Why do most people gravitate toward startups that build thin layers on top of AI right now?
There’s a view you often hear from more seasoned founders that the real value is in the last mile, not the model - that is actually solving the full problem end to end for a specific customer. I agree with that, but it doesn’t really explain why so many founders and early employees still go after these kinds of ideas.
It mostly comes down to what gets rewarded early.
If you build something quickly on top of frontier models, you can get signal fast. If it works, you can grow quickly and might even get acquired. If it doesn’t, you fold. That kind of early feedback loop lines up well with how VC portfolios are constructed and with what tends to get attention, so it’s not surprising that most people in startupland lean in that direction.
The alternative is just much harder. You’re picking a problem where solving it end to end is messy, slow, and requires real domain understanding. On top of building the product, you’re also dealing with skepticism, because the value isn’t obvious upfront and it’s not a hot area. That makes fundraising harder since you need more evidence before people are willing to believe the story, which in turn slows you down. It also requires a different kind of conviction, because you don’t get a lot of early external validation that you’re on the right path.
We ran into this with Cartesian. Physical retail wasn’t an obvious or hot space, and the default reaction was that it would be slow, hard to sell into, and unlikely to move quickly. But we kept hearing real pain from customers, even if we didn’t initially have the cleanest way to articulate it.
One thing we learned is that it was easier to convince customers than investors. So we focused on building the product, working closely with customers, and building an awesome team that could go deep on the problem. We stayed bootstrapped, which forced us to learn the space in a much more grounded way. We’ve already come a long way: 700 stores, 15 countries, and multi-million ARR.
Non-dilutive funding (from
@NSF @sbirgov) also helped bridge the gap early on, before the story became more obvious to the outside.
There’s always a lot of FOMO in the startup world, and it makes sense. But if you take the other route, you’re effectively trading speed of validation for depth of execution and defensibility. That trade is harder to make upfront, but so far it has been worth it for us.
MIT student asked a question earlier today that a lot of young founders are quietly wondering about:
"Won’t the frontier labs just do everything?"
Yes it's true that OAI/Ant are shipping at incredible pace, but it's quite easy to avoid their blast radius and build amazing startups:
OpenAI is not going to build a cattle-herding drone, buy an old F-150 and drive from ranch to ranch like the founder of one of the fastest-growing YC W26 startups, Graze Mate.
Anthropic is not going to integrate with dental insurance verification systems (Lance).
Google is not going to navigate NATO procurement (Milliray).
The value is in the last mile, not the model. Sales cycles require humans who understand the customer. And most importantly, the market is expanding, not shrinking: AI isn't cannibalizing the existing 1% software spend — it's unlocking the other 5-6% that was going to humans. That's a much bigger market for startups yet-to-be-founded than the one the labs are playing in.
Now, what DOES seem risky?
A thin UI layer on top of ChatGPT with no domain expertise; a general-purpose chatbot or assistant; or a product that gets obsolete when model capabilities improve.
But — tools for specific industries; "full-stack" AI companies that actually are the service (AI law firm, AI accounting firm, AI uranium exploration company); or generally products where the customer doesn't want a tool but an outcome — are defensible ideas for startups.