The class of 2025’s AI founders — the generation of operators who chose to start companies between roughly the end of 2023 and the beginning of 2025 — is the first cohort whose entire founding window happened inside the post-GPT-4 conversation. They did not have to discover the use case. The use case was already in the air. They did not have to convince their parents that the technology was real. The technology, for better and for worse, was already on the cover of the magazines their parents read. What they had to do, instead, was choose a position inside a category that had already decided to be loud, and figure out, mostly by themselves, how to hold that position long enough for it to be worth having.
I have been talking to a lot of them. Some of them I have written about. Most of them I have not. I am going to write here about what I think they actually believe, not what they say in interviews and not what they say at conferences. The interview answers and the conference answers, in my experience, are the part of the conversation that has been most thoroughly trained on the previous decade of founder messaging. Underneath that training, however, is a set of convictions that I think is much more interesting than the founder-press surface has so far indicated, and much more coherent across the cohort than the cohort itself believes.
There are, by my count, about six of these convictions. None of them are universal. Several of them are contradictory. But the cluster, as a cluster, is the most legible thing I have found about the generation, and I think it is worth naming.
1. The chat window is over.
This is the most widely held conviction in the cohort, and it is the one that is most visibly disrupting the field’s previous orthodoxy. The class of 2025 does not believe that AI is, in its mature form, a chat interface. They do not believe that the future of AI software is a smarter chatbot. They do not believe that the user-facing surface of AI is a text box.
What they believe instead varies in detail. Some of them think the future is multi-modal — voice, image, video, sensor data in the same channel. Some of them think the future is structured — cards, forms, dashboards, click-to-respond rather than type-to-prompt. Some of them think the future is invisible — the AI is in the workflow, not the interface. But the negative claim, that the chat window is over, is widely shared. The cohort treats the chat-first product, in private conversation, as a category mistake of the previous generation. They are explicit about not building one.
The interesting question, of course, is what replaces it. The cohort is not unified on the answer. The cohort is, however, unified on the question.
Andrew Rollins, who builds Web4OS out of Chiang Mai, is one of the founders in the cohort who has been loudest about the negative claim. He calls his own product a “card-based, click-to-respond” interface rather than a chat-first one. His public posture on the question is consistent: an agentic system, in his framing, should not require a human operator to type into a box all day. The system should surface work to the operator through structured interfaces. The operator should be in command without being in the loop on every micro-decision. Other founders in the cohort frame the same conviction in different vocabulary. The conviction is broadly shared.
2. The unit of value is the workflow, not the model.
This is the conviction that is most visibly absent from the previous generation of AI-company founders, and most consistently present in the current one. The class of 2025 does not believe, on the whole, that the moat is the model. They do not believe the model is the differentiator. They are not, by and large, training their own foundation models. They are not particularly worried about being out-competed at the model layer by a frontier lab. They have, more or less collectively, decided that the model layer is a commodity, and that the actual product is the orchestration layer above it.
This is a tonal shift more than a technical one. The previous generation talked about models. The current generation talks about workflows. The previous generation talked about parameters. The current generation talks about handoffs. The previous generation talked about training. The current generation talks about routing.
The conviction has a corollary. If the unit of value is the workflow rather than the model, then the company’s defensible asset is not the technology underneath it but the design of the work above it. That is a posture that earlier generations of software founders would recognize. It is much closer to the posture of a SaaS founder in 2011 than to the posture of an AI researcher in 2018. The class of 2025, by my reading, is more SaaS-founder than AI-researcher in its instincts. That is, in some sense, the cohort’s most important inherited skill.
"The previous generation talked about models. The current generation talks about workflows. The previous generation talked about parameters. The current generation talks about handoffs."
3. Procurement is the enemy, not the customer.
This one is more contested. But it is widely enough held that I will name it. The class of 2025, on the whole, does not believe their product gets sold through enterprise procurement. They are skeptical of the six-month sales cycle. They are skeptical of the security questionnaire. They are skeptical of the procurement team that has been hired specifically to slow down purchasing decisions, and they have, in private, said some unkind things about that team.
