Over the past eighty years or so, computing has shifted back and forth between centralisation and decentralisation. It went from mainframes to personal computers, before swinging back towards centralisation with the cloud and hyperscale data centres. AI has intensified this latest phase.
Training frontier models requires huge clusters of accelerator chips, networking, cooling and power. Five large tech companies spent more than $400bn in capital expenditure in 2025, with that figure expected to rise another 75 per cent in 2026. Data centre electricity use rose 17 per cent last year, with AI-focused data centres growing faster still.
But training a model and using one are different problems. For many applications, both happen in centralised data centres. But as AI moves into the physical world, sending data back to the cloud for processing can become impractical. For a robot working underground in a mine, the problem might be connectivity. For an autonomous vehicle processing real-time sensor data, it could be latency. A medical robot may face privacy and reliability requirements, while a drone has severe constraints on size, weight and power.
These constraints push computation in the opposite direction: towards the edge. Rather than sending everything back to a data centre, some inference happens locally, closer to where the data is generated.
This matters because generative AI has so far focused largely on cognitive tasks, while physical AI extends automation into physical work. This brings AI into contact with enormous pools of expenditure on labour, machinery, transport and industrial capital. While much of the current investment cycle is focused on building giant centralised computing factories, the potential endpoint is billions of intelligent machines operating in the physical world.
Traditional industrial robots excel at repetitive, programmed tasks in controlled environments. The familiar image is of robotic arms welding car bodies, sparks flying, on automotive factory floors. The next phase of AI, where it enters the physical world, promises something more difficult: machines that can perceive their surroundings, make decisions and adapt their behaviour as conditions change.
The scale of the opportunity is huge. There were already 4.7m industrial robots operating globally in 2024, with another 542,000 installed during the year, more than twice the annual installations of a decade prior. Toyota alone estimates that it and its major suppliers could need around 400,000 robots as it modernises its factories, with annual automation investment potentially reaching $6.4bn from 2028. That is just one industrial ecosystem. Transportation, defence and security, agriculture, construction, logistics, mining and healthcare all involve enormous amounts of physical work that could increasingly be automated.
However, the physical world is much less accommodating than a data centre. Industrial robots, military drones, autonomous vehicles and environmental sensors operate in different environments and perform very different jobs. Each has different constraints on energy, size and weight, and a reliable connection to the cloud cannot always be assumed. That creates demand not just for smaller chips, but for more efficient ways of processing and storing information, and for AI models designed around these constraints.
Several UK companies are working on this problem. Sheffield-based Opteran is developing machine autonomy inspired by insect brains, designed to run entirely at the edge on relatively modest silicon. Newcastle-based Literal Labs is taking a different route, developing logic-based models that use substantially less compute, memory and power than conventional neural networks. London’s Intrinsic Semiconductor Technologies is attacking the hardware itself, developing memory that can be integrated directly onto advanced processor chips, reducing the amount of energy-intensive data movement between the processor and memory.
The technologies vary, but the problem is essentially the same: how do you put useful intelligence into a machine when you can’t simply give it more computing power?
None of this means the data centre becomes less important. Training frontier models will continue to require enormous concentrations of computing power. However, deploying the resulting intelligence across the physical economy creates a different set of constraints for each application, and therefore a different set of technological and investment opportunities. In many cases, the hardware and software will need to be designed around the problem being solved rather than developed in isolation and matched to an application afterwards. The AI infrastructure build-out may therefore be highly centralised at one end and increasingly distributed at the other.
Hundreds of billions are being spent building the infrastructure needed to concentrate intelligence in data centres. I think that putting intelligence into the physical world may prove an even more interesting technical challenge, and an enormous multi-decade commercial opportunity.