What a wrapper actually is
A wrapper is a product whose core capability belongs to somebody else. You rent a model from OpenAI, Anthropic, Mistral or whoever, wrap it in a clean interface with a well-crafted set of instructions, and sell access.
To be clear: that is not bad. It was the right move in 2024. Distribution and interface genuinely matter, and the first movers who understood what these models could do captured real users while everyone else was still writing internal memos about AI strategy.
But a rented capability has three problems that all arrive at once.
You have nothing to defend. If your product is essentially a very good prompt, a competitor can rebuild it in a month. There is no accumulated advantage. Every morning you start from zero against anyone who decides to enter.
The platform eats you. This is the harsh one. A startup builds a tool that lets you interrogate your PDFs. Six months later the model provider ships that natively, for free, to two hundred million users. The startup was never a company. It was a feature the platform had not gotten around to yet.
Your margins belong to someone else. You pay per request and live on the difference. When your supplier changes pricing, your business model changes with it, and nobody asks your opinion. That is not a company, that is a reseller agreement with extra steps.
What we look at instead
The startups that hold our attention now have one thing in common: the AI has to leave the browser.
We mean that literally and figuratively.
Literally: applications that run offline, on the edge, inside a machine, inside a laboratory, inside a factory, inside a vehicle. Where you cannot simply call an API because there is no reliable connection, because latency matters, because the data legally cannot leave the building, or because the thing you are controlling moves faster than a round trip to a data centre. Edge computing is not a buzzword here, it is a constraint that filters out everyone who is not serious.
Figuratively: the hard part is not the model, it is the domain. The data nobody else has. The process nobody outside the industry understands. The regulation everyone else finds too boring to learn.
If your moat is the model, you do not have a moat. If your moat is the domain, the model is just plumbing.
This is a less glamorous kind of company. It takes longer to build. The sales cycles are slower, because hospitals and manufacturers and insurers do not buy on a Tuesday afternoon after seeing a demo. Fundraising takes more explaining.
It is also much harder to copy. That is the entire point.
The research side is saying the same thing
This is not only a view from the investment chair. Isabelle spent last month in Barcelona on the AI-driven innovation panel at the Workshop on Multimodal Foundation Models: From Research to Innovation, hosted by the Computer Vision Center, alongside Jordina Torrents Bardina and Lukas Fischer of NXAI, and much of that conversation was about the same gap.
Europe publishes some of the best AI research in the world. Turning it into companies that scale from here is the hard part. And the technology is no longer the bottleneck: multimodal models that reason across text, images, video and sensors are already moving out of papers and into products, from medical imaging to robotics, and increasingly they run on the edge.
One thing worth taking from that panel, because it contradicts a common assumption: there is no clean moment where you stop researching and start building. In deep tech the strongest ventures keep doing serious research inside the company. What decides the outcome is the interface between the two, and whether you actually execute.
One example: Vale Biolabs
Vale Biolabs in Schlieren is a good illustration of the pattern.
Drug discovery still leans heavily on animal models and endpoint assays that predict human response poorly. That is a well-known problem and an expensive one, because failures show up late, in clinical trials, after the money is spent.
Vale Biolabs builds a high-throughput organoid data platform that automates time-series phenotyping. In plain terms: they grow human tissue models, watch them over time in an automated way, and turn what happens into structured data you can actually act on.
Notice where the AI sits. It is not the product. It is what makes the volume of data usable.
Now look at the team, because this is where the defensibility actually lives. Ilaria Incaviglia, Co-founder and CEO, holds a PhD in Biophysics from ETH and Harvard and spent three years in early stage venture capital. Shenchen Wang, Co-founder and CTO, came from software engineering at Meta and is founding his second company in lab automation. Angeliki Damilou, Founding Scientist, is a postdoc fellow at Harvard Medical School with a PhD in Developmental Neuroscience from UZH.
Organoid biology, laboratory automation, data engineering and an understanding of how investors think, all in one room. That is not a combination you assemble over a weekend, no matter which model you have access to. It is also a useful reminder that in deep tech the team composition often is the moat.
They were founded in April 2026 and are backed by Venture Kick. They are also led by a female CEO, which in deep tech is still rarer than it should be, and part of why we pay attention.
The four questions we ask now
If you are building in AI and want to know how we will look at your company, it comes down to roughly this:
1. What do you know that nobody outside your industry knows? Not what your model can do. What you understand about a specific messy real-world process that took you years to learn.
2. What happens to you if the model providers ship your feature next quarter? If the honest answer is “we are finished”, that is the conversation. Better to have it now than in your Series A.
3. Where does your data come from, and does it get better as you grow? Public data is not an advantage, everyone has it. Data your customers generate through using you is a different thing entirely.
4. Why does this need to run where it runs? If the answer is offline, on-device, inside a regulated environment or under a latency constraint, that is usually where the real engineering, and the real defensibility, lives.
What this does not mean
A caveat, because we do not think in absolutes.
This is a thesis, not a law of nature. Some companies started as thin wrappers and became substantial businesses through distribution, brand, or data they accumulated along the way. Interface and go-to-market are real advantages, and dismissing them entirely is its own kind of naivety.
And “boring domain plus AI” is not automatically a good business either. Plenty of founders have picked a difficult industry, learned it properly, and still built something nobody wanted to pay for.
The point is narrower than it sounds: the model is no longer the differentiator, because everyone has access to roughly the same models. Whatever makes you hard to replace has to come from somewhere else. Find that first, then decide which model to use.