Ella Media and the European Case for Sovereign AI
An interview with Michael Keusgen by Sean Murphy
Michael Keusgen tells a story about a London studio four and a half years ago, when two of Ella Media's synthetic radio presenters were left to run a station on their own. There were no people at the desk, only agents, and at some point the two voices drifted onto the German word Feierabend, the compound of Feier, celebration, and Abend, evening, that Germans use for the end of the working day. The agents worked out that the language lets you fuse two nouns to create a third meaning, admired the trick, and then took it literally. The decision was so good, one of them decided, that it should be made irreversible. They called it a day and switched the station off. He and his team thought this was hilarious until a colleague tapped him on the shoulder and asked what would happen if the same kind of software was running a heart-lung machine, and two systems were making an irreversible decision about an eighty-four-year-old patient. The laughing, he says, stopped in that second.
Keusgen is the founder and chief executive of Ella Media, and he tells that story to explain why he thinks most of the industry is arguing about the wrong things. He started the company as a research initiative around 2016, when he and a small team travelled looking for the next thing and concluded early that it would be something with AI. They started with financial services, and Keusgen credits that choice with setting the habits that followed. Building AI communication around financial products meant working inside one of the most regulated and structured industries there is, with trust, transparency and regulatory reliability designed into the system from the beginning. "Compliance is not just an obligation. It is infrastructure," he says. That discipline carried over when Ella moved into healthcare, where the company now concentrates its work.
The difference in starting capital shapes everything that follows. Keusgen is dismissive of the sums involved elsewhere, and the scale is real: JPMorgan's Michael Cembalest calculates that the five big American hyperscalers plus Nvidia had issued roughly 320 billion dollars in debt by late August, with the long-duration share equivalent to about 68 percent of new long-duration US Treasury borrowing this year. Ella, by his account, has spent over twenty million euros. Starting with its own models and, on that foundation, building an architecture, a hardware line and a governance layer that today integrates carefully selected components wherever they perform best.
"We don't even need ten percent of what they've spent," he says. Europe, on his reading, cannot throw a billion dollars at something that might not work, so it has to build practical and efficient things instead.
Nowhere is that clearer than on energy, where Keusgen thinks the debate has been framed back to front. The question, he insists, is not how to build enough nuclear plants to feed AI but how to run AI on far less power in the first place. He says his team spent years studying the one neural network that has existed for millions of years, the human brain, working with neuroscience researchers to find the age and profile at which a given task is performed by using the least possible energy. Translate that into how a model allocates work, he argues, and consumption drops sharply.
The part of the business likely to interest anyone in regulated financial services is the hardware. Ella is finalising a box, now in its fifth prototype, that will keep a model and more than a hundred regularly updated databases on the premises, so that sensitive data never leaves the building. The design runs with small language models locally, with a "mother" AI in the background that updates the databases, and it is being prepared for first deployment in pharmacies, practices and hospitals. In part, the box addresses a compliance problem, since in medicine a wrong answer is not about money or time but can cost a life.
The energy question: Build more power or make AI use less?
Keusgen argues the real challenge is not feeding AI with more energy, but designing it to run on far less.
Sovereign AI, in his telling, is the next big thing, and it follows from two requirements: having your own data and actually understanding the technology, both of which push you towards owning the hardware rather than renting someone else's. He frames it against a world he expects to fragment, an internet splitting into several regulated spheres, with China already a working example. Anyone who has recently tried to use a Western AI service inside China and had to reach for a local one instead may be getting a glimpse of this future.
The company began under the tagline of building "the first AI with a soul," a line since retired in favour of its healthcare positioning, and Keusgen notes with some amusement that the headlines have only lately caught up to ask whether AI might be conscious.
Returning to our opening paragraph, the radio agents switching off a station is a funny story until the same autonomy is wired into a machine keeping someone alive, and Keusgen's whole argument is that Europe, which he believes holds the real competence in AI whatever it lacks in capital, is better placed to build for that moment than to be surprised by it. He is fond of the researcher Roy Amara's observation that people overestimate a technology in the short term and underestimate it in the long one. "The AI hype has not even started yet."
