Why I founded Kaldryn: Europe needs AI that stays in the building

In short

I founded Kaldryn in Stockholm in 2021 because European companies wanted to use AI and could not send their data to foreign clouds. Kaldryn One is the answer: an appliance that runs modern language models entirely inside the customer's own building. Five years later, the market has moved our way.

In 2021 I was a software consultant, and I kept sitting in the same meeting. A company wanted to use AI on its own documents. The technology was ready, the business case was obvious, and then the question came that ended it every time: where does the data go? The answer was always some version of a cloud API in another jurisdiction, and for a lot of European companies that answer simply was not acceptable. Legal said no, and legal was right to say no.

I watched good projects die that way, over and over, at different companies in different industries. That pattern is why Kaldryn exists.

What "the data cannot leave" actually meant in 2021

This was not paranoia. In July 2020 the European Court of Justice invalidated the Privacy Shield framework in the Schrems II ruling, which threw EU to US data transfers into legal doubt overnight (CJEU press release). At the same time, the US CLOUD Act compels American providers to hand over data on a valid US demand regardless of where that data is physically stored, which collides head-on with the GDPR.

So for a Swedish insurer or a hospital group in 2021, "we send it to a US API" was not a compliance answer. It was the end of the meeting.

The bet: put the model in a box

The founding thesis of Kaldryn is one sentence: if the data cannot go to the AI, then the AI must come to the data.

The second half of the thesis is just as important. Most companies will never assemble a GPU cluster, a model serving stack, and a machine learning team on their own, and they should not have to. So we package the whole thing as an appliance. A customer plugs in Kaldryn One, connects it to their systems, and gets modern language models working on their documents and records without a single byte leaving the premises.

In 2021 that sounded contrarian. The entire industry was moving the other way, toward bigger clouds and bigger APIs.

Five years later, the market caught up

The numbers have swung hard. Gartner projects worldwide sovereign cloud spending to reach 80 billion dollars in 2026, with European spending growing 83 percent year over year (Broadcom, 2026). Industry analysts project that more than half of enterprise AI inference will run on-premises or at the edge in 2026, up from under 10 percent in 2023 (Six Five Media).

Meanwhile the dependency problem has not moved at all: three US providers still hold roughly 70 percent of the European cloud infrastructure market, against about 15 percent for European providers (ASEE). Europe now regulates AI it mostly does not build, running on infrastructure it mostly does not own. I wrote about what the current stack looks like in what actually works on-premises in 2026.

The part nobody fixed: people leak data anyway

Banning cloud AI does not work either. Samsung learned this publicly in 2023, when engineers pasted confidential source code into ChatGPT three times in twenty days, and the company banned generative AI tools outright (Forbes). The problem has grown since: one 2026 analysis found 34.8 percent of the data employees share with AI tools is sensitive, up from 10.7 percent two years earlier (Unio).

Prohibition fails. Provision wins. Give people a safe internal AI or they will use an unsafe external one. That is the human argument for on-premises AI, and honestly it convinces more boards than the legal one. I walk through the money side in the honest math on running your own AI.

What I believe now

Five years in, three convictions have survived contact with reality.

AI you can point to beats AI you must trust blindly. A machine in your building, under your rules, answers the governance question by existing.
Compliance is an engineering feature, not a legal document. If the architecture keeps data inside, the paperwork writes itself.
Europe will not close its AI gap on rented infrastructure. Enterprise AI adoption in the EU reached 20 percent in 2025, up from 13.5 percent the year before (Eurostat), and every point of that growth raises the stakes on where the data lives.

The thesis has not changed since 2021. Only the urgency has. If your company is stuck where my clients were five years ago, I would genuinely like to hear from you: here is how to reach me.