At NVIDIA GTC Berlin 2026, Google Cloud announced a set of infrastructure and security updates built for the shift from single inference calls to continuous, multi-step agentic workflows. The company argues that scaling such systems requires not only raw performance but also model choice and verifiable control over data and infrastructure. The announcements span AI Hypercomputer compute and networking, expanded open-model and partner support, and new sovereign AI and Confidential Computing capabilities.

For the compute layer, Google Cloud introduced A5X instances based on the NVIDIA Vera Rubin NVL72 platform, with general availability coming next month and new EMEA regions for data residency and sovereignty requirements. These bare-metal GPU instances are paired with the new Virgo Network, a non-blocking intra- and inter-building fabric using mesh topology that supports clusters up to 80,000 Rubin GPUs within a single site and up to 960,000 across multi-site deployments. Storage is addressed through Google Cloud Managed Lustre with NVIDIA GPUDirect Storage, giving training and inference data a direct RDMA path from parallel file system to GPU memory without passing through the CPU. A5X also adds advanced KV-caching to scale context memory beyond GPU HBM, offloading and reusing context across storage tiers to reduce redundant computation, lower time-to-first-token, and improve inference goodput per dollar. On the orchestration side, A5X instances work with Google Kubernetes Engine, including GKE Managed DRANET for treating network interfaces as first-class Kubernetes resources and Kueue for topology-aware scheduling. Stephan Pretorius, CTO of WPP, said the combination lets WPP’s creative and production teams direct agents that perform frame-accurate quality checks and adapt brand assets, while relying on Google Cloud for reliability, scale, and sovereign security.

Google Cloud also expanded model and ecosystem options, noting that agentic workflows typically route tasks to the most efficient model rather than relying on one model. The Sovereign Cloud portfolio now supports open-weight models such as Gemma, Mistral, and NVIDIA Nemotron, allowing customers to run and govern them while retaining full custody of model weights and runtimes. NVIDIA Cosmos 3 and NVIDIA Nemotron 3.5 SuperVL with NVIDIA NeMo are now available on Gemini; Nemotron 3.5 SuperVL can be customized with Managed Training Clusters using an MTC-optimized recipe, and Cosmos 3 supports simulation-to-reality workflows for robots and autonomous vehicles to practice in virtual worlds before real-world deployment.

Sovereign AI for on-premises environments was expanded through Google Distributed Cloud (GDC). NVIDIA HGX B300 systems, announced earlier this year in preview, are now generally available on GDC for air-gapped deployments. The sovereign model catalog adds Gemini 3.5 Flash on GDC, accelerated by HGX B300, for multimodal reasoning across text, audio, video, and imagery inside an air-gapped perimeter. A preview of enterprise search with Gemini on GDC brings an out-of-the-box conversational interface powered by multimodal retrieval-augmented generation, letting knowledge workers query proprietary documents without custom engineering while keeping indexing and interactions within the sovereign boundary. For non-air-gapped needs, sovereign workloads can also run on NVIDIA AI infrastructure using Google Cloud Data Boundary or Google Cloud Dedicated, an isolated cloud run by partners such as S3NS in France and Thales in Germany.

Confidential Computing announcements target protection of IP during training and of user context and model IP during inference in the public cloud. Confidential G4 instances, enabled by NVIDIA RTX PRO 6000 Blackwell Server Edition GPUs, are now generally available. Confidential Computing is also available on A4 instances with NVIDIA HGX B200 and Intel TDX. Google Cloud further introduced Confidential Bare Metal, with A4X Max instances enabled by NVIDIA GB300 NVL72 using NVIDIA Confidential Bare Metal, extending hardware-rooted secure boot and attestation with encrypted DDR and NVLink interconnects to protect prompts and models during inference and training.

The broader positioning is that Google Cloud and NVIDIA co-engineer full-stack accelerated platforms into AI Hypercomputer, giving customers agent-native infrastructure with verifiable control over data. The company says that because hardware and software are designed together, models can run across architectures without rewriting code, keeping compute cycles focused on workloads and helping organizations ship AI products faster.