Since I started Genartml, I've probably heard the same pitch fifty times: “We're building our own AI model, fine-tuned on our proprietary data, for [very specific vertical].” Every time, the founder says it with the same tone — half pride, half fear that someone else will do it first.
Nine times out of ten, I want to grab them by the shoulders and say: don't.
What people mean when they say “build a model”
Usually they mean one of three very different things, and they're using the same word for all of them:
- Training a foundation model from scratch. (Extremely expensive. Do not attempt.)
- Fine-tuning an existing open-source model on custom data.
- Doing serious prompt engineering + retrieval on top of GPT / Claude / Gemini.
Only the third one is actually a startup-friendly move in 2026. And even calling it “building a model” is a stretch — it's building a system that uses a model.
The trap
Here's what happens to teams that go all-in on option 1 or 2 without needing to:
They spend eight months and a lot of money getting their model to be almost as good as GPT-4 was two years ago. Meanwhile the frontier lab just released a new model that's better than both — for cheaper API pricing than their own compute costs.
The team's competitive advantage evaporated overnight, and now they have a large training bill, a small team of ML engineers who feel demoralized, and no product.
The frontier labs are running at a speed that no vertical startup can match. If your moat is “our model is slightly better than the general one”, you don't have a moat — you have a lease.
What actually works
The AI companies that win — the ones I watch closely — do something less glamorous. They:
1. Own a very specific workflow. Not “AI for lawyers”. “AI that drafts the specific NDA revision comment your senior partner would write, in your firm's voice, in 90 seconds.” That's a workflow. That's defensible.
2. Own the last mile. Integrations. UI. Trust. Compliance. The unglamorous stuff that GPT can't build for you because it doesn't know your customer.
3. Own the data loop. Every time a user corrects the output, you learn. Over a year, your prompts + your retrieval + your workflow are 50% better because you've watched 10,000 users work with the tool. The frontier model helped — but the compounding is yours.
4. Own the audience. If you're the trusted name in your niche, the switching cost isn't the tech — it's the relationship, the workflow, the integration debt.
What we're doing at Genartml
Every product we ship uses off-the-shelf frontier models. We don't train anything from scratch. What we do is spend obscene amounts of time on the workflow around the model — the retrieval, the prompt strategy, the fallback behavior, the UI, the handoff to humans when the model isn't confident.
That work compounds. The frontier model changes underneath us every six months, and we mostly just get better for free.
Same with Extuter. The AI companion isn't “our custom model”. It's a very carefully engineered system that uses Claude and GPT for different tasks, with retrieval over the JEE/NEET syllabus, with a mistake-memory layer that's ours, with a UX built for 2 AM study sessions. That's the real product.
When you actually should build a model
There are cases. If you're:
- Working with data the frontier labs will never see (regulated healthcare, defense, physical robotics)
- Optimizing for latency or cost in a way that requires local inference at extreme scale
- Building for a language or domain the general models are genuinely bad at
…then yes. Fine-tune. Sometimes even train.
But if you're building for the general knowledge worker, the general SMB, the general consumer — you're competing with the labs. And they're better at building models than you are, by a factor no runway can fix.
The one-line version
Use their models. Build your moat somewhere else.
Most founders resist this because it feels less impressive. “We use GPT with clever prompting” doesn't sound like a startup. Fine — don't say it in your pitch deck. But build it that way anyway, and be here to laugh about it in three years.