It was the kind of talk that felt less like a presentation and more like stepping into a control room. Less hype. More steel, chemistry, sensors, pressure curves, alarm lists and the very real complexity of getting AI to work in industrial environments.
That is exactly what made it so good.
Because it is easy to talk about AI as something abstract, weightless and almost magical. What Geir showed instead was how grounded, messy and demanding it is when you try to make it useful in the real world.
And that is where it gets interesting.
If you want to catch sessions like this live, make sure to follow us on Luma: https://luma.com/RunwayFBU
From raw industrial data to actual value
Cognite was founded with a simple ambition: use data and AI to make industry better, safer, more efficient and more sustainable. Nine years later, that is still the core of the company.
That sounds straightforward. It is not.
Industrial data is rarely neat. It is scattered across hundreds of systems, describing the same assets with different names and in wildly different formats. Sensors. SAP records. PDFs. Engineering drawings. 3D models. Images. Point clouds. Vast streams of time series data. Some of the most important information still sits in formats that were never built for modern AI.
What Cognite has spent years building is the layer that makes all of this usable. A shared structure. A common language. A knowledge graph. Absolutely essential if AI is going to do anything more useful than look impressive in a demo.
AI that can actually keep watch
The most exciting part of the talk was not that AI can analyze industrial data. We know that. The shift now is how available and usable that capability is becoming.
Geir spoke about new time series models developed in partnership with NVIDIA. Models that can forecast, detect anomalies and help operators understand what is likely to happen before it happens and across thousands of assets and sensor streams.
This is where things start to feel real.
Not because industry needs more dashboards. It does not. But because industry needs systems that can spot patterns early, raise the right flags and eventually help humans understand what is going wrong and why.
One example Geir shared came from a chemical process where too little water causes a catalyst to precipitate and crash the reaction, while too much water drives unnecessary energy costs. That is not just a data problem. That is lost production, wasted energy and millions of dollars in energy costs.
That is where AI becomes interesting.
Three things that really stayed with us
1. Industry is not an easy place to “plug in” AI
There is no clean training dataset waiting for you. The data is inconsistent, the systems are old, and the context matters enormously. If you want to build here, you need to understand the environment you are building for.
2. A good model is not enough
It has to be safe, useful and deployed in the right kind of use case. Geir was clear on this. If AI gets an ad wrong, you lose a bit of money. If it gets something wrong in a critical industrial setting, the consequences can be much bigger.
3. Norway has a real edge
Silicon Valley has capital and talent. But it does not have industry in the same way. It does not have the plants, the infrastructure, the data, the domain experts and the operational reality close at hand. We do. And that is a bigger advantage than people often realise.
A few answers worth remembering
The Q&A at the end was especially strong, partly because the questions went straight to the hard parts.
What should early-stage companies do if they want to build for industry?
Geir’s answer was clear: work closely with real customers and real problems. Don’t build in a vacuum. In industrial tech, proximity to the actual operation is a competitive advantage.
What did he take away from NVIDIA’s big conference?
That the amount of capital flowing into AI right now is almost hard to comprehend. Not just into models, but into the entire stack around them. GPUs, data centers, power infrastructure. The scale is enormous.
Can these models be used directly in control loops?
Not yet, at least not in safety-critical settings. But they are already useful for monitoring, forecasting and supporting operators. The path forward clearly points toward more agentic systems.
Is Cognite using customer data for training?
Not by default. Geir was very clear that trust is foundational. Without that, there is no Cognite.
Where this is heading
Not everywhere, but in industrial settings where the upside is high and the friction is finally low enough.
This is exactly the kind of conversation we want more of at Takeoff Tuesday. Not just what is new, but what actually works. Not just what is theoretically possible, but what can survive contact with reality.
And afterwards, as always, there was lunch, conversation and the familiar feeling that someone in the room had probably already started building the next thing in their mind.
Are you an AI or robotics builder?
And want to follow Geir’s advice: building with real customers and real problems, not in a vacuum?
We’re actually building exactly that.
A new AI and Robotics Lab where you get access to real data, real customers and industry-leading partners from day one.
Get in touch at hello@runwayfbu.com if you’re interested.
We’ll share more very soon.
