When the model takes away the moat

There has been an explosion of AI-powered products released over the past four years, with most built on top of leading foundation models. Startups, enterprise software vendors, and even large incumbents route their core intelligence through APIs from the leading AI labs, and a growing but still narrow set of open-weight alternatives.

This arrangement has most certainly enabled a genuine burst of AI products, but it has also created a structural fragility which can be remedied through the democratisation of model creation and training.

Defining a wrapper

In its simplest form, an AI wrapper is a product whose core function is an interface sitting on top of a third-party foundation model's API, with little to no proprietary technology beneath that layer. 

That approach can have a short lifespan and Jasper can illustrate that. Once the standout of the AI writing boom with a $1.5 billion peak valuation, Jasper saw its core market vaporized when ChatGPT introduced free, native writing features. Its revenue fell from approximately $120 million in 2023 to approximately $55 million in 2024. Jasper did survive, by redirecting its efforts toward enterprise workflows and brand-specific training.  

Generally speaking, analysis shows that, among AI startups that have wound down in the US in 2025, 81% were "wrappers and apps". Copilots and assistants layered on foundation models, AI productivity tools, content generators, and vertical SaaS products whose core differentiation was an LLM front-end rather than proprietary data or infrastructure.

In Jasper’s case, and in many others, what pressured the company wasn't a competitor out-executing them. It was the platform they built on top of simply absorbing the feature they'd built their company around. 

Not all wrappers are equal

It would be a gross generalisation to say that all AI wrappers are the same. A thin prompt-and-display wrapper is different from a product that does heavy retrieval, tool orchestration, agentic workflows, or domain-specific fine-tuning. The latter can have real defensibility even while sitting on top of a third-party base model. The model call is one component, but certainly not the whole product.

Some companies also use multiple foundation models or fine-tune/distill their own smaller models for specific tasks after prototyping on a frontier API. Open-weight models add variety too as not all wrappers call a closed-source frontier model. 

The wrapper problem

Regardless of where an AI product sits on the wrapper complexity scale, they all share the same problem. When a company's core value proposition depends entirely on an API call to a third party, that company does not fully control its own product. 

It cannot control pricing changes, rate limits, deprecations, shifts in model behaviour, or the possibility that the platform provider decides to compete directly in the same vertical.

This is not a theoretical risk, though how much it affects any given company likely depends on how much of its value sits above the model layer.

The economics of the wrapper

So what is the defensibility of an AI wrapper in practice? 

The short answer is that it comes down to intellectual property.  

Where a company benefits from an existing distribution advantage, established customer relationships, regulatory licensing, network effects, or years of accumulated brand trust, model-layer differentiation may not matter as much. But if what makes a wrapper distinctive is observable from outside the organisation and replicable by a competent competitor within a limited period, it seems reasonable to expect it will eventually be replicated. 

That is the unfortunate economic truth beneath the wrapper model. The cost of replication is near zero for the platform that controls the underlying model, and near total for the startup that doesn't. A wrapper's moat, if it has one, has to come from something other than the model call itself. Absent those, a wrapper is unlikely to have a lasting competitive advantage. 

The barrier to the solution

At this point, you might be inclined to agree that model ownership is one solution to this problem. But that's easier said than done.

The reason so few companies own their models is not a lack of desire. It’s, more plausibly, financial. Training a frontier-competitive model requires enormous compute budgets, specialised talent, proprietary or expensively licensed data pipelines, and the infrastructure to run reinforcement learning at scale. These costs put model creation out of reach for all but the best-funded companies.

For companies to build on their own models as a norm, this barrier has to fall meaningfully. Several forces are already pushing in this direction, open-weight models being one of them. But even then, hosting and fine-tuning an open-weight model is no small task. It still demands technical talent and compute infrastructure most startups don't have access to. 

True democratisation of model ownership

Democratisation, in this context, doesn't mean every company training a frontier-scale model from scratch. That goal is neither realistic nor necessary. It means democratisation reaching the training layer the same way access to AI has been democratised for end users. 

Democratising model training and ownership could be what gives founders a lasting chance at turning an idea into a business that doesn't eventually get absorbed by a model provider as a simple feature. And if that happens, the stakes may go beyond individual business durability. 

As things stand, access to models is cheap. Access to training is not. The thesis that Boltzbit is built on, and the reason our General Learning Intelligence treats learning as part of the architecture, is that the ability to train an AI model should be fully democratised so that the value, the control, and the responsibility can be distributed rather than centralised.

‍