Details
- Google Cloud announces NVIDIA Vera Rubin-powered A5X bare-metal instances at Google Cloud Next, scaling to 80,000 GPUs per site and 960,000 in multisite clusters for AI factories supporting agentic and physical AI.
- NVIDIA and Google Cloud partnership introduces previews of Gemini on Google Distributed Cloud with Blackwell GPUs, confidential VMs, and Gemini Enterprise Agent Platform with Nemotron models and NeMo framework.
- A5X delivers 10x lower inference cost per token and 10x higher throughput per megawatt via codesign of Rubin chips, ConnectX-9 SuperNICs, and Google Virgo networking, building on Blackwell A4/A4X VMs.
- Expands from prior Blackwell portfolio including GB200/GB300 NVL72 systems; customers like OpenAI run inference on A4X Max, while Thinking Machines Lab trains on GB300, contrasting earlier generations with massive scale-up.
- Agentic AI tools like Managed Training Clusters with NeMo RL API aid customization; physical AI via Omniverse and Isaac Sim on Google Cloud Marketplace, used by Cadence, Siemens, CrowdStrike, Snap, and 90,000+ developers.
Impact
NVIDIA-Google infrastructure accelerates agentic AI adoption by enabling hyperscale training and secure inference for labs like OpenAI, driving multi-agent systems as seen in Google Cloud Next tools like Agent2Agent protocol. This bolsters developer ecosystems with open models and confidential computing, potentially shifting R&D toward physical AI in manufacturing and robotics. Over 12-24 months, expect surged funding for sovereign AI and industrial applications, intensifying competition with AWS and Azure in GPU supply chains.
