Physical AI & Opportunities for the Middle Market

Key Insights From This Episode

What is physical AI, and how is it different from generative AI?

Mitch Solomon's definition: artificial intelligence that enables machines to perceive and understand the physical world, make decisions, and independently take action within it. Generative AI produces a digital output a human can review and correct. Physical AI makes real decisions where a mistake can damage equipment or injure someone, so precision, control, and safety systems are not optional features.

There is no internet for physical AI to learn from, and that is the central bottleneck.

Generative AI arrived with decades of digitized human thought ready to train on. Physical AI has nothing equivalent, so every machine must be trained on every new task, either through simulation or through hundreds of teleoperated repetitions with tagged data. Finding the balance between what can be trained and what is commercially feasible to train is the hardest open question in the field.

The addressable market is larger than generative AI because most work is not done at a desk.

The World Bank estimates roughly one in five jobs globally is performed at a computer, leaving around three billion non-desk jobs, much of it dirty, dangerous, or dull. Add roughly 400 million workplace injuries a year, over 300,000 fatalities, and shortages of about 1.5 million healthcare workers, 600,000 manufacturing workers, and 300,000 truck drivers in the US, and the case extends well beyond labour substitution.

Separate the hype from the deployment reality.

When a humanoid robot handles a grocery bag on video, you do not know whether it was the first attempt or the ninety-ninth, or whether a human was teleoperating it with a headset and controller. For generalized tasks today, it is usually the latter. Meanwhile John Deere's See & Spray, which does one narrow task, is delivering reported 58% reductions in herbicide application and is priced on outcomes, charging by the acres you do not spray.

Near-term value favours enablers, long-term value favours whoever owns the data.

Chips, edge compute, simulation tools, and differentiated sensors are the picks and shovels, and many robotics companies will fail while still buying them. Over time value shifts to firms controlling the complete solution, the customer workflow, and the operating data. Mitch Solomon's contrarian note: a significant share may accrue not to technology suppliers at all, but to the companies deploying it most aggressively.

Diligence should test commercial viability, not technical capability.

Chris Rommel's filter: is it attached to an operational KPI the customer already cares about, such as downtime, scrap, labour, or safety incidents, and has it moved from pilot to production across different facilities and conditions? Near-term winners are focused systems, robotic arms, AMRs, vision-guided picking, rather than general-purpose humanoids. The market is voting with spend: AI development solutions growing 25% a year, edge compute above 40%, robotics revenue near 50%.