It’s funny, this physical AI stuff. Like, funny-complicated – right? At writing, Siemens is giving a press conference at Hannover Messe in Germany, which is all about industrial AI – from industrial design to supply chains to production work. It has a new “entry-level” industrial computer with an Nvidia GPU inside, geared for physical AI in factories. It has a new industrial AI agent, called Eigen, that goes inside engineering systems to understand projects, write code, configure systems, and “iterate until pre-defined performance benchmarks are met”. 

This is after writing, quite quickly, an interview with NTT Data that strips industrial AI back to its mathematical roots. NTT’s argument is corrective, actually – that factory-grade physical AI is not built on hallucinatory token-hungry language models but on bounded and deterministic maths and physics, with a narrow remit just to direct robot traffic. In that context, GPUs are surplus to requirements; CPUs suffice because robots solve for precision, not probability. The implication is that much of the current AI hardware narrative is misapplied to industrial automation.

But both arguments are correct. Siemens just reframes the discussion, and takes it further. In Hanover, Cedrik Nieke, in charge of its digital industries business, is talking about AI like a “brain in a jar” – powerful, but isolated and disconnected. Its value only materialises when embedded into industrial systems – like electricity was transformative only when it was applied to trains and factories, and all the other places his dynamo forbears at the German juggernaut-firm found for it 180 years ago. The challenge with AI is not model size but the system integration.

As an aside, the message for Europe, squeezed between the fast-digitalizing industrial economies of China and the US, is to make the applications – to apply AI in industrial applications. Because the models will be commoditized, and the value will be to make them work. It has a fight on its hands, all the same – but it has some of the biggest automation companies, and an urgent agenda to get it right. The point is that the Siemens is bigger; it is taking the industrial AI story way upstream, above the private 5G network running the humanoid robot. 

Its talk, here, about engineering agents and digital twins says that complexity is migrating away from the shop floor into design, orchestration, and lifecycle management – where AI is less about control loops and more about coordinating workflows, automating tasks, making data usable. Industrial data is vast, siloed, and untapped, says Nieke; that’s the bottleneck, he says – not guessing games about where to put GPUs. He has an analogy about all the manufacturing data in the world – 1.9 zettabytes, apparently, the equivalent of 36 million years of Netflix movies.

Like Netflix needed to be named. The thing is, 80 percent of these industrial flicks are never seen – and so maybe name-drop is right. “PLM silos, ERP silos, CRM silos,” he says. This is the layered view of industrial AI. At the edge, it looks like IoT-plus: efficient, local, and CPU-driven. Above it, it becomes agentic and system-wide. Private 5G, in case we’re still interested, is just connective tissue to bind these layers at the edge – into something that finally works, on the shop floor.