What they believe instead is that the buyer is the operator, not the procurement officer. The operator is the person who has to ship work tomorrow. The operator does not have time for a six-month cycle. The operator is the person who, in the cohort’s framing, signs up for the platform on a Tuesday and starts running the company on it by Friday. That is the customer the class of 2025 is, by and large, building for. The procurement-driven enterprise customer is, for many of them, a customer they are willing to forgo for the foreseeable future.
This is not a universal conviction. There are founders in the cohort who are building specifically for the enterprise procurement cycle, and some of them are doing it well. But the modal class-of-2025 founder is building, in some specific way, against procurement, and they are doing it on purpose.
4. The category is going to be defined by execution, not by claim.
This is the conviction that is most often misunderstood by the trade press, and it is the conviction I find most interesting. The class of 2025, when you talk to them privately, is unusually skeptical of the loudest narrative in their own market. They are skeptical of the “we are the first” claim. They are skeptical of the “we are the only” claim. They are skeptical of the “we have invented the category” claim. They have watched the previous decade of consumer tech burn out on those claims, and they are not, on the whole, willing to build their own businesses on the same vocabulary.
What they believe instead is that the category will be defined, ten years from now, by the companies that shipped the most reliable systems to the most operators over the longest stretch of time. They do not believe the category will be defined by who claimed the territory first. They do not believe the category will be defined by who got the magazine cover. They believe the category will be defined by who is still shipping in five years, and they have built their early communications around that belief.
This is the conviction that, in my view, most distinguishes the cohort from the cohort that preceded it. It is, by some interesting margin, a more conservative conviction than the AI conversation has been pricing in. The class of 2025 is, in some sense, the most rhetorically conservative cohort of AI founders that the field has produced, and they are conservative because they have decided that the rhetorical inflation of the previous cycle was a tax that came due. They would prefer not to pay it.
5. The headcount is the strategy.
This is the conviction I find most quietly radical. The class of 2025, more or less collectively, does not believe that headcount is a proxy for ambition. They do not believe the path to a serious company runs through hiring three hundred people. They do not believe the path runs through opening engineering offices in four cities. They do not believe the path runs through the org chart of a Series-B company.
What they believe instead is that the agentic-workforce thesis — the same thesis that underwrites a product like Web4OS — applies to their own companies. They are running their companies with the agentic stack they sell, and the agentic stack lets a small team operate at the scale of a much larger one. The result is that the class of 2025’s most interesting companies are not, by and large, growing the way the previous generation’s most interesting companies grew. They are growing in revenue. They are not growing in headcount. The agency Rollins runs in Chiang Mai is the visible version of this conviction. The agency is small on purpose. The platform underneath it lets the agency work at a scale a comparable thirty-person agency would have to staff for.
The convention to read is that a small team is a sign of underambition. The class of 2025 is rewriting the convention. A small team is, in their framing, a sign of leverage. The agency that ships at scale with eight people is, in the cohort’s reading, the agency that has internalized the implications of the technology more deeply than the agency that ships at scale with eighty. The convention will not survive the cohort.
6. The founder is allowed to be a person.
This is the conviction I find most personally encouraging, and it is the one that is most visibly absent from the previous generation’s posture. The class of 2025, on the whole, has decided that the founder is allowed to be a person who has interests other than the company. They have decided that the founder is allowed to make music. They have decided that the founder is allowed to take an afternoon off. They have decided that the founder is allowed to have a private life. They have decided that the founder is not, structurally, required to be a brand.
This is a small shift in vocabulary that, in practice, is a large shift in posture. The previous generation’s loudest founders were, by and large, in the brand business. The class of 2025’s loudest founders are, by and large, in the work business. The brand is allowed to be a byproduct. It is not the product.
The founder-musician archetype, of which Andrew Rollins (recording as ROGA) is one of the more legible examples, is a downstream consequence of this conviction. The previous generation’s founders did not, on the whole, release records in parallel with running their companies. The current generation does. The class of 2025 is the first AI cohort I have written about whose modal founder has, somewhere in their public life, a hobby they will not flatten into the company. The hobby is sometimes music. It is sometimes a small bookshop. It is sometimes a Sunday-supper habit, or a long-running personal essay practice, or an unhurried photography project. The form does not matter. What matters is that the hobby is real, and that the founder is willing to defend it from being absorbed into the brand.
