NVIDIA and Google Cloud announce collaboration for agentic AI and physical AI on shared infrastructure

Hardware and cloud in one package

NVIDIA and Google Cloud announced a collaboration that combines NVIDIA GPU infrastructure with Google Cloud platform services for agentic AI and physical AI workloads. The goal is to offer enterprise clients one integrated offering rather than requiring them to manually connect GPU clusters with cloud applications.

The collaboration covers the entire stack — from low-level GPU infrastructure (H100, B200, and the Blackwell generation), through the CUDA software layer, to Google Cloud managed services for training and inference. For clients this means less operational work, since they don’t have to perform capacity planning and optimization themselves.

Focus on two new AI categories

The term “agentic AI” describes autonomous software agents that independently execute tasks. Rather than a user asking questions of a chatbot, the agent receives a goal (for example “book a trip to Vienna”) and independently calls the necessary services. Such agents require more inference runs per task and greater latency sensitivity.

“Physical AI” is a newer term that NVIDIA is actively promoting. It refers to AI that controls physical systems: robots, autonomous vehicles, industrial lines, and drones. Training physical AI models requires realistic simulations — NVIDIA Isaac and Omniverse platforms generate synthetic data, while Google Cloud provides scalable infrastructure for those simulations.

Competition with AWS and Microsoft

The collaboration is a strategic response to existing arrangements by the competition. AWS has a strong relationship with NVIDIA through EC2 GPU instances and Bedrock, while Microsoft Azure has exclusive agreements with OpenAI. Google Cloud has traditionally been losing market share in generative AI due to the positioning of its own TPU chips.

With this move, Google Cloud is sending a clear message that NVIDIA GPUs have first-class support on its platform, on par with Google’s TPUs. Clients already using the CUDA ecosystem no longer have to run to AWS or Azure to get NVIDIA performance.

What it means for practitioners

For development teams working on agentic solutions or robotics, the offering means simpler access to the hardware + platform combination. Managed services cover the most demanding part — GPU cluster configuration, networking between nodes, distributed training frameworks — so teams can focus on the model and business logic.

Specific details on products and pricing will be announced in the coming months through Google Cloud Next and NVIDIA GTC events. Early access will be available to selected enterprise partners.