Physical AI funding cannot bypass deployment economics

PitchBook reports that robotics and physical AI startups raised $27.6 billion across 1,009 deals in 2025, more than doubling their haul from the previous year. Writing in The Robot Report, investor Janis Skriveris frames this capital concentration as the start of an inevitable disruption curve, arguing that general-purpose models will do for hardware what large language models did for enterprise software. That comparison conflates software distribution with physical operations. A language model can deploy globally through an API overnight. A physical agent must interact with physical surfaces, survive mechanical wear, and navigate dynamic environments where an unhandled failure halts an entire facility.

Skriveris identifies the second phase of this transition as the corporate adoption wall, where startups run into enterprise procurement and integration friction. The actual obstacle runs deeper than slow procurement cycles. Enterprise robotics pilots frequently stall because training a foundation model in simulation does not eliminate sensor drift, mechanical wear, or edge integration issues. When a language model hallucinates a summary, a human rereads the draft; when a robot arm misjudges a friction threshold on a warehouse conveyor, the line stops and human operators must intervene manually. Those operational interruptions keep maintenance overhead high enough to erase projected labor savings.

A multi-billion-dollar valuation reflects capital chasing a sector rather than verified reliability on a factory floor. Long-term advantage belongs to systems with closed operational feedback loops that make maintenance, fleet telemetry, and edge recovery cheap enough to run without constant hand-holding. Until vendors demonstrate deployment economics that hold up after the initial integration team leaves the facility, the metric that matters for buyers is operational uptime rather than funding milestones.