As enterprises transition to continuous, multi-step agentic workflows, traditional infrastructure is hitting its limits. But raw performance is only part of the equation. To truly scale agentic AI, enterprises also need the flexibility to run the right models for the right tasks, alongside the data and operational sovereignty required to maintain verifiable control over both their sensitive intellectual property and the underlying infrastructure.
Today at NVIDIA GTC Berlin 2026, we are announcing new advancements across the stack to deliver this foundation: minimizing bottlenecks, expanding ecosystem choice, and introducing new sovereign AI and Confidential Computing capabilities to protect your most sensitive workloads with hardware-enforced security.
Architecting for the agentic era
To help you train, serve, and operate models at scale, Google Cloud’s AI Hypercomputer is built on compute, storage, networking, and software that we co-designed as a single system. Today, in collaboration with NVIDIA, we are announcing new updates to this agent-native infrastructure:
-
Accelerated computing with A5X instances based on NVIDIA Vera Rubin platform: These bare-metal GPU-accelerated instances feature the latest NVIDIA Vera Rubin NVL72 platform to deliver performance at scale. With general availability, coming next month, we are also expanding our infrastructure footprint to support EMEA data residency and sovereignty requirements with availability in new local regions.
-
Scale-out A5X networking with Virgo Network: Virgo Network supports a non-blocking intra- and inter-building scale-out fabric via mesh topology. This new capability will enable extreme horizontal scaling, allowing clusters to expand to 80,000 Rubin GPUs within a single site and up to 960,000 Rubin GPUs across multi-site deployments.
-
Storage with Managed Lustre and NVIDIA GPUDirect Storage: Google Cloud Managed Lustre will support NVIDIA GPUDirect Storage on A5X instances, giving training and inference data a direct RDMA path from the parallel file system into GPU memory, bypassing the CPU and system memory.
-
Long-context memory with KV Caching: A5X instances will also support advanced KV-caching capabilities designed to scale context memory beyond GPU high-bandwidth memory (HBM) for long-context and multi-turn agentic workloads. This high-speed caching layer will offload and reuse context across storage tiers, reducing redundant computation, lowering time-to-first-token, and improving inference goodput per dollar.
-
Workload orchestration with A5X instances on Google Kubernetes Engine (GKE): GKE brings a suite of features for distributed AI workloads, such as GKE Managed DRANET, which simplifies RDMA management by making network interface cards first-class native Kubernetes resources; and Kueue, which boosts performance of distributed workloads through topology-aware scheduling and resource management. This makes GKE and A5X instances a powerful combination for agentic workload orchestration.
“The true promise of agentic AI isn’t just speed, it’s freeing up human creativity from the friction of modern production pipelines. With Google Cloud’s A5X instances accelerated by NVIDIA Vera Rubin NVL72, our WPP Creative and Production teams can direct intelligent agents that perform frame-accurate quality checks and adapt brand assets seamlessly. Google Cloud gives us the reliability, scale and sovereign security required to meet our clients’ most ambitious creative demands.” – Stephan Pretorius, Chief Technology Officer, WPP
Expanding open models and ecosystem innovation
Agentic workflows rarely rely on a single model. They route tasks to the most efficient model for the job. Your infrastructure should expand your options rather than restrict your roadmap. With Google Cloud, you can deploy infrastructure on your choice of software and NVIDIA AI infrastructure across multicloud, hybrid, and highly regulated environments. Today, we announced several updates to give you the freedom to build and adapt quickly:
-
Open models: We are adding support for models like Gemma, Mistral, NVIDIA Nemotron, and others to our Sovereign Cloud portfolio. This allows you to run and govern open-weight models, while retaining full custody of model weights and runtimes, avoiding lock-in while maintaining high-performance acceleration.
-
NVIDIA Cosmos 3 and NVIDIA Nemotron 3.5 SuperVL with NVIDIA NeMo now available on Gemini: You can now customize the new NVIDIA Nemotron 3.5 SuperVL model with Managed Training Clusters. Use an MTC-optimized training recipe to jumpstart your Nemotron post-training work with NVIDIA NeMo RL and NVIDIA NeMo Gym on Google Cloud. With NVIDIA Cosmos 3, you can generate virtual worlds where robots and autonomous vehicles can try, fail and improve before real-world deployment, supporting simulation-to-reality workflows.
