The AI boom that started with ChatGPT isn’t just about the big AI labs developing foundation models. Many startups have appeared, that leverage the foundational models for practical applications. At first, these were dismissed as “just ChatGPT wrappers” that would be eaten big the big labs as soon as they found a profitable model. But 2 or 3 years later, some of these wrappers are still doing very well for themselves (Perplexity and Cursor come to mind). So, are AI wrappers here to stay ?
“It’s just a wrapper”
The dismissing of AI wrappers relies on the assumption that the only difficult part of an AI application is the AI, and therefore wrappers are easily reproducible. So the big AI labs would just need to wait for a wrapper to become successful, and then copy that wrapper and integrate it with their product. As an ace in their sleeve, they would be able to cut access to the model for their now-competitor.
In the end, the underlying assumption is that they bring very little value to the market.
But the wrappers are still here
Despite the strong doubts some cast about them, those “wrapper companies” kept building, and they still thrive today. They broadly fall in 3 categories: IDEs (applications to write code) such as Cursor, Lovable, Replit; search interfaces like Perplexity and You.com; and graphics design helpers (to make slides, videos, web pages, etc).
As predicted by the crowd critical of wrappers, the big AI labs have copied the successful wrappers and integrated it tightly with their own models:
- Google launched Firebase Studio, its AI-powered IDE
- The big players now all have a Web Search and Deep Research mode inspired from Perplexity in their chat interface
- The image generation and canvas/artifact feature to design web pages have also been implemented by most of the big labs
These attempts at copying wrappers have had more or less success, but even the successful ones haven’t killed the wrappers they copied. They just exist next to each other, trying to capture different market segments.
As for the ace in the sleeve of the big labs, their monopoly on good models, it has been shattered by the publication of highly-performant open-source models such as Deepseek and Qwen, and the flourishing of inference providers that serve them.
AI Scaffolding is the new big thing
When the initial excitement around large language models began to fade, many companies realized that the groundbreaking productivity gains and industry disruption they’d been promised weren’t materializing as quickly as expected. Some did see improvements, but only after investing significant effort into integrating AI into their existing workflows. This integration work received the name of “AI scaffolding”, marking its importance in building around AI. The definition of “AI scaffolding” is in fact the same as that of “AI wrapper” (a piece of software that uses AI), but with a positive connotation.
It turns out that doing this engineering work actually brings value to the market. The valuation of AI wrapper companies proves it, and their growing user bases prove it. AI scaffolding is getting praised for good reason: it is necessary in order to make AI useful.
In the end, AI wrappers and AI scaffolding are the same thing. It’s honest software engineering that happens to have AI as part of its toolbox, and it deserves as much respect as other software engineering.