Proprietary models
Passport
- Rock
- Assets
- Depth
- 3.5 · Mine
- Time to dig
- 10+ years
- Capital
- ◉ · extreme
- Solo
- ✗ no
- AI
- ↑ AI-resistant
- Rent
- ✓ rentable
Sample
- Share of apps
- 12.3%
- No-rate
- 71%
- Median price
- $18.99
Figures from the canivibecodeit sample. No-rate is the share of apps carrying this tag that cannot be vibe-coded — a proxy for structural strength. Only the first thirteen mechanics were measured.
Essence
Models you trained yourself — frontier or narrowly specialised — plus the compute and inference underneath them. The tag is not for wrappers around someone else's API: the moat exists only where the model is genuinely your own.
How it is built
OpenAI and Anthropic — the frontier as a capital race
The moat has three layers: billions in training compute, research talent of the rarest kind, and accumulated feedback data — RLHF, human preferences — that exists in no open dataset. Every generation of models raises the entry threshold, and inference infrastructure serving hundreds of millions of users is a second moat inside the first.
Synthesia and HeyGen — a narrow model on data you gathered yourself
Do not try to out-train the frontier; train a specialised model — photoreal avatars, lip sync — on data the company collected and licensed itself, including studio shoots of actors under contract. A narrow task plus a proprietary dataset plus a production-grade pipeline: frontier models are "smarter in general" and worse at this exact point.
Midjourney — a model sharpened by community feedback
Midjourney trains its aesthetic on millions of user choices: which variants get upscaled, which get rerolled. The product with the largest audience gets the most signal for sharpening the model — the data flywheel again, but at the level of the weights. Staying on that wheel is hard for a challenger even starting from the same base checkpoint.
How it is bypassed
Open source catches up — the capability is commoditised
Stable Diffusion devalued the uniqueness of closed image generation within months; Llama and Mistral did the same for good-enough text; DeepSeek showed that the cost of training at the frontier falls by an order of magnitude too. For most applications the gap between the best model and the free model has narrowed to something the user cannot see — which means the moat only earns at the very top edge of quality.
Distillation and fast-follow
The leader's outputs are the follower's training data: distillation compresses the lead from years to months. You cannot fully defend against it — the outputs are in the users' hands — so leaders are forced to run faster rather than build the wall higher. For the attacker it is a legal grey zone, but commercially it works: each generation of the leader is quickly followed by "almost the same, far cheaper".
Cursor and Perplexity — the value leaks into the application layer
The most important bypass: do not build a model at all. Cursor rents the best models over an API and owns what is not the model — the developer's workflow, the context of the codebase, the distribution. As models commoditise, margin and loyalty collect in the application layer and the model owner turns into an electricity supplier. The incumbent's moat is not broken; it simply stopped being the place where the contest is decided.
Verdict
A model moat is real at the frontier and in narrow domains with proprietary data. It is bypassed by open source, by distillation and — strategically — by moving the game to the application layer, where the model is just a rented component.