Artificial intelligence in machine lifecycles
My prediction: by 2032 there will be about 40 distinct applications of AI will touch the lifecycle of a typical off-highway machine.
Not one AI. Not one platform. And certainly not one “AI strategy”.
Caterpillar recently announced in Q2 2026 an expanded relationship with NVIDIA, including the use of the Blackwell-based Jetson Thor platform for on-machine AI.
It is an extraordinarily capable module. It is also currently listed at around $5K ($3K at time of Cat's announcement!), even in volumes of 1,000 units.
That may make sense for autonomy, multi-camera perception or an advanced operator assistant, and of course it has the “safety island” capability that you need to support functional safety (e.g. ISO23232). The price tag does mean that there is little chance that we will see this kind of compute in every mini excavator. Also worth remembering that powerful compute, in a physical system, only makes sense if you have a lot of data, and that means a lot of sensors. Again, this is not something that we expect in every commoditised machine.
This is not a criticism of Caterpillar. They are getting on with it, deploying technology and learning.
But I predict that the greatest impact of AI on off-highway will not come from one powerful central computer.
AI will appear throughout the machine lifecycle:
• Concept and requirements
• Engineering and simulation
• Manufacturing and quality
• Embedded inside bought-in components, modules, sensors.
• Operator assistance and machine control
• Diagnostics, service and aftermarket
Before the machine ships, much of this AI will run in the cloud.
Once the machine is working, the balance will shift towards the edge. Connectivity, latency, cost, reliability and functional safety will often make local processing essential. Cloud AI will still support model training, fleet optimisation and remote service.
Some applications will use LLMs. Others will use machine vision, deterministic models, anomaly detection or tiny neural networks monitoring a single sensor.
They will be selected by dozens of different teams and produce hundreds of different outcomes. Some, probably a minority of the model training will be done in house by OEMs, whilst the rest will be done using open-source or collaborative data sets.
Off-highway companies need direction, governance and shared infrastructure to progress in AI. But “our AI strategy” is the wrong level of abstraction.
The real question is:
“Where across the machine lifecycle could intelligence create value, and what is the smallest, safest and cheapest form of AI capable of delivering it?”
