# The model was never your moat A model you can run on a laptop now scores within a few percent of the best closed frontier model. That sounds like earth-shaking news, and for anyone building products the correct reaction is a shrug — because the model was never the thing defending you. Here's why the frontier going free changes almost nothing about how to build, and what actually compounds into an advantage. - **Author:** Fedor Chishchin - **Published:** 2026-06-03 - **Tags:** ai-native, business, agents - **Source:** https://fedorthinks.com/blog/the-model-was-never-the-moat --- A thing that would have sounded absurd two years ago is now just true: you can download a model, run it on a well-specced laptop, and get output within a few percent of the most expensive closed frontier model on the planet. The open-weight pack — DeepSeek, Llama, Qwen, Mistral — has [closed the gap to single digits](https://www.digitalapplied.com/blog/open-weight-models-h1-2026-retrospective-deepseek-qwen-llama) on the benchmarks businesses actually care about, and [70B-class models now run on consumer GPUs](https://www.modemguides.com/blogs/ai-infrastructure/best-open-source-llms-hardware-april-2026) at interactive speed, with quantization landing within 2–3% of full-precision cloud APIs. The frontier, in other words, is becoming free. Every headline frames this as a seismic event — and it is, for a few labs. But if you're building products on top of these models, the honest reaction is a shrug. Here's why. ## A capability everyone has is not an advantage Ask the uncomfortable question: if your product's edge is "we use the best model," what exactly stops your competitor from making the same API call? Nothing. It never did. The frontier was always a shared resource — an endpoint anyone with a credit card could hit. The only thing changing now is that it's becoming an endpoint anyone can hit _for free_, on their own hardware. That's the whole point. The model was never your moat, because a moat is something your competitor _can't_ easily have, and the model is the thing everyone has by definition. Whether it sat behind a $30-per-million API or a free download, it was always the commodity input, not the differentiator. The open-weight wave didn't take your advantage away. It just made it undeniable that the advantage was never there. ## Static data isn't the moat either The usual retreat is "fine, the model is a commodity, but our proprietary data is the moat." In 2026 that's mostly an [outdated idea too](https://www.alpha-matica.com/post/the-data-moat-2-0-escaping-llm-commoditisation-by-owning-the-endless-feedback-loop). A static dataset — even a big one — is increasingly replicable: synthetic generation, fine-tuning on public corpora, or just buying a comparable set from a marketplace. Owning a pile of data you collected once is not a wall. It's a head start that erodes. So if it's not the model and not the data sitting in a bucket, what's left? ## What actually compounds The defensible things share one property: they get _stronger with use_, in a way a competitor can't shortcut by downloading something. A few that hold up: - **The feedback loop.** Not data you have, but data you keep generating: every user action produces signal that makes the product better, which drives more use, which produces more signal. A competitor can copy your model and even your dataset; they can't copy a loop that's been spinning for two years. - **Grounding and domain logic you own.** The deterministic source the model sits on top of — the engine that computes the right answer — is yours, and it's the part the model can't author. My own products lean entirely on this: the language model is interchangeable, but the [chart-computing engine underneath](/blog/grounding-is-a-constraint) and the domain rules around it are the actual asset. Swap the model for next month's better one and the product just improves; the moat doesn't move. - **Workflow integration.** Being embedded in how a team actually works — the permissions, the data, the ten adjacent steps — is far stickier than any model. People don't rip out the thing their week runs on because a rival has a marginally smarter model. - **Taste and distribution.** When the building blocks are free to everyone, the scarce things are knowing _what_ to build and being able to get it in front of people — which is the [whole argument](/blog/building-got-cheap-ideas-didnt) that as building gets cheap, judgment and reach become the constraint. None of those arrive as a download. That's exactly what makes them moats. ## Commoditization is good news if you build right Here's the reframe. The fact that intelligence is collapsing toward free is a _threat_ only if access to a model was your entire pitch. For everyone building the system _around_ the model, it's pure leverage: your single most powerful input is getting cheaper and better every quarter, for free, while the assets you actually defend on — the loop, the grounding, the workflow — keep compounding. There's one condition: you have to be built to take the gift. If your code [hard-codes one model](/blog/cheap-model-most-of-the-work) deep in the logic, each new open-weight release is a migration project instead of a free upgrade. Keep the model behind a clean seam — a swappable dependency — and commoditization works _for_ you automatically. Every time a better free model ships, you get cheaper and stronger by changing a config value, while teams who welded themselves to one vendor watch their cost advantage evaporate. ## The real question So when the next "open model matches the frontier" headline lands, notice that it doesn't actually move your moat — because your moat was never the model. The leaderboard race is a race to own the one component that's commoditizing fastest and that everyone gets for free anyway. The question was never "which model is best." It's "what do I have that doesn't arrive as a free download" — and if the honest answer is "nothing," then the open-weight wave didn't destroy your moat. It just told you the truth about one you never had.