Over the past three years, the term wrapper has been thrown around endlessly.
While effective at describing the first generation of AI apps that emerged following the launch of ChatGPT, things have come a long way since. Continuing to use the term in an attempt to highlight a supposed lack of defensibility in the application layer today feels lazy. I'd argue that in the current age of AI, great products have never been harder to replicate.
As we're already seeing, and will continue to see over the year ahead, value is rapidly accruing higher up the stack. Think @cursor_ai, @EvidenceOpen, @gamma, @Lovable, @meetgranola, @WeAreLegora, @ModelML_.
The pre-AI software of the past twenty-plus years was largely WYSIWYG. Click button X and Y happens. Features were built on deterministic code, preconfigured to do the same thing over and over again. As a result, functionality was relatively easy to copy one-to-one, regardless of implementation approach. The constraint was the supply of engineering talent.
The current generation of software is fundamentally different. It is built on top of generative models that can take infinite inputs and produce infinite outputs unless constrained by design.
And unlike the 2023-2024 wave of AI software, what we've started to see en masse over the past twelve months are products built not on a single model, but on many. Open and closed models. Fine-tuned models. Models routed based on task complexity and cost. Products that orchestrate multiple model calls under the hood, sequentially for predictable workflows and dynamically for open-ended ones. Products that grant models access to external data and tools, allowing them to search the web, query databases, call APIs and execute code. Products built on obsessively crafted prompts written by domain experts. Products underpinned by thoughtful context engineering and whose behaviour is continuously tested against internal evals. Products built to evolve as technology advances.
As @karpathy noted in his 2025 LLM Year in Review, this new cohort of applications sits at a "much thicker layer above the base models". These are not thin wrappers but complex systems.
What you get (WYG) in the application layer today vastly exceeds what you see (WYS). In this new paradigm, it is nearly impossible to infer what's actually happening under the hood, because the product is no longer the software artefact itself. The product is the invisible layer that dictates the bulk of value: models, orchestration, context. In AI, the devil lives in those hidden details. The last mile of work that consumes the majority of development effort.
When people talk about a lack of moats in software today, they are failing to account for this shift in what constitutes product. They are judging defensibility based on what they can see rather than what they cannot, conflating the shift towards an abundance of software, driven by increasingly capable coding agents, with eroding moats at the application layer. But moats most certainly still exist.
They sit in deep domain understanding of real workflows and edge cases; in rapidly packaging that understanding into product through intuitive UX, product design, and deliberate context choices; in continuously repackaging as models evolve to improve performance, latency, and cost; and in moving fast enough to get into customers' hands, where real usage generates the feedback that fuels the next iteration. The moat is no longer a single defensible asset. It's the organisational ability to run this cycle repeatedly, faster and more effectively than everyone else. To win today, you need to be:
The widespread perception that moats have disappeared in the application layer is understandable. Coding agents are rapidly transforming non-engineers into good engineers and good engineers into phenomenal ones, vastly expanding the pool of capable builders. Every obvious and semi-obvious market is now flooded with "me too" products. This noise creates the illusion that nothing is defensible.
But while building is getting easier, knowing what to build, translating that judgement into product, taking that product to market quickly, ensuring end customers derive real value from usage, and doing it again and again as the technology evolves at breakneck speed is incredibly hard. This is a war of attrition. Teams with the organisational muscle to keep the loop running will build meaningful businesses. The rest will fade.