Physical AI: When Intelligence Leaves the Screen
Embodied systems alter the economics of labour, safety and supply chains in ways that language models do not. What to plan for now, and what can safely wait.
A different cost curve
Software intelligence has a marginal cost approaching zero. Physical intelligence does not. Every deployment involves hardware that wears, environments that vary, and safety cases that must be argued to a regulator rather than demonstrated in a video.
This is why physical AI diffuses more slowly than software AI and why its effects, once they arrive, are more durable. A capability that takes five years to deploy also takes five years for a competitor to copy.
Where it lands first
Adoption is concentrating where three conditions hold together: the environment is structured enough to be predictable, labour is scarce or hazardous, and the cost of an error is contained. Warehousing, container handling, inspection, agricultural harvesting and certain classes of construction all satisfy this.
General-purpose humanoid robotics attracts far more attention and satisfies none of the conditions particularly well. It may eventually matter enormously. It is not where the next five years of realised value sits, and planning assumptions should reflect that.
The labour question, stated honestly
The displacement debate is usually conducted at a level of generality that makes it useless. The specific pattern in physical AI is that tasks are automated before jobs are, and that the tasks automated first are disproportionately the physically damaging ones.
The policy consequence is that the transition is manageable if it is anticipated and severe if it is not. Nations with functioning vocational retraining systems and portable benefits will experience this as a productivity gain. Nations without them will experience the same technology as a shock, and will be tempted to respond by obstructing it — which forfeits the productivity gain without preventing the displacement.
The window for building that capacity is now, while deployment is still concentrated in a small number of sectors.
What to do this year
For manufacturers: instrument the processes you would eventually want to automate, because the data requirement precedes the deployment by years. For ministries: resolve the safety-certification pathway, since regulatory ambiguity is currently a larger brake on adoption than technical capability. For everyone: separate the humanoid narrative from the deployment reality when allocating capital.
Physical AI is arriving on an industrial timescale rather than a software one. That is not a reason to ignore it. It is a reason to prepare on the same timescale.