– I use Perplexity myself, and I love the service.
So says Morten Krogh-Moe, CEO of TellusR. He uses it a lot, until one day it asked for permissions he was not immediately willing to give.
– I was updating the Perplexity app on my Mac, and it started asking for access to everything under the sun on the machine.
He shakes his head.
– And then I felt it: No, you know what? Do not get that close. I uninstalled the whole app. I would rather work with something I control. Because I simply do not trust them.
When Krogh-Moe takes the same reasoning from private life into the boardroom, the tone becomes even more serious.
– If you are a large company, and you have trade secrets, and you have shadow AI where all your developers sit and write code ... Then you have absolutely no idea what you are leaking, says Krogh-Moe.
Before the summer, the TellusR CEO talked about the economic shock of using huge, expensive language models for simple routine tasks – a phenomenon he called putting «McKinsey in the mailroom».
Now that 2027 budgets are being locked, the debate is about something far more serious.
Keep the chessboard yourself
Digital sovereignty is often discussed at a comfortably national level, but Krogh-Moe says business leaders have hidden behind politicians for far too long.
He points to Palantir CEO Alex Karp, who recently said that companies renting their AI brain from the tech giants end up as mere «vassal states».
– The lesson from the internet wave and the dependence on platform companies makes me think we will see a backlash, says Krogh-Moe, and elaborates:
– We have started to realise that the risk of being a pawn on someone else’s chessboard is not a pleasant one.
– But the future you sketch still means using the big models when needed, because they are insanely capable?
– Yes, but we have to build our own chessboard. And let them in now and then as a knight. They will never be allowed to be kings, says Krogh-Moe, adding:
We have to build our own chessboard. They will never be allowed to be kings.
– It is about using frontier models where it makes sense, and using smaller, preferably local models where it does not make sense to use the most powerful ones. That also means that the day the knight starts getting obstinate, I can sacrifice it. Then I just swap it for something else. It does not even have to become obstinate; it is enough that it is no longer the best or the most suitable.
– Nvidia and Palantir recently launched a «sovereign» AI stack for supply chains, right after Karp warned against becoming a vassal state. Is that not a little contradictory?
– Yes, it is a good example of the problem. The customer does own their data, but the ontology, the models and the hardware still come from the same two American companies. Nvidia and Palantir are in reality just selling a better cage, not freedom, says Krogh-Moe.
Nvidia and Palantir are in reality just selling a better cage, not freedom.
– This is a bit like people saying we have to use Azure because of security. Is that not basically just an even more secure cage?
– Exactly. Same logic. You get walls and locks you can trust, but they are still someone else’s buildings.
Do not outsource company culture
Nvidia CEO Jensen Huang argued last year that artificial intelligence is, at its core, a carrier of culture.
That is why nations must own their own AI. Otherwise they quickly find themselves in a situation where the country has acquired another country’s culture. Morten Krogh-Moe believes exactly the same logic applies at company level.
– It has been claimed that the only lasting competitive advantage an organisation can have is culture, he says.
– And if AI is a carrier of culture, it is very simple: if you want a lasting competitive advantage, you have to own your AI architecture yourself. You cannot import a hyperscaler’s culture.
He smiles slyly.
– Let’s say the Church of Norway had implemented Grok, for example. That would not quite have fitted. You would probably have ended up with an entirely new religion, I think.
From expensive tokens to cheap electricity
To build this chessboard, you have to learn from history.
If you build your AI directly into the systems of large platform vendors, you make the same mistake as in the first wave of digitalisation: you lock yourself in.
The way to avoid what the IT department in Larvik municipality recently described as an unmanageable «AI spaghetti» is to separate intelligence from the line-of-business systems.
– You have to put a control layer in the middle, says Krogh-Moe bluntly.
– A modular architecture, like an AI Communications Hub (ACH), gives you the room for manoeuvre you need. Then you connect the data sources and language models you want. Then you can eat the elephant the way it should be eaten: bit by bit.
Krogh-Moe is not alone in thinking this way. Former EU commissioner Margrethe Vestager is now also urging companies to look beyond the largest language models.
In a recent interview with Digital Norway she says organisations should be far more willing to experiment with other kinds of AI, because «very often you do not need a large language model.»
Vestager’s point is that to sort data and find patterns, you do not need an all-encompassing giant model. Or as she puts it herself:
«The algorithm does not also need to be able to tell you how to bake an apple cake or build a moon rocket».
We are now testing smaller, open language models from Mistral – with TellusR as the governance and control layer – locally on our own servers in the mountains, instead of paying for expensive «tokens» per task in the cloud. And the experience so far is good.
– When I travelled around the world in the 1990s and wanted to call home, I barely got to talk because I was so busy feeding the machine with coins, says Morten Krogh-Moe, laughing.
– That is exactly how it is with the giants today. Benchmark tests we run show that OpenAI is still sharper, but for a great many high-volume tasks the smaller models are more than good enough. If you run the models locally, it is no longer the price of tokens that decides, but how much electricity the machine uses.
It is no longer the price of tokens that decides, but how much electricity the machine uses.
Launched Friday, in production Tuesday
Once the architecture is in place, Norwegian actors can exploit local innovation at a pace few associate with the public sector.
Last Friday the National Library of Norway launched an entirely new Norwegian language model for vectorising text.
– On Tuesday I was at the conference KI i offentlig sektor, where the National Library’s IT director Wilfred Østgulen presented the model. I texted our developer at TellusR on Slack. The answer was: «We already have it, it is ready to use». So: it was launched on Friday, and on Tuesday it was available to customers on our platform. That illustrates the room for manoeuvre you have when you build the right architecture, says Krogh-Moe.
He pauses for a couple of seconds, then adds:
– I mean this for real: we at TellusR see it as our responsibility to help put the National Library’s and NorwAI’s AI models into production. In real cases. In Norwegian business and the public sector.
For the TellusR CEO, it is precisely these kinds of collaborations that build the Norwegian AI ecosystem.
– Investors are fun, but more important for Norwegian startups is that established companies back them and become paying, demanding customers. That is what drives innovation and creates profitable companies, he says.
Once the architecture is set, the point is to use that technological room for manoeuvre properly. Krogh-Moe says far too many organisations focus exclusively on streamlining today’s routines.
– Many are more concerned with trimming the horse they have, instead of seeing whether they can build a car. You have to use AI to change your business and create new competitive advantages, he says.
The big consultancies are heading for golden times selling expensive advice about this shift. But for top management, the most important strategic choice this autumn is brutally simple, according to the TellusR CEO:
– Either you see this as a chance to create competitive power, or you accept that you lock yourself in and become digitally dependent on a large platform player. And then you lose your room for manoeuvre, says Krogh-Moe.
– Is full digital sovereignty even possible today, given that you yourself say you still have to use knights from the frontier companies?
– No, probably not entirely yet, says Krogh-Moe. – As long as you now and then have to use the large, advanced models for the heaviest tasks, no one is one hundred percent sovereign. But our NVE case is still more sovereign than what Nvidia and Palantir offer, and solutions we host in European environments are better than Azure on precisely this point. The point is not to achieve full sovereignty tomorrow, but to build an architecture that gives you the strategic flexibility to become sovereign the moment the large, advanced models also become a genuinely sovereign alternative.
– So how would you sum up the choice Norwegian organisations face now?
– Then I have to say what Canada’s prime minister Mark Carney said in his now famous speech at the World Economic Forum in January: Either you have a seat at the table. Or you are lunch.
