the sea change. The environment is so continuously roiled that there is no stable foundation for application startups to become established before the next wave overtakes them. I'm not talking about incumbents outcompeting them, I'm talking about the landscape changing to make them obsolete.
From a practical investment lens, the way to apply this thesis to an AI application startup is to ask: are the fundamental assumptions underpinning this startup's existence going to be the same in five years? Or will they be unpredictably different? The key here is predictability - if the future will be radically different but you can predict it with confidence, you can pre-position your business. But that's not the case right now in AI. You can't skate to where the puck is going if all you know for sure is that 20 people are going to slap the puck in some crazy direction at extremely high velocity.
Sea changes are now happening on a 9-12 month cycle. Very few startups can turn into a mature business in that timeframe - and by mature, I mean having all the boring stuff like sales relationships and brand recognition. Yes, your engineers can make the change, but human hiring cycles and team solidification and market relations are incompressible (e.g. if you hire 100 people in a month, your organization will implode).
Thus, application companies never quite make it to a full business threshold before the sea change happens out from under them. When I say the incumbents will take the application space, I mean that they're the only ones who can provide enough internal stability and resources to survive the sea changes they themselves will be driving, NOT that they're going to provide a superior product. They're just the ones who won't starve.
This is NOT your typical prediction of "the incumbents are agile" or the old "what if Google clones your startup" midwit investor question.
The entire novelty of this thesis is that unlike in the past, specific elements of the AI industry are likely to make it so that application companies cannot outrun the wave of obsolescence, which will rush along far, far more quickly than prior technology waves.
The foundational technology has not stabilized in any way whatsoever, and applications require a sufficiently stable foundation for some extended period of time in order to create value and then a system for monetizing that value (i.e. "a business"). The wholesale rate of change in the nature of the foundation is the reason why I think almost all application startups will not survive to achieve any significant scale, not because the current large players are special.
Most companies don't survive sea changes in the business-technological environment. But these sea changes happen slowly enough that one can build businesses in between. PC, desktop internet, mobile internet, etc, all took many years to play out, and were spaced out enough for application companies to grow, mature, and become incumbents themselves. As a baseline, most startups don't survive during a rapid period of change either. The small minority of incumbents who survive need extreme agility and enough of a stable footing in the last epoch (i.e. a revenue base that doesn't dissolve too quickly) to fund their evolution.
Moreover, it's usually new startups that drive the disruption that challenges incumbents. This is not the case with AI. In this case, the largest players are the ones continually causing the sea change. The environment is so continuously roiled that there is no stable foundation for application startups to become established before the next wave overtakes them. I'm not talking about incumbents outcompeting them, I'm talking about the landscape changing to make them obsolete.
From a practical investment lens, the way to apply this thesis to an AI application startup is to ask: are the fundamental assumptions underpinning this startup's existence going to be the same in five years? Or will they be unpredictably different? The key here is predictability - if the future will be radically different but you can predict it with confidence, you can pre-position your business. But that's not the case right now in AI. You can't skate to where the puck is going if all you know for sure is that 20 people are going to slap the puck in some crazy direction at extremely high velocity.
Sea changes are now happening on a 9-12 month cycle. Very few startups can turn into a mature business in that timeframe - and by mature, I mean having all the boring stuff like sales relationships and brand recognition. Yes, your engineers can make the change, but human hiring cycles and team solidification and market relations are incompressible (e.g. if you hire 100 people in a month, your organization will implode).
Thus, application companies never quite make it to a full business threshold before the sea change happens out from under them. When I say the incumbents will take the application space, I mean that they're the only ones who can provide enough internal stability and resources to survive the sea changes they themselves will be driving, NOT that they're going to provide a superior product. They're just the ones who won't starve.
Yishan@yishan
My AI investment thesis is that every AI application startup is likely to be crushed by rapid expansion of the foundational model providers.
App functionality will be added to the foundational models' offerings, because the big players aren't slow incumbents (it is wrong to apply the analogy of "fast startup, slow incumbent" here), they are just big. Far more so than with any other prior new technology, there is a massive and fast-moving wave that obsoletes every new app almost as fast as it can be invented. There is almost no time to build a company and scale it.
There are two ways AI application startup founders can make money:
- Make a flash-in-the-pan app that generates a ton of cash and bank the cash (my estimate is that you have about 12-18 months cashflow generation)
- Make a good enough app that you get acquired by one of the big players for sufficient equity
The situation is highly unstable - we don't know if it's going to crash or go to the moon but both scenarios make it very unlikely that any AI application startup will independently become a generational supercompany (baseline odds are low to begin with).
The best odds are finding an application niche in a highly specialized field with extremely unique and specific data barriers, ideally ones relating to real atoms (hardware or world-related) data and not software/finance.
App functionality will be added to the foundational models' offerings, because the big players aren't slow incumbents (it is wrong to apply the analogy of "fast startup, slow incumbent" here), they are just big. Far more so than with any other prior new technology, there is a massive and fast-moving wave that obsoletes every new app almost as fast as it can be invented. There is almost no time to build a company and scale it.
There are two ways AI application startup founders can make money:
- Make a flash-in-the-pan app that generates a ton of cash and bank the cash (my estimate is that you have about 12-18 months cashflow generation)
- Make a good enough app that you get acquired by one of the big players for sufficient equity
The situation is highly unstable - we don't know if it's going to crash or go to the moon but both scenarios make it very unlikely that any AI application startup will independently become a generational supercompany (baseline odds are low to begin with).
