Every studio is founded on a bet. Ours fits in one sentence: convenience is rented; control is built. This note explains what we mean by that, why we think the trade-off has genuinely changed, and why we chose to build this company in Gothenburg rather than anywhere else.

The rented path

The default way to add machine learning to an organisation today is to rent it. Data and prompts leave your perimeter for a third-party cloud or a model API. Pricing, rate limits, and model behaviour change underneath you, on someone else's schedule. Compliance rests on a vendor's terms you can't fully audit. And switching provider means rebuilding. The lock-in is the product.

None of that is a scandal. It's a trade: speed now, control later. For a demo or an internal experiment it's often the right trade. Our argument is narrower: for the systems that matter most, it's the wrong one.

Because for a manufacturer or a public-sector organisation, data (and the models trained on it) is the most strategic asset in the building. A model trained on your production line encodes years of process knowledge. A model serving citizens encodes decisions the public is entitled to scrutinise. Route that asset through infrastructure and APIs you don't own, and you're renting your own advantage back, on terms you didn't set and can't hold still.

What open weights changed

For a long time the rented path was defensible, because the alternative was worse: the capable models sat exclusively behind proprietary APIs, and running your own meant accepting a large capability gap.

Open-weight models ended that era. Today you can take a model whose weights you hold, inspect it, fine-tune it on your own data, run it on your own hardware, and keep it running indefinitely: no licence renegotiation, no deprecation notice, no behaviour change you didn't approve. For the bounded, well-specified problems production systems are actually made of, such as inspection, prediction, classification, and document understanding, an open-weight stack is strong enough to build on.

And once that's true, the calculation flips. A model you own that solves the problem beats a slightly shinier one you rent, the moment the system has to run for a decade. Ownership compounds; convenience depreciates.

Jurisdiction is an engineering requirement

Our two verticals, Swedish industry and the public sector, share something beyond geography: both answer to regulation with teeth. GDPR, the EU AI Act, and NIS2 are not legal footnotes to be handled after the build. They are system requirements, as real as a latency budget.

Where data may flow determines the architecture. What must be documented determines the pipeline. What must be auditable determines the logging. If your compliance story depends on a foreign vendor's terms of service, it is promised, not provable, and a promise is not something you can hand a regulator, or a citizen.

So we treat jurisdiction as an input to design: data residency decided at the whiteboard, models running where the data lives (on-prem, private cloud, or at the edge), and compliance evidence produced as we ship, not reconstructed months afterwards. One legal framework, fully auditable, end to end.

Why Gothenburg

You could build this company anywhere. We built it on Sweden's west coast on purpose.

This region quietly built world-class automotive, manufacturing, and deep-tech engineering, industries where systems are designed to work for decades, where safety margins are calculated rather than hoped for, and where documentation and handover are part of the craft, not an afterthought. That culture is precisely what production machine learning needs, and precisely what the demo-driven side of the AI industry lacks.

It's also where our work physically is. The factory floors we deploy to are here. The public organisations we serve operate under Swedish law, accountable to the people of this region. We'd rather be thirty minutes from the machines and the municipalities than close to anything else.

What this means in practice

Sovereign ML, as we practise it, is a short list of commitments:

  • Models run where your data lives. Your data never leaves the perimeter.
  • Open-weight models by default. Nothing critical behind someone else's API.
  • Full source code and documentation handed to your team, with an exit path at every layer.
  • GDPR and EU AI Act compliance designed in from the first commit, with evidence produced as we ship.
  • Every engagement ends in a system your own team runs.

That's the bet Quince is built on. The way to test it is deliberately small: a two-week ML Feasibility Sprint, fixed scope, published price, and an honest answer about whether machine learning is worth it for your problem, on your own terms.