This artificial intelligence (AI) stuff is coming at us fast, isn’t it? Seems like just a year ago (oh, it was) that talk of large-language models (LLMs) was all the rage. And before I could fully wrap my head around it, I found myself interviewing experts and writing articles about agentic AI. That was just this spring. Now, I’ve spent a good part of the last month or so knee deep in physical AI trends—at conferences, in articles and on podcasts.

Quick reality check: we’ve been at this for about 50 years now. The bots and agents aren’t sneaking up on us;  they’ve just become useful enough that we need to pay attention.

What exactly is physical AI? In short, physical AI is artificial intelligence grounded in the laws of physics, chemistry and thermodynamics. The concept is different than LLMs and agentic AI, which have no inherent understanding of why, for example, a distillation column can’t violate component mass balances, or why a compressor has hard operating limits. When design limits are respected, reliability constraints can be honored, and decades of engineering standards become guardrails embedded in AI. So, it’s easy to see why many consider it a significant leap forward.

I recently sat in on a joint presentation by Honeywell Technologies’ Rahul Negi and Nvidia’s Ahsan Yousufzai at Honeywell User Group in Phoenix, and they put the benefits and the journey for physical AI in context. On AI up to this point, Negi, the company’s head of AI and industrial autonomous operations, explained, "Small step changes brought us to where we are, but they’re not going to get us that next step forward."

Yousufzai, Nvidia’s global head of business development and strategy, explained why the moment for physical AI is now. Physical AI is where process operations become the primary use case. He pointed to existing plant instrumentation such as acoustic sensors, vibration monitors, cameras and analytical instruments that become the perceptual layer for an AI that can maintain a real-time, physics-consistent model of plant states and act on them.

Get your subscription to Control's tri-weekly newsletter.

Because physical AI operates within a consistent envelope, it can anticipate upsets before alarms are triggered, rather than reacting afterward. It adapts in real-time to feed composition changes, temperature excursions and pressure disturbances. It also coordinates across the control and optimization stack.

And, Negi and Yousufzai aren’t the only ones discussing physical AI, particularly as edge technology advances. Across the globe from Phoenix, in Taipei, experts at Advantech’s Edge AI conference reported the edge has great advantages over the cloud in enabling physical AI. Miller Chang, president of embedded sector for Advantech, said the U.S. market alone for edge AI is projected to grow to $197 billion by 2034. He referenced the many diverse robotics solutions already proliferating in the market.

While all this seems exciting, there are reasons for pause. Before physical AI is adopted full steam, real-world models must be consistently accurate, and most experts agree we aren’t quite there yet. Another hindrance is the continued and persistent disconnect between operational and information technology architectures. Until these issues are solved physical AI use probably won’t take off in industrial settings.

But that doesn’t mean anyone should abandon their quest for physical AI. It offers promise and safeguards against many of the fears that keep humans awake at night these days.

We may be a half century into this AI journey, but in many ways, we’re still just getting out of the starting blocks. Autonomous operations are the ultimate destination. Physical AI just may be the enabling technology to get us there.