The ChatGPT Generation, Four Years In

A reporter asked me this week how the founders who started companies after ChatGPT's public release in November 2022 are doing. Not the dozen names everyone can recite, but the other several thousand. It is a better question than it sounds, because almost nobody is keeping score on that group. The trade press tracks rounds, and the companies I'm describing don't raise very many of them.

Right now, there are really two groups wearing one label. A handful of founders are building at the center of the storm: models, horizontal infrastructure, and the copycats that spring up around both. A much larger, quieter group is applying the same technology to industries that were built long before any of this existed. Almost all of the coverage is about the first group. Almost everything durable is being built by the second.

We've spent the last few years assessing this cohort, backing a few dozen of them, and passing on a great many more. Here is what I'd say about them that I haven't said before.

It is a generation defined by what was abundant

Every founder generation is shaped by whichever input was cheapest when they started. The class of 2011 to 2021 had cheap capital, and it left fingerprints on everything: headcount and top-line growth as status symbols, growth purchased at any price, a decade of companies that never had to learn what their product was worth. The class that started after November 2022 has had something different in abundance: capability. Anyone with a laptop can now build software that would have taken a funded team a year.

On net, that is a wonderful fact and it has a predictable side effect. When building is nearly free, it stops being the scarce skill, and founders who came up believing it was the key differentiator are disoriented. The bottleneck has moved to two places where this generation is structurally weaker: knowing an industry from the inside, and getting an arm’s-length stranger to pay. The founder who can write a production agent in a weekend is now roughly as common as the founder who could stand up a Rails app in 2012. The founder who has sat through a Tuesday morning at a workers' comp carrier is not.

The failure modes look nothing like 2021's

The ZIRP generation died of overspending. This one is failing in quieter, more interesting ways, and we see the same four repeatedly.

1. Revenue that isn't revenue. A surprising share of this cohort's ARR is innovation budget: pilots, proofs of concept, and "AI initiatives" sponsored by someone whose sole job is to have AI initiatives. It renews until the sponsor changes roles or the CFO asks what it did. We've learned to ask, line by line, who approved each contract and whether that person owns a P&L. The companies selling to an operator with a number to hit behave completely differently from the companies selling to a Chief AI Officer, even at identical revenue.
2. The wedge that became a feature. A lot of companies in this generation were built on a capability gap in the models: something GPT-4 couldn't quite do that clever scaffolding could. Those gaps close on a release schedule the founder does not control. The question we ask now is simple: if the underlying models get 10x better and 10x cheaper over the next eighteen months, does your business get better or worse? It's remarkable how many pitch decks are, on reflection, inherently short the very foundation on which they're built. We push on this from a different angle too: what about you, and the specific segment you've chosen, could not be replicated by pointing a sophisticated model at the problem? Founders who can't answer that quickly are usually building a wrapper and calling it a company. We're now seeing the same realization land inside the enterprises these founders sell to. A lot of large companies spent the last two years trying to build these tools themselves, on the theory that a model plus an internal team was enough, and many are quietly moving from "we'll build it here" to "I just want to buy something that already works at scale." Enterprises are living through the same abundance the founders are; they believe a model can do everything, and mostly it can't, at least not without the domain-specific work a founder who has actually done the job can bring.
3. Pricing nobody has solved. Seat-based software pricing assumes a human in the seat. When the product removes the human, the founder is left choosing between charging per task, per outcome, or per some unit of "work" the buyer has never budgeted for. Most of this cohort is still improvising, and improvised pricing shows up later as margin surprises, especially once inference costs are modeled at 10x today's volume rather than at pilot scale. Part of the problem sits on the other side of the table too: most buyers don't actually know what they want to pay for yet, so founder and buyer are both guessing at the same time. That isn't a reason to wait for the market to settle before picking a model; it's a reason to expect the first two or three pricing attempts to be wrong, and to watch how fast a founder notices and changes rather than how confident they sound about the current one.
4. Demo-quality distribution. The product demos beautifully; the founder has never sold anything to the person who would actually use it. Cheap capability means the demo is easy to produce, so it no longer tells you much. What it doesn't tell you is whether anyone in the industry will pick up the phone. We now weigh a founder's first ten customer conversations far more heavily than their first ten features. We also discount the easy wins. A lot of early sales momentum in this generation comes from friendlies, warm intros, or other startups in the same accelerator or cohort buying from each other. That is fine for getting a product built, but it proves nothing about long-term product-market fit; we don't start to believe in traction until the buyers stop being people the founder already knew.

What we ask now that we didn't ask in 2021

Our diligence has changed more in the last four years than in the prior ten. A partial list of the questions that have become standard:

1. Who on the founding team has held the job this product performs or replaces? Not consulted for it, or had a friend in it. Held it personally.
2. Of your current revenue, how much would survive a new CFO who has never heard of you?
3. What did your product do a year ago that a frontier model now does natively, and what did you do about it?
4. What is your gross margin if usage grows 10-fold and model prices don't move? And if they halve?
5. Show me the sales conversation where the buyer said no. What was the objection, and was it about the technology or about you?
6. How many people work here, and why isn't it fewer?

That last one would have sounded absurd in 2021. It's now one of the most revealing things we ask. This generation can run remarkably lean, and the ones who choose not to are usually telling you something about how they think about capital.

A fair grade

That's the diligence. The founders themselves still deserve a fair grade on their own terms.

On balance, I'd give the ChatGPT generation a better grade than the one before it, and I don't think that's close. They are more capital-conscious by necessity, more technically fluent by default, and far quicker to ship. Their median company is a real business much earlier than the ZIRP median ever was. We are seeing companies in our Fund II creating revenue at roughly 3x the pace of our Fund I companies.

Where they're weaker is precisely where abundance made them weak: distribution, domain, and the patience to learn an industry rather than abstract it. The winners in this class will mostly be founders who either already had those things or went and acquired them the hard way, in the field, before they wrote much code. That's a less glamorous story than the one about the models. It is also, I suspect, the one that will hold up. For us, that is the whole underwriting question: finding founders who already have distribution and domain, and backing the rest before the market teaches them the hard way.

Rob

‍

Photo by Daniel Andraski / Pexels

‍