AI Infrastructure Race Intensifies: Custom Chips, Physical AI, and Usage-Based Billing Reshaped
인프라/플랫폼 | Tue Jul 21 2026 00:00:00 GMT+0000 (Coordinated Universal Time) | 4 sources
Google's next-generation custom AI chip development, Nvidia's physical AI ecosystem buildout in Japan, and Gemini's billing overhaul are reshaping the AI infrastructure landscape.
Analysis
[Google] developed next-generation custom AI chip 'Frozen v2' [1]
- Deployment targeted for 2028
- 6-10x power efficiency versus existing AI chips
- Optimized for Gemini models
- Strategy to reduce Nvidia dependence
[Nvidia] signed Vera Rubin AI factory and physical AI partnerships in Japan [2]
- 13
- 750 Vera CPUs and 27
- 500 Rubin GPUs
- 140MW data center
- Operations targeted for 2028
- 44 companies including SoftBank
- Sony
- NEC
- and Honda participating
[Nvidia Cosmos Coalition] unveiled Cosmos 3 Edge for Japanese robotics and manufacturing giants [2]
- Fanuc
- Yaskawa
- Kawasaki Heavy
- Fujitsu
- Hitachi participating
- On-device execution on Jetson Thor chips
- Honda R&D and Omron already using the tools
- Collaborative control system testing underway
[NVIDIA Cosmos 3 Edge] released 4B-parameter open world model on Hugging Face [3]
- Runs locally on Jetson Thor
- 4-billion-parameter world model
- Real-time reasoning and robot action generation
- Adaptable to specific robots in about a day
[Google Gemini] overhauled billing structure to compute-based usage model [4]
- Measured by compute resources rather than request count
- Four tiers: Free
- Plus
- Pro
- Ultra
- AI Plus $8
- Pro $20
- Ultra $100-$200
- Plus 2x
- Pro 4x
- Ultra up to 20x limits