Physical AI moves from demo floor to factory floor as robots face the real world

The key constraint is not investor enthusiasm. It is real-world data, battery life, edge chips, safety certification and the cost of deploying machines into messy industrial environments.

Humanoids are attracting the heat, but nearer-term ROI still sits in purpose-built automation, warehouse AMRs and specialized robotics systems.

The durable winners are likely to be companies with proprietary deployment data, clear labor-bottleneck solutions and Robotics-as-a-Service models that reduce upfront customer costs.

Robots face the real world

Citi’s Robotics & Physical AI Leadership Conference left a clear message: physical AI is no longer just a laboratory story or a venture-capital slide deck. It is starting to move from proof of concept toward commercial deployment. But the catch is just as important. This is not the same scaling curve as the chatbot boom. Robots do not live in clean digital sandboxes. They work in warehouses, factories, ports, trucks, construction sites and defense environments, where every inch of the real world adds friction.

The demand side is getting easier to understand. Labor shortages, reshoring, tighter supply-chain control and friendlier regulatory winds are pushing companies to automate more of the physical economy. The problem is that deploying a robot is not like releasing a new software model. It means hardware, safety certification, batteries, sensors, chips, maintenance, customer training and a data loop that has to be built in the wild rather than scraped from the internet.

That is the real distinction between digital AI and physical AI. In large language models, the base model can carry much of the value. In physical AI, the value sits much closer to the ground. Proprietary real-world data, task-specific deployment history, safety performance and the ability to solve one expensive labor bottleneck at a time matter more than a sweeping promise that robots are coming for everything.

The conference also put a hard number around the size of the bet. Roughly $20 billion has gone into physical AI over the past two years, with applications stretching across warehouses, logistics, trucking, construction, aviation and defense. Humanoids are getting the glamour bid, especially as companies like BMW test upgraded humanoid robots on factory floors. But the nearer-term return on investment still looks more likely to come from purpose-built systems, warehouse autonomous mobile robots and specialized automation platforms than from a general-purpose humanoid revolution arriving overnight.

Data remains the choke point. Even tens of millions of hours of robot data expected to be collected in 2026 may still represent only basis points of what is ultimately needed for high-level robotic performance. That line matters. It tells you this is not a trade where the market simply throws capital at the theme and waits for magic. Physical AI needs repetition, scars, edge cases and live operating history. The robot has to learn not just how the task looks in a demo, but how it fails on a bad day, in a crowded aisle, with poor lighting, low battery and a human worker standing too close.

The hardware bottleneck is just as important. Power, battery longevity and chip architecture are all constraints. Most current semiconductor platforms were built for datacenter workloads, not real-time edge inference on mobile platforms. That means the physical AI stack still needs its own industrial nervous system. The robot needs to think locally, move safely, conserve power and respond instantly. That is a very different problem from serving tokens from a server rack.

For investors, the industrial read-through is that automation-exposed companies should remain long-cycle beneficiaries. Rockwell Automation, Emerson Electric, Honeywell, Symbotic, Ralliant and Belden were highlighted as preferred exposures across pure-play automation, warehouse automation, sensors, test and measurement, and industrial networking. The logic is straightforward. If companies are going to automate more of the physical economy, the picks and shovels are not only robots. They are controllers, sensors, software, networking, safety systems and the industrial architecture that allows machines to work with less human intervention.

Robotics-as-a-Service may be one of the more important business-model shifts. This matters because the sticker price of automation can scare off small and mid-sized customers. RaaS changes the conversation from a heavy upfront capital decision into a more manageable operating-cost decision. That should help broaden adoption, especially in warehouse and logistics environments where the labor pain is obvious, the tasks are repetitive and the ROI can be measured in throughput, uptime and accuracy.

My view is that this is where the trade becomes more interesting and less obvious. Physical AI is not a meme version of robotics with a new label slapped on it. It is the slow wiring of AI into the physical economy. That makes the theme potentially bigger than the humanoid headlines, but also slower, messier and more industrial than the market wants to admit. The money will not only chase the robot on the factory floor. It will chase the companies that own the deployment data, manage the safety envelope, lower the adoption cost and turn automation from a science project into an operating tool.

The danger is extrapolating the digital AI curve onto a physical world that refuses to scale cleanly. Chatbots can improve by serving billions of prompts. Robots improve by surviving real jobs. That is a harder flywheel to build, but a more defensible one once it starts turning.