From humanoids to multimodal models to HBM shortages, where is the Physical AI gold rush moving next? Below is an AI stack framework (from Chamath Palihapitiya), paired with recent news highlights across three key layers: deployment, intelligence, and hardware.

Deployment Layer

Humanoid Manufacturing: OpenAI-backed 1X Technologies has opened a massive factory in California with the goal of building 10,000 consumer-facing humanoids in its first year.

Humanoid Robot: Meta Platforms Acquires Humanoid Robot Startup Assured Robot Intelligence. “We acquired Assured Robot Intelligence, a company at the frontier of robotic intelligence designed to enable robots to understand, predict, and adapt to human behaviors in complex and dynamic environments,” a Meta spokesperson said in a Friday statement.

Intelligence Layer

Physical AI model: NVIDIA released the Nemotron 3 Nano Omni model on April 30. It unifies vision, audio, and language into a single system, allowing robots and agents to process sensor data 9x faster than previous multi-model setups.

Synthetic Simulation: Generalist AI unveiled its GEN-1 model, which uses synthetic data to improve robotic task success rates to 99% with only one hour of real-world training data.

Hardware Layer

Physical AI Chip: Intel hired a top Qualcomm executive on May 4 to lead a new “Physical AI” computing group, signaling a pivot toward on-device processing for robots and PCs.

Supply Chain Bottlenecks: Reports from May confirm that HBM (High Bandwidth Memory) supply is fully sold out through the entirety of 2026, with SK Hynix and Micron unable to take new orders.

Energy Inputs: Meta signed a deal on May 3 to reserve 1 GW of space-based solar power from Overview Energy, a futuristic bet to bypass constrained terrestrial power grids.