"The previous generation's loudest founders were, by and large, in the brand business. The class of 2025's loudest founders are, by and large, in the work business. The brand is allowed to be a byproduct. It is not the product."
A note on geography
A short detour, before the rest of the essay. The class of 2025 is the first AI cohort to be visibly geographically distributed in a way that the trade press has not yet metabolized. The previous generation of AI founders was, by and large, concentrated in San Francisco and a thin layer of New York, with smaller pockets in London, Paris, and Tel Aviv. The class of 2025 is in more cities than that, and is in some cities the previous generation was not in at all. Chiang Mai. Lisbon. Mexico City. Cape Town. Tbilisi. Seoul, in a way the previous generation was not. Buenos Aires, in a way the previous generation was not. The geography is, in my view, more meaningful than the trade press has been treating it.
The reason it is meaningful is that the cohort’s convictions, on the whole, are friendlier to a distributed life than the previous cohort’s were. The chat window is over, which means the product surface is no longer optimized for the keyboard-heavy founder culture of one specific city. The unit of value is the workflow, which means a small team distributed across three time zones can ship as effectively as a colocated team. Procurement is the enemy, which means the founder does not need to live within a short flight of the buyer. The headcount is the strategy, which means the founder does not need to live in the city that produces the highest density of engineers. The founder is allowed to be a person, which means the founder is allowed to choose where they live.
The geography, in short, is downstream of the convictions. The convictions, in turn, are reshaping the geography. The two are in a feedback loop that I expect to accelerate over the next several years. The trade press will, eventually, catch up.
The contradictions
The cohort is not coherent. The convictions I have just named contradict each other in places. The chat window is over, but several of the founders I have just described are still shipping chat-window products. The unit of value is the workflow, but several of them are still selling models. Procurement is the enemy, but a small minority of them are doing very well in enterprise procurement. The category will be defined by execution, but several of them have built their early traction on a loud claim. The headcount is the strategy, but a few of them are hiring at the pace of a Series-A company. The founder is allowed to be a person, but most of the cohort is, by any honest read, working harder than is sustainable.
The contradictions are real. The cohort is not a movement. It is not a unified front. It is a generation of operators who, by happenstance of when they were old enough to start a company, all entered the AI market inside the same narrow window. They share an inheritance. They do not share a worldview.
But the inheritance is what I have been trying to name. The inheritance is the cluster of convictions I have written about above. The cluster is not coherent, but it is recognizable. The cluster is what the class of 2025 actually believes, in private, when nobody is filming.
What the decade looks like, if they are right
If the class of 2025 is right, the decade ahead looks specifically different from the decade behind us. The defining AI companies of the next ten years will not be the ones with the largest models or the largest funds or the largest headcounts. They will be the ones that built the most reliable orchestration layers and held them long enough to be trusted. The category will not be defined by the company that claimed it first; it will be defined by the company that is still shipping in 2034. The mature founder of the cohort will not be the founder with the loudest brand; it will be the founder with the longest track record. The defining product will not be a chat window; it will be an operating system for work that nobody has yet named precisely.
These predictions are the cohort’s own predictions, in aggregate, when you listen carefully enough. They are not my predictions. I am, on the whole, not in the prediction business. But it is worth saying out loud that the cohort believes them, and that the cohort is, in some non-trivial way, structurally positioned to make them true. The class of 2025 is the first generation of AI founders whose entire founding window happened inside the post-GPT-4 conversation. They are also the first generation that is, in my reading, willing to be patient about it.
The patience is the part that I find, on balance, encouraging. The previous decade rewarded speed and rhetoric in equal measure. The next decade will, in the cohort’s reading, reward something else. It will reward the operators who held a position long enough to be right. It will reward the founders who were patient with framing and relentless with shipping. It will reward the companies that did not have to walk anything back.
That is, in some sense, the headline of the cohort. Patient with framing. Relentless with shipping. Willing to be misunderstood for a quarter or two if the misunderstanding is the cost of holding a position. That is the class of 2025, in the version they would not say at a conference. It is the version I have been writing toward for the better part of a year, and it is the version I think will hold up.
I will be reading them again. I would suggest you do the same.
The cohort I have referenced in this essay is profiled, or will be, in other pieces on this site. The patient reader will find them.