For decades, "smart automation" promised to do for factories and warehouses what software did for the office: strip out repetitive human work and replace it with something faster, cheaper and tireless. 

It never quite delivered on that promise at scale. 

Now Physical AI - AI that perceives and acts in the real world rather than just processing text and pixels - is being pitched as the technology that finally closes the gap. Is it? 

I think the answer is yes, but the road there is longer, harder and more expensive than most people appreciate, and understanding why, matters if we want to actually get there.

Why Physical AI is different — and why it's the missing piece

Traditional automation struggled for years against a stubborn reality: industrial environments are wildly varied, and so are the processes running through them. 

A human operator can walk into a new site, become operational in minutes, adjust to a layout change overnight, and do five different jobs in a single shift. 

A traditional machine can't. It is brilliant at one repetitive task in a controlled environment, and nearly useless the moment conditions change. 

That's the core reason so much of materials handling is still done by people driving forklifts rather than by robots.

Physical AI is the first real candidate to close that gap. It combines human level visual perception, understanding of a dynamic 3D world with computer vision and AI algorithms, and AI powered decision-making that was trained on real and simulated data and can adapt on the fly. 

These are the same qualities that make a human operator valuable, but embedded in a machine that doesn't get tired, distracted or hurt. 

That's why I believe it has the potential to be the next industrial revolution: it targets the exact limitation that kept automation boxed into narrow, controlled use cases for the last thirty years.

The gap between promise and reality

Here's the tension. We've watched AI transform the digital world in a matter of months - a new model ships, and overnight millions of people are using it. Physical AI hasn't moved at that speed, and the contrast has made plenty of people question whether it will ever be commercially viable.

The numbers back up the scepticism. The last few years have seen a steady stream of casualties among robotics companies and projects. Rethink Robotics (twice), READY Robotics, Bossa Nova Robotics, Dorabot, Small Robot Company, Kivnon, RoboTire and others have shut down, been sold off cheaply, or filed for bankruptcy, often after raising tens of millions of dollars. 

Even category leaders like iRobot have filed for Chapter 11 bankruptcy (in December 2025). 

At the same time, real deployment is happening. Market trackers estimate the installed base of AI-enabled autonomous mobile robots in warehouses is now in the hundreds of thousands and climbing fast. 

So the picture is genuinely mixed — real, growing deployment sitting right next to a long casualty list. That's not a contradiction; it's the signature of a technology that works, but whose business of building and selling it is still brutally hard.

Why the gap exists

The gap is explained less by whether the technology works and more by how long, and how expensive, it is to prove that it works commercially. 

A typical software start-up can find product-market fit in 12 to 18 months. A Physical AI company usually cannot: reaching a customer is hard, and even once you do, the path from pilot to revenue can take three to four years, with heavy R&D investment required the whole way to meet the reliability and completeness a real deployment demands. 

At best, maybe half of those efforts convert into revenue. Meanwhile, funding for start-ups in this space is scarce outside the earliest rounds or before crossing the USD10 million mark, leaving a brutal middle stretch where many companies simply run out of runway. 

Large automation incumbents face a mirror-image problem. They already have the customer relationships and don't depend on outside capital, so they can absorb a long time-to-revenue. 

But their margins are typically thin, and their investors increasingly question the ROI of big innovation bets like deploying AI at scale - especially in markets where there's no urgent competitive threat forcing the issue. 

Contrast that with defence, a traditionally conservative sector with very long cycles for innovation adoption, where AI-enabled autonomy has been adopted remarkably fast. 

When there's urgency, there's speed - and there's no greater urgency than protecting your people from physical harm.

How the gap gets bridged

I see three shifts that can close it. 

First, funding patterns are already changing: a new generation of start-ups born after the GenAI wave is raising seed rounds above USD50 million (often in the hundreds of millions), because they understand the runway must last until repeatable revenue, not just a demo. 

Second, the industry needs investors who judge these companies on a different type of economics - total contract value (TCV) and potential TCV, not ARR (annual recurring revenue). And with different time horizons - 10–15 year instead of three to five. And with a different goal - becoming large, profitable companies, not sale at a 10x multiple. M&A in this space rarely happens at strategic multiples, and the list of acquirers is short, so the old venture playbook doesn't fit. 

Third, the market itself needs to become more competitive. I expected Chinese manufacturing - often cheaper at increasingly better quality - to force that competition faster than it has. I don't have a clean answer for how to accelerate it but the industry needs a forcing function to move.

Physical AI's fundamentals are sound. What it needs now isn't more proof that it works — it's an industry, and a funding model, patient enough to let it prove that it pays.