Securing deployment with Sovereign AI and Confidential Computing
As agents take on more autonomous, mission-critical tasks, the data they process becomes increasingly sensitive. Organizations deploying these advanced workloads require hardware-enforced protection and operational control that don’t sacrifice innovation. Whether safeguarding intellectual property during model training in the cloud or executing mission-critical workloads in fully air-gapped environments, Google Cloud and NVIDIA help provide verifiable security and local custody.
Today, we are announcing new Sovereign AI capabilities for on-premises environments, alongside advancements in Confidential Computing for the public cloud:
Sovereign AI on-premises with Google Distributed Cloud
For mission-critical workloads that must remain or strictly isolated from the internet, Google Distributed Cloud (GDC) brings a managed cloud stack directly to your data center or edge facility. To advance sovereign AI initiatives, we are delivering new compute accelerators, multimodal models, and turnkey search capabilities across GDC air-gapped deployments:
-
NVIDIA Blackwell Ultra GPUs on GDC: Following our preview announced earlier this year, the latest NVIDIA HGX B300 systems are now generally available (GA) on GDC. This brings industry-leading compute density and energy efficiency on-premises, providing the high-throughput platform foundation that demanding local inference workloads require.
-
Gemini Flash models on GDC: We are expanding our sovereign model catalog with Gemini 3.5 Flash on GDC, accelerated by NVIDIA HGX B300. Organizations can now execute rapid, high-throughput multimodal reasoning, spanning text, audio, video, and imagery, entirely inside their air-gapped perimeter.
-
Enterprise search capabilities on GDC: Available in preview, our enterprise search capabilities with Gemini on GDC air-gapped delivers an out-of-the-box conversational interface powered by multimodal retrieval-augmented generation (RAG). These capabilities are powered by Gemini on GDC air-gapped with NVIDIA HGX B300. Now, analysts and knowledge workers can instantly query, summarize, and cross-examine proprietary documents, media, and unstructured archives without custom engineering, keeping all indexing, embeddings, and interactions strictly within your sovereign security boundary.
Beyond air-gapped on-premises deployments, you can also run sovereign AI workloads on NVIDIA AI infrastructure using Google Cloud Data Boundary to enforce controls such as regional data residency on our public cloud, or Google Cloud Dedicated for an isolated cloud run by a local independent partner such as S3NS in France or Thales in Germany.
Confidential Computing in the public cloud
As enterprise workloads scale, safeguarding highly sensitive intellectual property during active training, and user context and model IP during real-time inference is critical to maintaining a secure AI architecture. By combining Google Cloud’s advanced security runtimes with NVIDIA Confidential Computing, organizations can deploy high-performance generative AI models inside verified and trusted execution environments without sacrificing processing speed.
-
G4 instances with Confidential Compute are GA: We are expanding access to privacy-preserving AI with Confidential G4 instances enabled by NVIDIA RTX PRO 6000 Blackwell Server Edition GPUs. G4 Confidential GPUs, with NVIDIA Confidential Computing provide upgraded security without compromising on performance.
-
Confidential Computing is available today on A4 instances enabled by NVIDIA HGX B200 and Intel TDX
-
We are introducing Confidential Bare Metal to Google Cloud, recognizing the value of single-tenant bare-metal performance and isolation for frontier model AI serving. Our innovation leads to A4X Max instances (enabled by NVIDIA GB300 NVL72) using NVIDIA Confidential Bare Metal. By extending hardware-rooted secure boot and attestation with encrypted DDR and NVLink interconnects, bare-metal AI workloads leverage this trusted execution boundary to protect user prompts and models during inference and training.
The Google Cloud advantage
Our co-engineering collaboration with NVIDIA integrates their full-stack accelerated computing platforms directly into AI Hypercomputer. This gives you the agent-native infrastructure you need to train, serve, and operate models at scale. Today’s announcements deliver this unified foundation, combining the NVIDIA Vera Rubin platform and high-throughput networking with expanded support for open models.
Together, we also give you verifiable control over your data. With new sovereign AI deployments and confidential computing across NVIDIA infrastructure, you can protect your most sensitive workloads using hardware-enforced security and local data residency.
Because we design hardware and software together as a single system, you can run models across different architectures without rewriting code. Your compute cycles stay focused on the workload, so you ship AI products faster.
To get started with these new capabilities, please contact your Google Cloud sales representative.