Moat types
AI Moats: What Survives When Everyone Has the Same Model
Updated 21 October 2026
An AI moat is a competitive advantage that stays in place even when every competitor can access the same AI model. Models are rented, cheap and improving for everyone. A real AI moat lives in what surrounds the model: proprietary data, an owned workflow, distribution, network effects or earned trust that a rival cannot buy.
The quickest way to check for one is the empty database test. Imagine a well-funded competitor plugs in exactly the same model you use, with the same prompts, but starts with an empty database: no customers, no history, no integrations, no labeled outcomes. If their product would be about as good as yours on day one, your value is in the model, and the model is not yours.
Why AI moats are a hard question
Every software product can be copied. AI products can be copied faster, because the hardest part (the model) is available to everyone through an API.
Three forces make this worse:
- Models are commoditizing. In May 2023, a leaked internal Google document published by SemiAnalysis argued "We Have No Moat, And Neither Does OpenAI", warning that "people will not pay for a restricted model when free, unrestricted alternatives are comparable in quality."
- Open-weight models keep closing the gap. Meta made Llama 2 free for research and commercial use in July 2023. In January 2025, DeepSeek released R1 under the MIT license, claiming performance on par with OpenAI o1 at an API price of $2.19 per million output tokens.
- Prices are collapsing. Epoch AI found the price of reaching a fixed level of model performance has fallen between 9x and 900x per year depending on the task.
Investors noticed early. In January 2023, a16z wrote that many generative AI apps are "relatively undifferentiated, since they rely on similar underlying AI models" and have not found network effects or data and workflows that are hard to duplicate. Later that year, Sequoia revised its own data flywheel thesis and concluded that "the moats are in the customers, not the data", pointing to workflows and user networks as the more durable advantages.
Startups vs big companies
For a large company, AI is usually a feature added to something it already owns: a user base, a dataset, a distribution channel. For a startup, AI is often the whole product. That asymmetry is the problem. When an incumbent adds the same capability, it ships it to existing customers at near-zero acquisition cost. A startup has to answer what it owns that the incumbent does not.
Types of AI moats
None of these come from the model itself. They come from what the model is connected to.
| Moat type | What it is | Why the same model cannot copy it |
|---|---|---|
| Proprietary data | Data created by normal use that rivals cannot buy or scrape | An empty database has none of it |
| Workflow or system of record | The product is where the work and records live | Replacing it means migrating history and retraining teams |
| Distribution | An owned channel to users | Rivals must pay to reach the same users |
| Network | Users make the product better for each other | A new entrant starts with no network |
| Trust and compliance | Security reviews, certifications, integrations approved by regulated buyers | Takes years, not a model upgrade |
| Model ownership at scale | Training your own models on usage data you uniquely have | Only works once usage is large and proprietary |
The last row is the exception that proves the rule. Owning a model becomes an advantage only after you have the users and data to train it, which means one of the other moats came first. See proprietary data and the data flywheel for how the first and last rows connect.
AI moat examples
These companies show the range from strong to absent.
- OpenAI (ChatGPT). Its biggest advantage is arguably distribution, not the model: Sam Altman said in October 2025 that more than 800 million people use ChatGPT every week. Competing models can match benchmarks; matching that habit is harder.
- Jasper. An early AI writing tool built on OpenAI models. After ChatGPT launched, it became a low-cost competitor practically overnight; Jasper cut its 2023 revenue forecast by at least 30% and its internal valuation. The textbook failed empty database test.
- Cursor. An AI code editor that started on third-party models and now says its in-house models generate more code than almost any other LLMs, alongside crossing $1B in annualized revenue in November 2025. Usage at that scale is the data needed for model ownership.
- GitHub Copilot. Distribution through GitHub and Microsoft: Microsoft reported over 26 million Copilot users and 180 million developers on GitHub, with 80% of new developers starting with Copilot in their first week.
- Harvey. Legal AI that embeds in law firm workflows and security requirements; Sacra estimates it reached $195M in annual recurring revenue by the end of 2025 with 1,000+ customers. Its advantage is trust and workflow with elite firms, not exclusive access to a model.
- Glean. Enterprise search with 275+ connectors that inherit and enforce source-system permissions. A rival with the same model still has to rebuild every connector and the permission logic that security teams approve.
- Abridge. Ambient clinical documentation that works inside Epic; it co-developed a nursing product with Mayo Clinic through Epic's Workshop program. Integration is a moat, but it also creates platform risk: Epic launched its own native AI Charting in February 2026, putting standalone scribes on notice.
