Nvidia and Google Cloud are expanding cooperation to build hyperscale infrastructure for next-generation agentic AI and physical AI. The companies plan to expand the Google Cloud AI Hypercomputer and offer A5X bare-metal instances based on Nvidia’s Vera Rubin platform. Nvidia said A5X can scale to 80,000 GPUs at a single site and up to 960,000 GPUs across multi-site clusters, while improving throughput per watt and reducing inference costs per token.

Nvidia said it will shift competition in AI semiconductors from chip specifications to end-to-end efficiency across pre-training, post-training, inference and agents. It also disclosed measured results showing Blackwell-based GPUs deliver 55 times faster mixture-of-experts inference than the prior Hopper generation. Nvidia highlighted efficiency gains from a new numeric format and software advances, and introduced curriculum-based post-training results for its Nemotron 3 Nano model. It also said Korean partners are participating and unveiled a Korean-focused synthetic dataset.