Why we built DeclassifAI in Norway

There is a version of this piece that opens with a fjord. We are not going to write that version. Norway has plenty of fjords and they are not why we are here. We are here because of a specific combination of regulation, infrastructure, and culture that turns out to be unusually well suited to building AI for regulated work.

This is the perspective piece in our writing list. The other articles take the legal and technical lines carefully. This one is allowed to say what we think.

The regulator is ahead of the field

The single thing that matters most for an AI company building into regulated work is the quality of the supervisor. A regulator that has thought hard about deployment-time AI questions, published its work in detail, and accepts dialogue with the industry is worth more than a tax break or a procurement programme.

Datatilsynet is that regulator. The 2024 AI strategy^[1] commits the agency to building real technical competence on AI. The regulatory sandbox, running since 2020, has produced eight published final reports. The Norwegian DPA's posture toward generative AI, articulated through the ChatGPT moment in 2023, treats deployment questions as engineering problems that admit of structured answers, not as press-release material.

This matters operationally. When we built DeclassifAI's redaction pipeline, the questions we asked ourselves were the same questions the NTNU Copilot sandbox report asked: where does the data flow, who can access it, what is the lawful basis, what is logged, who reviews. Those questions are public, on the regulator's website, in worked form. Building for them is more work than building for "GDPR compliance" as a marketing claim. It is also clear what done looks like.

Sovereignty is a real constraint, not a slogan

Norwegian regulated work has an additional layer most other EU countries do not: the Norwegian Security Act (Sikkerhetsloven).^[2] The Act imposes specific requirements on entities that handle sensitive information about state security, including supply-chain due diligence and location-of-processing constraints.

Most cloud AI offerings cannot meet these requirements without modification, and many cannot meet them at all. For Norwegian defence suppliers, parts of the energy and telecoms sector, and chunks of the public sector, sovereignty is not a preference but a procurement constraint. Designing software for that constraint from the start, in the jurisdiction that defines the constraint, is a different problem than retrofitting a global SaaS for it after the fact.

The same applies, with adjustments, to Swedish, Finnish, Danish, and increasingly French and German regulated work. The architectural pattern that satisfies Sikkerhetsloven generalises to most of Europe's sovereignty-conscious sectors. Building it native here is cheaper than reinventing it after a contract win.

Hydropower is a quiet AI input

This is the underestimated practical fact. Norway produces around 98% of its electricity from renewable sources, primarily hydropower, with one of the cleanest grids in Europe by carbon intensity.^[3] Industrial-grade power prices in many regions are among the lowest in Europe. The climate is cool enough that data centres do not pay for cooling at the rates they do in southern Europe.

For an AI company, this is not abstract. Inference and training compute is the largest variable cost in the business. Locating it on a clean, cheap grid is a competitive advantage that compounds. It is also a customer-facing benefit: regulated buyers increasingly account for the carbon intensity of their AI suppliers, and an honest Norwegian footprint reads better than carbon-offset arithmetic on a US-Midwest grid.

The result is a small but growing AI infrastructure ecosystem in Norway, particularly around Stavanger, Oslo, and the Gjøvik corridor. We are part of it, deliberately.

The public sector buys software the way you would want it bought

Norwegian public-sector software procurement has its frustrations like any other. But the underlying pattern is healthy in a way that matters for an AI company: the public sector buys software, including AI software, with detailed technical scrutiny, real consultation cycles, and a willingness to fund development of solutions that meet specific public-interest constraints.

The Digitaliseringsdirektoratet (the Norwegian Digital Agency) has been articulating a strategy for public-sector digitalisation^[4] that treats AI as both a tool and a regulatory subject. The combination of a competent regulator (Datatilsynet) and a competent buyer (Digdir and the line ministries) produces an unusually serious public-sector AI market.

For a company building privacy-preserving AI, this is the right kind of demand. Buyers ask the questions that force you to build the right things. Talking to a Norwegian municipality about a redaction pipeline is materially different from talking to a US enterprise SaaS buyer. Both can be productive; only one produces software that holds up under regulator scrutiny without major redesign.

A population that already expects privacy

Norway, like the rest of the Nordics, has a baseline cultural expectation that personal data is handled carefully. The default assumption when a software product handles sensitive information is that there should be a documented privacy posture, that consent should be informed, that retention should be minimised. This is not unique to Norway, but it is more deeply embedded here than in many markets.

For an AI company, this is a user-facing benefit. The features that matter for regulated buyers (visible audit logs, source attribution, defaults that favour data minimisation, configurable retention) are also features ordinary Norwegian users notice and reward. Building software that respects the user is the same engineering work as building software that survives a Datatilsynet review.

The trade-offs we accept

We should be honest about what Norway is not.

It is not the largest market. The economics of building from Norway depend on selling into the rest of Europe and the EEA, which we do. Most of our customer base lives outside the country.

It is not a venture-capital epicentre. Building here means accepting that the capital infrastructure is thinner than in the UK or the Bay Area. Companies that succeed from Norway tend to fund growth from revenue more than from equity. For an AI company in regulated work, with steady B2B contracts and lower marketing burn than consumer AI, this fits.

It is not a flat tax jurisdiction. Norwegian corporate tax is in line with EU averages; personal tax is higher than in many tech-export markets. The compensating offer is a quality of public services and infrastructure that lets the company recruit and retain people without having to compete on cost-of-living alone.

It is not the warmest climate. We work indoors. The data centres are happy about it.

What Norway is, for an AI company

Put together, the picture is specific. Norway gives an AI company:

• A regulator that has published its work in detail and accepts structured dialogue.
• A legal framework that takes sovereignty seriously enough to produce real procurement constraints, which become differentiating capabilities.
• An energy infrastructure that makes the compute economics work.
• A public-sector market that buys software the way you would want it bought.
• A cultural baseline that rewards the right defaults.

None of those are unique. Each exists somewhere else in some form. The combination is unusual. It is the reason DeclassifAI is built in Gjøvik and not in Berlin or Dublin or Palo Alto.

A closing observation

There is a recurring assumption in AI commentary that the only places that matter are the US (for capital and talent) and China (for state-scale deployment). Europe in that frame is a regulator and a market, not a builder. The assumption is wrong, and the next few years will make it more visibly wrong.

The most consequential AI deployments in regulated work over the next decade will be in places where the regulator has thought clearly, the procurement is serious, the infrastructure is sovereign, and the cultural defaults reward privacy. Norway has all four. So do, in different mixes, the Netherlands, Estonia, France, Germany, and a handful of others.

For a company building AI into regulated work, the question is not "where is the centre of gravity for AI" but "where is the centre of gravity for the kind of AI we are building." For us, that answer has been the same since we started: here, on this grid, under this regulator, in this market. Two years in, the answer still looks right.