- Midjourney. Built its early community inside Discord, where users generated and shared images in public channels; founder David Holz said the company was already profitable in 2022. The community is a strength, but it lived on a platform Midjourney did not own until its web app launched in 2024.
- Character.ai. Strong engagement (Sequoia singled out AI companionship as the exception to weak retention), yet in 2024 Google hired its CEO and took a non-exclusive license to its technology. A model lead alone did not translate into independence.
- Duolingo Max. Duolingo launched Max in March 2023 built on GPT-4. Any rival can call GPT-4; none starts with Duolingo's learners, streaks and curriculum. The AI rides on an existing moat.
- Notion AI. The model is generic; the value is that it works over the documents your team already keeps in Notion. A rival with the same model and an empty workspace answers nothing useful.
- Grammarly. Years of distribution as a browser extension and an installed habit, but its core job (rewriting text) became something any chatbot does for free. A clear case of a distribution moat under pressure from a model-level feature.
- Perplexity. Answer search built on models it largely does not own, competing directly with Google and OpenAI. Its defensibility depends on product speed, brand and distribution deals, which are real but less durable than owned data.
How a startup builds an AI moat from zero
- Pick a narrow workflow and own it end to end. Become the place where the work is done and the record is kept, not a text box beside it. This is what Sequoia meant by the moat being in the customers.
- Design for outcome data on day one. Capture what happened after the model answered: accepted, edited, rejected, won the case, closed the ticket. Outcome-labeled data is what an empty database lacks.
- Integrate deeply into systems customers already use. Connectors, permissions and approved integrations take months to replicate and survive model upgrades.
- Earn trust where it is expensive. Security reviews, compliance work and regulated buyers are slow to win and slow to lose.
- Stay model-agnostic. Swap in cheaper or better models as they appear. Falling prices should widen your margins, not erase your product.
- Train your own models only once the data is yours. Model ownership is a reward for usage, not a starting point.
Common mistakes and fake AI moats
- "Our prompts are proprietary." Prompts can be reverse-engineered or rewritten in an afternoon.
- "We fine-tuned a model." The next general release often beats it. A fine-tune counts only if it is trained on data no one else has.
- "We are first." Being early in AI buys months, not years. Jasper was early.
- "We have data." Public, scraped or purchasable data is available to every rival. Check it against the proprietary data test.
- Ignoring the model provider. If OpenAI, Google or Microsoft could ship your product as a feature, that is platform risk, not a moat.
- Confusing a better UI with a moat. Good design matters, but it is the easiest layer to copy.
How Moat Check scores it
Moat Check scores this as the "Beyond the AI model" test, one of ten tests on the moat hub. The question is: "Would it stay valuable if a competitor plugged in the same AI model with an empty database?" The reasoning: if the value comes only from an AI model, anyone with the same model can match it.
- Pass (10 points): The product stays valuable when a competitor uses the same model with an empty database, because of data, workflow, integrations, network or trust that exist today or by design.
- Partial (5 points): The advantage is plausible but not yet in place. Examples: a data flywheel that only matters "at scale", integrations on the roadmap, or a workflow customers could still leave easily.
- Fail (0 points): The product is mainly an interface on top of a general-purpose AI model.
The scoring rule is strict: pass only if it is true today or by design, and "at scale" is a partial at best. Moat Check uses AI and web research to apply this fixed framework, so treat the result as a starting point for your own thinking, not investment advice. The methodology explains all ten tests, and what is a moat covers the basics.
Questions
Do AI startups have moats?
Some do, but rarely because of the model. The durable AI moats come from proprietary data, ownership of a workflow or system of record, distribution, network effects, and trust, all of which a competitor cannot get by calling the same API.
What is a GPT wrapper?
A GPT wrapper is a product whose main value is a prompt and an interface on top of a general-purpose model from OpenAI, Anthropic, Google or an open-weight provider. If the model provider or any rival ships the same feature, the wrapper has little left to defend.
Is a fine-tuned model a moat?
Usually not on its own. Base models improve and get cheaper every year, so a fine-tune is often matched by the next general release. It helps only when it is trained on data competitors cannot get and keeps improving with usage.
What is the empty database test?
Imagine a competitor plugs in exactly the same AI model you use but starts with an empty database: no customers, history, integrations or labeled outcomes. If their product would be roughly as good as yours, your value is in the model, and you do not have an AI moat.
Can OpenAI or Google copy my AI startup?
If your product is mainly a feature on top of their model, yes, and they own the distribution to ship it to hundreds of millions of users. Products tied to a specific workflow, private data or regulated buyers are much harder for a model provider to absorb.
Keep reading
Who will copy your startup?
Real competitors, the likely copycat and a moat score. Free, about a minute.