There is a quiet revision happening in how investors and operators talk about AI companies. In 2023, the first question in a pitch meeting was often “which model are you using?” By 2026, the more telling question, as Linking Row’s defensibility map frames it, has become: “what becomes uniquely yours as customers use the product?” That shift, from the model you rent to the asset you accumulate, is the whole argument in miniature. The capability is increasingly commodity. The moat has to come from somewhere else.
This used to be a contrarian position. It is fast becoming consensus, and for good reason: the companies treating the model as their differentiator are the ones showing up on this year’s shutdown lists.
The moat myth
Walk into most early-stage AI pitches and “our model” turns out to be an API call to OpenAI, Anthropic, or Google, wrapped in a system prompt and a clean interface. There is nothing inherently wrong with that, plenty of valuable products are thin layers over powerful infrastructure, but it is not a moat. It is a configuration. And configurations are trivial to replicate, because your competitor is calling the exact same endpoint with a slightly different prompt.
The economics are unforgiving. According to commentary from IdeaProof’s 2026 startup failures analysis (directional rather than precise), thin AI wrappers have been seeing gross margins compress below roughly 20% within about twelve months, squeezed from two directions at once. Inference costs rise or refuse to fall as fast as hoped, while the underlying model providers ship the very feature you built your company around, natively and for free. When that happens, your differentiation evaporates overnight and your unit economics never recover. These are the businesses dominating 2026’s shutdown lists.
The cruel part is the timing. There is a shrinking window between launching a clever prompt-based feature and watching an incumbent absorb it into their platform. If your entire value proposition can be described in a single sentence that begins “it uses AI to,” assume someone with more distribution is already building it as a checkbox feature. Speed of replication has collapsed; the model is the least defensible part of your stack precisely because everyone has equal access to it.
What actually defends a business
If the weights aren’t yours and the capability is shared, defensibility has to be manufactured elsewhere. Four sources hold up under scrutiny.
Distribution into a niche horizontal players won’t target. The big model labs and the well-funded horizontal tools chase large, generic markets. They are not going to build the bespoke compliance workflow for Indian chartered accountants, the inventory-reconciliation tool for regional pharma distributors, or the vernacular customer-support layer for tier-two lending. Owning a specific, unglamorous wedge, with the trust and channel relationships that come with it, is a real moat because the giants have no incentive to come for it.
Proprietary and regulated data flywheels. This is the heart of the new defensibility question. If every customer interaction makes your product measurably better in a way competitors cannot copy, because they don’t have the data, you have a flywheel. The strongest versions sit on regulated or hard-to-access data: financial records, clinical notes, legal filings, transaction histories. The moat isn’t the model touching the data; it’s that the data accrues to you, exclusively, and compounds over time.
Workflow lock-in and switching costs. A product that lives inside a team’s daily process, holding their templates, their history, their integrations, their approval chains, is far harder to rip out than a point tool. Once your software is the place where work happens, switching means retraining staff and rebuilding process, not just swapping a login. Lock-in is earned by becoming infrastructure, not by being clever.
Compliance as a moat. In regulated industries, the painful work of certifications, audit trails, data-residency guarantees, and sector-specific approvals is a barrier that no foundation model confers. A competitor with a better model still has to spend years and significant capital clearing the same regulatory bar. That friction protects you, and in India’s tightening data and financial-regulation environment, it is an underrated source of durability.
The counter-argument, fairly stated
It would be dishonest to claim the model never matters. There are domains where a genuine model edge, or control over infrastructure, is decisive. If you are operating at frontier scale, fine-tuning on data nobody else can assemble, or running inference cheaply enough through owned infrastructure to change the cost structure of an entire category, the model and the stack beneath it absolutely contribute to your defensibility. Cost-per-token and latency, at scale, can become competitive weapons rather than commodities.
And the “wrappers always die” thesis has counterexamples that deserve airtime. Manus, frequently dismissed early on as a wrapper around existing models, scaled anyway, on the strength of product execution, a sticky experience, and distribution momentum rather than any proprietary model breakthrough. The lesson is not that wrappers are doomed; it’s that a wrapper can win if it converts an initial capability lead into something durable, retention, habit, a data loop, before the window closes. The danger is staying a wrapper. The opportunity is using the wrapper as a wedge.
So the honest version of this argument is narrower than the slogan. The model is rarely a sufficient moat, and almost never the only one worth having. But execution velocity and a real capability edge can buy you the time to build the moats that last.
What founders should do
The practical reframe is simple: own the workflow, not the weights. Your job is not to have the best model; it is to be the system your customer cannot imagine working without. Build for the place where decisions get made and work gets done, and let the underlying model be an interchangeable component you can swap as the market shifts. Treating models as commodities is a feature of a healthy strategy, not a weakness.
From there, stack at least two moats. Any single source of defensibility can be eroded: a niche can be entered, a data advantage can be matched by a better-funded rival, a workflow can be cloned. But moats compound when they reinforce one another. Distribution into a niche feeds a proprietary data flywheel; that data deepens workflow lock-in; the workflow generates the compliance footprint that keeps competitors out. Each layer buys time for the others to thicken.
Concretely, founders should be able to answer the 2026 question with specifics:
- What compounds? Name the asset, data, relationships, integrations, that gets more valuable with every customer and every week of usage.
- What’s the switching cost? If a customer left tomorrow, what exactly would they lose or have to rebuild? If the answer is “not much,” you have a feature, not a company.
- Who won’t come for this? Identify why the incumbents and model labs have no reason to target your wedge, and build before that calculus changes.
- What can’t be copied with the same API key? Whatever survives that test is your actual moat.
The model was never going to be the moat for most companies; it was always going to be the table stakes. The businesses that endure this cycle will be the ones that used cheap, shared intelligence as a starting point and then built something proprietary on top, distribution, data, and lock-in that no API call can replicate. Rent the intelligence. Own the relationship.
