NVIDIA and Google Cloud Partner to Advance Agentic and Physical AI

NVIDIA and Google Cloud are building on a full-stack AI partnership spanning more than a decade to bring agentic AI and physical AI into full production. At Google Cloud Next, this collaboration expands further, with several announcements strengthening Google Cloud AI Hypercomputer as infrastructure for AI factories.

The core idea is to combine next-generation infrastructure, security, open models, and industrial AI into a single stack. Newly unveiled details include NVIDIA Vera Rubin-based A5X bare-metal instances, a Gemini preview on Google Distributed Cloud, and configurations using Blackwell and Blackwell Ultra GPU.

On the infrastructure side, A5X is built on the Vera Rubin NVL72 rack-scale system, targeting a 10x reduction in cost per inference token and a 10x increase in tokens per megawatt compared to the previous generation. It is also designed, via ConnectX-9 SuperNICs and next-generation Google Virgo networking, to scale up to 80,000 NVIDIA Rubin GPUs in a single-site cluster and up to 960,000 GPUs in a multi-site cluster.

Google Cloud's NVIDIA Blackwell portfolio is broadly configured as follows.

• A4 VMs with HGX B200 systems
• A4X VMs with GB200 NVL72
• A4X Max with GB300 NVL72
• fractional G4 VMs with RTX PRO 6000 Blackwell Server Edition GPUs

This structure lets customers connect multiple NVL72 racks to scale to tens of thousands of GPUs, use up to 72 Blackwell GPUs in a single rack, or even allocate resources in 1/8 GPU increments. This flexibility supports a wide range of workloads, from mixture-of-experts reasoning and multimodal inference to data processing and complex simulations.

Security and sovereign AI are also being strengthened. Google Gemini models are being previewed on Blackwell and Blackwell Ultra GPU within Google Distributed Cloud, and NVIDIA Confidential Computing keeps prompts and fine-tuning data encrypted so that even infrastructure operators cannot view or alter them. In the public cloud, a preview of Confidential G4 VMs with RTX PRO 6000 Blackwell GPUs extends these protections to multi-tenant environments.

On the model and agent development side, the platform supports Gemini, Gemma, NVIDIA Nemotron open models, and the broader open-weight ecosystem. In particular, NVIDIA Nemotron 3 Super is available on the Gemini Enterprise Agent Platform, and Managed Training Clusters now include a new managed reinforcement learning API based on NVIDIA NeMo RL that automates cluster sizing, failure recovery, and job execution.

Real-world use cases are already emerging.

• Thinking Machines Lab is accelerating Tinker API training using A4X Max VMs and GB300 NVL72.
• OpenAI is using GB300 and GB200 NVL72 systems on Google Cloud for large-scale inference, running high-load workloads including ChatGPT.
• CrowdStrike is using NeMo Data Designer, NeMo Automodel, and NeMo Megatron Bridge to generate synthetic data and tune domain-specific cybersecurity models.

For industrial and physical AI, NVIDIA Omniverse libraries and open-source NVIDIA Isaac Sim are available on the Google Cloud Marketplace, enabling precise digital twins and robotics simulation pipelines. In addition, NVIDIA NIM microservices, such as Cosmos Reason 2, can be deployed on Vertex AI and Google Kubernetes Engine so that robots and vision AI agents can see, reason, and act in the real world.

Ultimately, this platform serves as a foundation for turning experiments into production systems across fields such as code review, security, factory optimization, drug discovery, generative imagery, and video intelligence. Companies like Snap, Schrödinger, and Salesforce, along with startups such as CodeRabbit, Factory, Aible, Mantis AI, Photoroom, and Baseten, are already using this combination, and the joint developer community has grown to more than 90,000 in just over a year.