The most expensive engineering argument of 2026 isn't an engineering question

Build your own AI model or wrap the API — the worked decision brief, with a verdict.

Same argument, every team I've talked to this year, and it plays out the same way each time.

Someone senior says: "We're spending real money on API calls. We have data. We should train our own model." Someone else says: "Have you priced GPUs?" The meeting turns into a benchmark-off about tokens per second and eval scores, and everyone walks out feeling technical. Nobody walks out with a decision.

Build-vs-wrap is a strategy call that teams keep running as an engineering meeting. Run it as the strategy call it is and the default gets clear fast.

Start with the doors. Jeff Bezos's old rule: two-way-door decisions should be made fast, by the people closest to the work. One-way doors deserve slow, senior deliberation. Wrapping a frontier API is a two-way door. You can switch providers, add a fallback, or move to your own model later, all at contained cost. Training and self-hosting is closer to a one-way door: the data pipeline, the eval harness, the GPU commitments, and the MLOps hires are hard to unwind once you've built and hired around them. When one option is reversible and roughly as good, take the reversible one unless you have a specific, tested reason not to. That asymmetry alone tells you the default.

Default wrap. The question worth your time is what would flip it. Three bars, and a decision to build should clear all three, not just the one that flatters your roadmap.

One: moat. Is the model the reason customers choose you, provable in a sentence? If a customer pays you because your model is better, building may be core. If the model is a means to a product they value for other reasons like workflow, data, or distribution, then the model is plumbing, and you rent plumbing.

Two: economics. There's a crossover point where rented API cost (linear with usage) meets self-hosted cost: a high fixed floor of GPUs, MLOps, and eval infra, then cheap at the margin. Below roughly $50–100K a month of inference spend, self-hosting rarely pays back the team it takes to run it. Most teams arguing about this haven't computed their own crossover number. They're reacting to the sticker shock on the API bill.

Three: capability gap. Do you have proprietary data a general model has never seen, and have you run the test where a fine-tune on it beats prompt-engineering plus retrieval on the same task? Almost nobody has run that test. Most teams that think they have a data advantage have a data pile.

Fail any bar and the honest answer is: wrap, set a reminder for two quarters out, and re-decide. That's the option nobody frames. Treat the API as a paid experiment that shows you which narrow slice, if any, is worth owning. You can't know what to build until production traffic shows you where the general model actually fails you.

I wrote this up as a full worked decision brief: the same seven-section structure (door classification, key questions, frameworks applied, decision criteria, sources, next steps, escalation triggers) that YourBrief generates on any decision. It's the shape I wish more of those benchmark-off meetings had walked in with.

Read the full worked brief:
https://yourbrief.io/blog/build-your-own-ai-model-or-wrap-the-api-2026?utm_source=john-substack&utm_campaign=revenue-raid&utm_content=build-your-own-ai-model-or-wrap-the-api-2026

And if you're staring at this same fork with your own numbers, your inference bill, your data, and a real question about whether the model is your moat, you can generate the same brief against your situation for $1:
https://yourbrief.io/brief?plan=promo&utm_source=john-substack&utm_campaign=revenue-raid&utm_content=build-your-own-ai-model-or-wrap-the-api-2026&decision=Should%20we%20build%20our%20own%20AI%20model%20or%20wrap%20a%20foundation-model%20API%3F

— John