AI infrastructure is entering a new phase.

Over the past 24 hours, three developments point to a major shift: capital is moving beyond training massive models and toward inference, robotics, and real-world AI deployment.

• Sony + TSMC make a $6.32B physical AI bet: The companies are reportedly planning a 1 trillion yen joint venture in Kumamoto, Japan, focused on next-generation chips and advanced image sensors for robotics, spatial vision, and physical AI systems.
• Inference is becoming the bigger cloud opportunity: Gartner now projects AI-optimized IaaS spending will jump 96% in 2026 to $42B, with inference accounting for 55% of spending—overtaking model training as enterprises deploy AI agents and applications at scale.
• Apple tests China’s CXMT memory chips: As memory manufacturers prioritize lucrative HBM for AI servers, Apple is reportedly evaluating CXMT DRAM to secure supply for its consumer devices.

Why it matters:
The AI boom is moving downstream.

The first phase was about training bigger models. The next is about running those models everywhere—inside enterprise workflows, consumer devices, factories, robots, cameras, and autonomous systems.

That means the biggest infrastructure opportunities may increasingly shift from training compute to inference and physical AI hardware.