The AI infrastructure conversation has been dominated by centralised compute: hyperscale data centers, GPU clusters, grid connections measured in hundreds of megawatts. Physical AI changes this picture fundamentally. When AI moves from the cloud into the physical world — into robots, humanoids, autonomous vehicles and industrial automation systems — the power infrastructure challenge becomes distributed, local, and architecturally different from anything the data center industry has built.

WHAT PHYSICAL AI ACTUALLY IS

Physical AI refers to AI systems that perceive, reason about and act in the physical world. Unlike language models that process text and generate responses, physical AI systems must interpret sensor data in real time and translate it into physical actions — often in environments where a 100-millisecond response delay is unacceptable.

Physical AI Categories — Power Infrastructure Implications Industrial robots: fixed location, high power, continuous operation
Humanoid robots: mobile, battery-dependent, charging infrastructure critical
Autonomous vehicles: mobile, en-route charging, depot infrastructure
Edge inference devices: distributed, low individual power, massive aggregate
Smart manufacturing: embedded in factory electrical infrastructure

Companies like Figure, Boston Dynamics, Tesla (Optimus), 1X Technologies and Sanctuary AI are developing humanoid robots that require not just compute infrastructure but physical charging and power management infrastructure at industrial scale.

THE DISTRIBUTED POWER CHALLENGE

A data center concentrates 100–500 MW of electricity demand at a single point. Physical AI distributes power demand across thousands or millions of endpoints. This creates infrastructure challenges that are not solved by building large substations — they require a different architecture entirely.

Industrial robot charging infrastructure

A modern automotive factory deploying hundreds of AI-enabled robots requires significant upgrades to its internal electrical distribution. The power density of robot charging zones — where multiple robots charge simultaneously between shifts — can create local demand spikes that exceed the original factory electrical design. Industrial facilities deploying physical AI at scale are discovering that their existing electrical infrastructure, including internal transformers and switchgear, was not designed for this use case.

Humanoid robot deployment at scale

Tesla's stated ambition is to produce millions of Optimus humanoid robots annually. Each robot requires regular charging. A facility deploying 1,000 Optimus robots — a plausible near-term scale for a large warehouse or factory — requires charging infrastructure equivalent to a small data center in terms of power management complexity, if not raw wattage.

Autonomous vehicle depot infrastructure

Electric autonomous vehicle fleets return to depots for charging. A depot of 500 autonomous trucks, each requiring 200 kW of charging capacity, creates a 100 MW demand peak — requiring the same HV transformer infrastructure as a medium-sized data center, but at a location that was previously a parking lot with minimal electrical infrastructure.

THE EDGE COMPUTE LAYER

Physical AI requires local compute at the point of action — the latency of sending sensor data to a cloud data center and waiting for a response is incompatible with real-time physical decision-making. This creates a distributed compute infrastructure requirement: edge AI processors embedded in or adjacent to physical AI systems.

Nvidia's Jetson platform, Ambarella's AI vision processors, and Ouster's intelligent LiDAR systems represent the hardware layer of this edge compute infrastructure. Each individual device consumes modest power. Deployed at the scale that physical AI implies — millions of robots, billions of edge devices — the aggregate power infrastructure requirement is significant and largely unmodelled.

Edge AI Power — Aggregate Scale Individual edge AI device: 10–100 watts
1 million deployed devices: 10–100 MW aggregate
10 million devices (2030 scenario): 100 MW–1 GW
Infrastructure requirement: distributed grid upgrades at thousands of locations
vs data center model: single large connection at one location

WHY THIS MATTERS FOR GRID INFRASTRUCTURE

The physical AI power infrastructure challenge is different from the data center challenge in a critical way: it cannot be solved by finding a single large site and connecting it to a 225kV substation. It requires upgrading thousands of medium-voltage connections at industrial sites, logistics facilities, transport depots and urban locations.

This distributed upgrade wave will stress medium-voltage switchgear and distribution transformer supply chains — the same equipment that is already under pressure from data center demand. The switchgear shortage that we have previously identified as an underreported constraint will be compounded by physical AI deployment demand hitting the same supply chains from a different direction.

THE FRANCE PHYSICAL AI INDUSTRIAL OPPORTUNITY

France's industrial base — automotive (Stellantis, Renault), aerospace (Airbus, Safran), manufacturing — is an early adopter of physical AI automation. Renault's ElectriCity complex in northern France is an example of a facility integrating advanced automation that requires grid infrastructure upgrades beyond its original design.

Industrial facilities in France deploying physical AI at scale will need power infrastructure assessment and upgrade services. This is a direct extension of the grid readiness analysis that GridReadiness already provides for data center projects.

WHAT TO WATCH