From Code to Muscle: Physical AI and Data Center Infrastructure

At CES 2026, the spread of humanoids emerged as a key topic alongside Physical AI. AI that used to only generate text or images is moving into a stage where it understands the physical world and acts within real environments.

The core of Physical AI consists of three elements.

• Physical understanding: reflects real-world laws such as gravity, friction, and acceleration.
• Sensor-based perception: turns surroundings into real-time data using cameras, LiDAR, and tactile sensors.
• Autonomous execution: calculates motor movements to fit the situation rather than following pre-programmed trajectories.

So if Physical AI is the brain, humanoids are the most general-purpose body implemented on top of it. Their strength lies in being able to make direct use of rack structures and work paths designed around human standards, while also being able to enter high-heat, high-power environments.

Data centers are also affected by this shift. AI is becoming an intelligent infrastructure that optimizes cooling and predictive maintenance, while at the same time operational automation is progressing as robots take on on-site inspection and operations.

• Google applied DeepMind AI to cooling control, reducing cooling energy use by up to 40% and lowering overall PUE by 15%.
• NTT Data x ugo stated that it performs 24/7 unmanned monitoring using 4K cameras and sensors for temperature, humidity, airborne substances, sound, and thermal imaging, and plans to deploy ugo mini at 15 sites to cut inspection time by up to 80%.

The adoption roadmap has three stages.

• Observation: accumulates data through patrolling, thermal imaging, and noise logging without judgment or action.
• Assistance: reduces human movement by carrying tools and small parts, guiding procedures, and recording work.
• Substitution in hazardous spaces: enters high-heat, high-power zones first and relays smoke, gas, and temperature anomalies.

Technically, ultra-low-latency communication and Edge AI are essential, and by NVIDIA's Omni Physics standard, the control loop must maintain at least 100Hz (10ms) or higher. Also, as Goldman Sachs has forecast, AI data center power demand could grow by more than 160% by 2030, meaning thermal management must be solved in tandem.

Sim-to-Real, where training happens first in a virtual environment before being transferred to an actual robot, is also important. NVIDIA Isaac Lab and NVIDIA Research's Learning Physical Skills in Simulation (2025) are presented as a path to safely learning complex cabling and unstructured environments. Ultimately, kt cloud's high-density AIDC reads as both a space where humans face limits and a proving ground where Physical AI can demonstrate its strengths.