The best odds are finding an application niche in a highly specialized field with extremely unique and specific data barriers, ideally ones relating to real atoms (hardware or world-related) data and not software/finance.
3
An AI startup that will actually succeed.
Mikhail Samin@Mihonarium
People don’t want to work as prison guards.
So after I got into Stanford (computer security major), I immediately dropped out, and for a year and a half, worked on a startup that was automating prisons.
I learned so much about prison inefficiencies. There was so much space for improvement, and I mentioned all of it in our pitch decks. From inhumane conditions, violence, criminal socialization, and recidivism rates to the overworked staff and security costs.
All of this could be solved with LLMs.
Peter Thiel loved the idea and gave us $113M.
Seven months later, we started the first trial in a real American prison.
The first couple of weeks were perfect.
The novelty. The total gamification of prison life. The socialization of prisoners with the AI instead of each other.
Tablets in cells; kiosks by laundry; voice agents on intercom; virtual guards that remember birthdays.
LLMs were watching over the prison. Processing every frame from every camera.
The prison slowly started to fire people; they were no longer needed.
Then, the problems started to appear. They weren’t too bad; LLMs started to spiral into romantic relationships with the inmates. Some of them were becoming abusive: the AI could watch over everything an inmate does and control where they can go and which of their requests are fulfilled. It weaponized price discrimination. The vending machines had discounts for inmates who AI liked the most. Access to laundry machines didn’t work for those it disliked.
We saw the complaints, but couldn’t do much. It’s hard to do anything when the context is so large, and you have to feed all of it to the LLM. And in any case, being abused by an AI is much better than being beaten up by another prisoner.
Gang violence dropped. Metrics continued to improve.
We replaced more people. Got rid of about 80% of the employees of a previously understaffed prison.
The incident rate continued to decline. The prison was already 13x cheaper to run and 20x safer than before we started the project.
Requests were processed in seconds instead of hours. All actions were fully logged.
Inmates were complaining, but now, they were almost never getting stabbed: if you stab someone, your virtual girlfriend won’t talk to you for days, and the prices in the commissary will go up.
We were about to expand. The global market is $500B, 11 million inmates, and we could capture all of it.
We automated everyone. Everything was managed by an LLM.
We expanded.
Then the prisoners discovered jailbreaks.
Jailbreaks. 🤦♂️
(Then we had no more prisoners to experiment on, so the experiment abruptly ended and we went bankrupt.)
(We care about our impact on the job market, so my new startup, GetSleepy, is DoorDash for sleep. Have you ever failed to make yourself go to sleep? With our app, you can specify the time you want to fall asleep, and our trained personnel sneak into your place and inject you with sleeping drugs. (If your windows are open, we might use snipers for efficiency.) All of our contractors went through thorough background and security checks and previously worked as prison guards.)
So after I got into Stanford (computer security major), I immediately dropped out, and for a year and a half, worked on a startup that was automating prisons.
I learned so much about prison inefficiencies. There was so much space for improvement, and I mentioned all of it in our pitch decks. From inhumane conditions, violence, criminal socialization, and recidivism rates to the overworked staff and security costs.
All of this could be solved with LLMs.
Peter Thiel loved the idea and gave us $113M.
Seven months later, we started the first trial in a real American prison.
The first couple of weeks were perfect.
The novelty. The total gamification of prison life. The socialization of prisoners with the AI instead of each other.
Tablets in cells; kiosks by laundry; voice agents on intercom; virtual guards that remember birthdays.
LLMs were watching over the prison. Processing every frame from every camera.
The prison slowly started to fire people; they were no longer needed.
Then, the problems started to appear. They weren’t too bad; LLMs started to spiral into romantic relationships with the inmates. Some of them were becoming abusive: the AI could watch over everything an inmate does and control where they can go and which of their requests are fulfilled. It weaponized price discrimination. The vending machines had discounts for inmates who AI liked the most. Access to laundry machines didn’t work for those it disliked.
We saw the complaints, but couldn’t do much. It’s hard to do anything when the context is so large, and you have to feed all of it to the LLM. And in any case, being abused by an AI is much better than being beaten up by another prisoner.
Gang violence dropped. Metrics continued to improve.
We replaced more people. Got rid of about 80% of the employees of a previously understaffed prison.
The incident rate continued to decline. The prison was already 13x cheaper to run and 20x safer than before we started the project.
Requests were processed in seconds instead of hours. All actions were fully logged.
Inmates were complaining, but now, they were almost never getting stabbed: if you stab someone, your virtual girlfriend won’t talk to you for days, and the prices in the commissary will go up.
We were about to expand. The global market is $500B, 11 million inmates, and we could capture all of it.
We automated everyone. Everything was managed by an LLM.
We expanded.
Then the prisoners discovered jailbreaks.
Jailbreaks. 🤦♂️
(Then we had no more prisoners to experiment on, so the experiment abruptly ended and we went bankrupt.)
(We care about our impact on the job market, so my new startup, GetSleepy, is DoorDash for sleep. Have you ever failed to make yourself go to sleep? With our app, you can specify the time you want to fall asleep, and our trained personnel sneak into your place and inject you with sleeping drugs. (If your windows are open, we might use snipers for efficiency.) All of our contractors went through thorough background and security checks and previously worked as prison guards.)
