There's a single sentence that has killed more side projects and stalled more funding rounds than any actual competitor ever has:
“But couldn't OpenAI just build this?”
It's a reasonable fear on its face. The big AI labs have the models, the talent, and the distribution. So the logic goes: anything you build on top of them is a sandcastle, and the tide — GPT-6, the next Claude, a new ChatGPT “app” — is coming in. Why build a product when the model that powers it could swallow you as a feature?
Here's why that fear, for most products, is wrong. And why what's left after you clear it away is one of the better opportunities in software.
The fear rests on two beliefs nobody checks
“The labs will absorb everything” only holds if two things are true: that the chatbot is the universal interface for every job, and that eventually everyone just builds or prompts their own tools. If both were true, the application layer really would be a rounding error on the way to a few god-model monopolies.
Both are wrong. Let's take them in turn.
Belief #1: chat is the right interface. It isn't.
The chat box did not win because it's the best way to interact with software. It won because it was the fastest thing to ship. As one widely-shared design essay put it:
“The chat box isn't a UI paradigm. It's what shipped.” — Adi Leviim, UX Collective
Language models emit text; text boxes accept text; so the text box became the default AI interface. That was a decision about build speed, not user outcomes. And for a huge share of real jobs, a blank prompt box is friction, not magic. When you need speed, precision, or confidence, typing a paragraph and hoping the model inferred what you meant is worse than pressing a button that does the known thing. As Forbes framed it, whether ChatGPT replaces your favorite apps depends on UI, not AI.
A dedicated product encodes the expertise so the user doesn't have to supply it every time. A general-purpose model can write you a workout, sure — if you re-explain your history, equipment, injuries, goals, and last week's soreness in every session. A product built for that job already knows. The interface is the value, and a chatbot throws that value away by making you rebuild context from scratch each time you open it.
Belief #2: everyone will build their own. They won't.
The other half of the fear is that AI turns everyone into a builder, so who needs your product? History is remarkably unkind to this idea. As Ivan Turkovic argues in “No, Average People Will Not Build Their Own Software With AI,” every technology generation has promised that programming was about to become universal — COBOL, the fourth-generation languages, Visual Basic, no-code — and every time it stayed a niche.
The reason isn't capability. It's desire. Most people do not want to design software. They want the job done. Building something good — the flows, the edge cases, the hundred small taste decisions — is work, and it's work the overwhelming majority of people neither enjoy nor want to own. The real effect of AI, Turkovic notes, isn't that people build their own tools; it's that the tools built for them get better, more personalized, and ship faster. That's not a threat to people who make products. That's the job posting.
Where the value is actually going
The “labs eat everything” story assumes value flows to whoever owns the model. But that's not what the people funding this stuff are seeing. On a16z's own stage, Martin Casado and Sarah Wang made the case that value is accruing across every layer of the stack — models, infrastructure, and applications — directly contradicting the earlier assumption that foundation-model companies would capture it all.
It makes sense once you say it plainly: intelligence is getting cheaper and more abundant every quarter. The scarce, valuable things are everything around the intelligence — a UI shaped to one job, the workflow, the earned trust, the taste, the distribution to people who will never open a terminal. Those are exactly the places a general chatbot can't reach and a trillion-dollar lab won't bother to go.
The honest part: thin wrappers really do die
It would be dishonest to tell you nothing gets absorbed. Plenty does. If your “product” is a system prompt and a text box sitting on top of someone else's model, then yes — you will get swallowed, either by the lab or by the first user who realizes they can paste the prompt into ChatGPT themselves. Thin wrappers are renting land, and the landlord is paying attention.
So the absorb-everything fear isn't wrong so much as it's a filter. It kills the lazy version and spares the crafted one. The correct response isn't “nothing is safe” — it's “build something that isn't a wrapper.”
So what do you actually do?
- Build a product, not a prompt. If a user could get 90% of your value by pasting your system prompt into a chatbot, you haven't built anything yet. Go make the other 90%.
- Pick a job where chat is the wrong shape. Anything that needs speed, repetition, precision, or trust wants a real interface — not a conversation you restart every time.
- Own the last mile the model can't. Workflow, proprietary data, the way it feels, and distribution to people who don't live in an AI console.
- Meet users where they already are. You will not retrain a few billion people into software designers. Build the thing they'd rather just open — and, per our take on the death of the MVP, make it genuinely lovable while you're at it.
Why we bet on building products
We make standalone products on purpose, not chatbots. Kiron exists because nobody wants to prompt an AI mid-set — you want a trainer that already has your plan and is counting your reps. B-Side exists because nobody wants to interrogate a chatbot about what to listen to — you want five new albums waiting Friday morning, each with a reason you'll like it. Could a frontier model technically do either? With enough babysitting, probably. That's exactly the point: nobody wants to babysit. They want the thing done well, in an app shaped for it.
The labs are going to keep making the underlying intelligence better and cheaper — which is fantastic news, because we get to build on top of it. The opportunity was never in owning the model. It was in knowing what to build with it, and having the taste to make it worth opening twice.
Sure, a general-purpose model could probably fetch what our apps fetch — the way a Labrador could, in theory, herd sheep. But you want the dog that was bred for the job, not the one improvising with great enthusiasm and mixed results. We're very much bred for the job. 🐾
