AI no longer needs to be dependent on the cloud. When intelligence moves directly into machines, sensors and industrial equipment, new opportunities for efficiency, safety and predictive maintenance open up. Stefan Månsby, our CEO at Stream Analyze, explains why the next big step in AI is happening at the edge.
_____________________________________________________________________________________________________
The real value of edge AI only arises when all AI work can be done directly in the device.
A forklift in a warehouse already knows most of itself, down to that half-second of hesitation in the hydraulics that wasn't there last week. Yet it sends all its data to a data center hundreds of kilometers away to learn what it already has on board.
Intelligence is moving into the physical operations that companies already run: the machines, sensors, and equipment that have constantly generated data and sent it out for someone else to interpret.
We’ve stopped thinking it’s weird. Collect, send, process, send back – we’ve been doing it for so long that it no longer seems like a choice. But that’s changing, and it’s time to review where and how AI is actually being used out at the edge.
Most of what is called edge AI doesn't run at the edge.
Most of what is called edge AI today is not actually run where the machine is, i.e. in the mine, on the truck or in the factory. A model is trained in the cloud, frozen and sent out to a device where it is allowed to perform a single narrow task. Everything else, such as training, adaptation and monitoring, is still done centrally. What is happening now is that the entire chain is moving into the device. You can build the model, adapt it, ask it questions and run it on site, on real-time data, while the machine is in operation – without the data first going to the cloud, on to the business and back again.
There has been surprisingly little talk about this. The AI conversation is mostly about copilots, chatbots and agents, i.e. software working with other software. The IoT conversation is about sensors and connectivity. The interesting thing happens where the two meet: real AI inside the physical assets that a company already has. I tend to frame it as a question – what happens when the toaster gets its own built-in AI? To me, that is the next big step for the edge.
Why has it been delayed?
For many years, true intelligence in the device required more computing power and memory than the devices had. AI therefore took center stage, and we got used to calling the detour architecture.
What has changed is undramatic but crucial. The entire chain can now run on completely standard hardware, both on powerful edge servers and on microcontrollers with a few kilobytes of memory – the kind that are already in equipment that has been in operation for years. The threshold that once made the cloud the obvious choice is much lower today, and in many cases gone.
What it means in practice
A machine that can adapt in real time to its own wear, its own route, and its own factory makes better decisions than a machine running a model trained on the average of the entire fleet six months ago. It also affects margins for companies that currently send out technicians to retrieve data, update, and fix devices on site, as much of that work disappears.
Then there’s accessibility, which means more than you might think. When the intelligence is in the device, it works even where the cloud can’t reach, like in a vehicle between two coverage areas or in a remote location with poor connectivity (and even worse budget). The cost of the cloud also means that many potentially valuable use cases never come to fruition. When AI can run directly in a toaster, they suddenly become economically viable.
It already works
In one of Boliden's mines, at a depth of about 1600 meters where the connection is too poor to send anything at all, a model is running directly on a Volvo wheel loader. It predicts gearbox failures ten to twelve hours before they occur. The model is built, deployed and updated on the machine, in collaboration with Halmstad University , RISE and AI Sweden . The advance warning is enough to drive the loader out on its own, instead of sending a technician down to a machine that has already stopped.
On Autoliv's production lines, it used to take months to move a quality control unit from one factory to another. Now it runs as a unified fleet in France, Mexico and the US.
Together with HMS Networks, real-time anomaly detection is running, catching the kind of intrusions that hide in industrial network traffic. All of this fits into seven megabytes in a device with 128 megabytes of memory, without a server and without the data having to leave the device.
Where we are going
For ten years, “doing AI” has meant building yet another thing in the cloud. The next phase is less noticeable but bigger. Intelligence is moving out into the physical operations that companies already run: the machines, sensors, and equipment that have been constantly generating data and sending it off for someone else to interpret.
The detour via the cloud was reasonable as long as the intelligence could not be found anywhere else. Now it can. For those who operate machines in the field, it is worth reviewing which decisions today await a round trip to the cloud, and how many of them the machine could just as easily make itself.
You can see the original article on IT Branschen