NVIDIA and Google Cloud have maintained a collaborative relationship for over a decade, co-engineering a comprehensive AI platform that spans every layer of the technology stack. This foundation allows developers and enterprises to transition agentic and physical AI from experimental labs into production environments. At Google Cloud Next in Las Vegas, the partnership announced significant advancements designed to power the next generation of AI factories. These updates include new bare-metal instances powered by the NVIDIA Vera Rubin architecture and a preview of Google Gemini running on NVIDIA Blackwell GPUs. The collaboration also introduces confidential VMs and agentic AI capabilities on the Gemini Enterprise Agent Platform, leveraging NVIDIA Nemotron open models and the NeMo framework.
Optimizing Inference with Vera Rubin Architecture
The core of this infrastructure upgrade is the introduction of A5X instances powered by the NVIDIA Vera Rubin NVL72 rack-scale systems. Through extreme codesign across chips, systems, and software, this architecture delivers up to 10x lower inference cost per token and 10x higher token throughput per megawatt compared to the prior generation. For cloud engineers managing large-scale inference workloads, this represents a critical efficiency gain. The system utilizes NVIDIA ConnectX-9 SuperNICs combined with next-generation Google Virgo networking, scaling to support up to 80 GPUs per rack. This configuration is particularly relevant for professionals preparing for cloud architecture certifications, as it demonstrates how hardware-software co-design directly impacts operational metrics like cost-per-token and energy efficiency. Engineers working on Kubernetes clusters or containerized AI services will find the networking improvements essential for reducing latency in distributed training and inference pipelines.
Deploying Gemini on Distributed Cloud
Google has also provided a preview of Google Gemini running on Google Distributed Cloud, specifically optimized for NVIDIA Blackwell and NVIDIA Blackwell Ultra GPUs. This deployment model allows organizations to run large language models closer to their data sources while maintaining the flexibility of a distributed architecture. For DevOps professionals, this implies a shift in how they manage model serving across hybrid environments. The integration of NVIDIA NeMo frameworks enables developers to fine-tune and deploy models more efficiently. This setup is ideal for enterprises requiring high-performance computing for digital twins on the factory floor or complex workflow management agents. The ability to run these models on confidential VMs adds a layer of security crucial for industries handling sensitive data, aligning with security-focused certification requirements.
Agentic AI and Enterprise Integration
The partnership extends to the Gemini Enterprise Agent Platform, which now supports agentic AI using NVIDIA Nemotron open models. These agents are designed to manage complex workflows autonomously, a capability that is rapidly becoming standard in enterprise automation. For AI engineers, the availability of open models like Nemotron provides a flexible starting point for building custom agents without being locked into proprietary ecosystems. This flexibility is a key consideration for professionals studying for AI engineering certifications, as it highlights the importance of model interoperability. The platform supports both agentic AI and physical AI, enabling robots to interact with the physical world while digital twins simulate factory operations. This dual capability allows for a seamless transition between simulation and real-world deployment, a critical skill for modern infrastructure engineers.
What This Means For You
For cloud engineers and DevOps professionals, this collaboration signals a maturation of the infrastructure required to support agentic and physical AI at scale. The move from experimental prototypes to production-grade systems requires a deep understanding of both hardware constraints and software optimization. The introduction of Vera Rubin and Blackwell GPUs necessitates updates to deployment strategies and resource allocation models. Professionals should review their current architectures to ensure compatibility with these new standards. Understanding how to leverage these specific hardware features, such as the SuperNICs and Virgo networking, will be essential for maintaining competitive efficiency. As the industry moves toward AI factories, the ability to optimize inference costs and throughput will define the success of large-scale deployments. Engineers should focus on mastering the integration of these new components into existing cloud environments to stay ahead of the curve.
