NVIDIA and Google Cloud have introduced a new set of technologies aimed at building large-scale AI factories capable of running advanced agentic and physical AI workloads. The announcement includes A5X bare-metal instances powered by NVIDIA Vera Rubin systems, along with expanded support for Google’s Gemini models operating on NVIDIA Blackwell and Blackwell Ultra GPUs. These systems are designed to handle everything from AI model training to real-time inference across demanding environments.
🔑 Key Highlights
- A5X instances scale up to 960,000 NVIDIA Rubin GPUs
- New systems deliver 10x lower inference cost per token
- Gemini models run on distributed cloud with Blackwell GPUs
- Over 90,000 developers joined joint ecosystem in one year
The A5X infrastructure brings significant performance gains, delivering up to tenfold improvements in both cost efficiency and processing throughput compared to earlier generations. These systems integrate advanced networking through ConnectX-9 SuperNICs and Google’s Virgo network, allowing deployments to scale from tens of thousands to nearly one million GPUs across multiple locations. This setup enables organizations to run large, complex AI workloads using optimized cloud infrastructure.
Alongside compute advancements, the partnership introduces new capabilities for secure AI deployment. Gemini models are now available on distributed cloud environments powered by Blackwell GPUs, allowing organizations to operate AI systems closer to sensitive data. Confidential computing ensures that prompts, models, and training data remain encrypted and inaccessible to unauthorized users, even within shared cloud environments.
The platform also expands support for agentic AI development through open models and APIs. NVIDIA Nemotron models are integrated into the Gemini Enterprise Agent Platform, enabling developers to build systems that can reason, plan, and execute tasks. Additional tools, such as managed training clusters and reinforcement learning APIs, simplify large-scale model training while reducing operational complexity.
Beyond digital applications, the collaboration targets industrial and physical AI use cases. Tools like NVIDIA Omniverse and Isaac Sim allow developers to create digital twins and simulate robotic systems before real-world deployment. These capabilities support industries ranging from manufacturing and aerospace to robotics and autonomous systems, enabling companies to design, test, and optimize operations in virtual environments before implementation.
📊 What This Means (Our Analysis)
This development signals a clear shift from isolated AI tools to integrated systems that can operate at industrial scale, combining compute, models, and deployment into a single platform.
By unifying infrastructure with agentic and physical AI capabilities, the partnership positions enterprises to move faster from experimentation to real-world execution.
📌 Our Take: AI is no longer confined to models—it is becoming the operational backbone of how systems are built and run.