Carl Lender from Sunrise, USA · CC BY 2.0 · Wikimedia Commons · Source · License
Google Cloud and NVIDIA officially announced on October 21, 2026, that the Google Cloud A5X instances, powered by the next-generation exascale AI infrastructure platform ‘NVIDIA Vera Rubin,’ will transition to general availability (GA) next month. This collaboration supports high-performance workloads capable of handling sophisticated autonomous agent operations and large-scale reinforcement learning, with plans to expand into global public cloud data center infrastructure, including regions in Europe.
The two companies explained that by combining NVIDIA’s next-generation Vera Rubin architecture with Google Cloud’s custom hardware liquid cooling solutions, they have dramatically increased efficiency for the training and inference of large-scale AI models. This commercialization announcement marks a milestone where next-generation accelerator hardware and specialized computing infrastructure are now directly available to global frontier AI research labs and enterprises.
Architectural Synergy: Vera Rubin NVL72 and 4th Gen Liquid Cooling
Google Cloud A5X instances are designed around the NVIDIA Vera Rubin NVL72 system. For ultra-high-speed data transfer, the NVIDIA ConnectX-9 SuperNIC and Google Cloud’s Virgo petabit-class network have been integrated. To manage the heat generated during sustained high-load computations, Google’s 4th Generation Advanced Liquid-Cooling technology has been incorporated.
Google Cloud plans to expand its Vera Rubin-based infrastructure across multiple data centers worldwide. According to the announcement, Google Cloud intends to scale the infrastructure in phases to accommodate up to 960,000 Rubin GPUs, providing an unprecedented large-scale reinforcement learning cluster within a single public cloud environment.
‘Superlearner’ Development and Reinforcement Learning at Ineffable Intelligence
The first early-access customer for the A5X instances is ‘Ineffable Intelligence,’ an AI technology firm founded in London by AlphaGo developer David Silver. Ineffable Intelligence is actively utilizing the A5X platform to build a ‘Superlearner’ system that improves its own performance through interaction and continuous collection of experiential data.
In the announcement, CEO David Silver emphasized that for autonomous agents to continuously enhance performance, there must be a tightly coupled feedback loop that generates data through environmental interaction and reflects it back into model training. He explained that the ultra-high-bandwidth connection between Vera CPUs and Rubin GPUs, as well as the organic communication structure between GPUs within the rack, plays a core role in expanding the scope of large-scale reinforcement learning research.
Unified Creative Agent Workflow: The WPP Case Study
Global marketing and communications firm WPP also demonstrated a creative agent reference architecture on the Vera Rubin-based A5X instances. WPP combined the NVIDIA Nemotron model with ‘NVIDIA OpenShell,’ a sandbox framework that supports secure agent execution, to automate complex creative tasks.
This agent system establishes multi-step creative workflows—such as video generation, resolution upscaling, and style transfer—based on natural language requests and executes them following an approval process. Initial deployment results confirmed that tasks previously requiring large multi-node clusters were performed seamlessly within a single A5X instance.
Rack-Scale Confidential Computing and Sensitive Data Protection
As enterprise AI agents become more sophisticated, security techniques protecting core confidential models and corporate data within cloud infrastructure are also being strengthened. Google Cloud announced it will be one of the first cloud providers to commercially offer ‘NVIDIA Confidential Computing’ at the rack scale via bare-metal-based A4X Max instances (equipped with NVIDIA GB300 NVL72).
Furthermore, confidential computing will also be officially available on G4 virtual machines (VMs) equipped with NVIDIA RTX PRO 6000 Blackwell server edition GPUs. Global tech companies such as Vertebrae, Polaris Cloud, and Confidential Core AI are already adopting this technology for secure, privacy-preserving inference and model fine-tuning.
Power Efficiency Optimization and Future Roadmap
To address the power consumption issues arising from large-scale infrastructure construction, both companies are continuing to collaborate on energy efficiency management technologies. They are evaluating the integration of ‘NVIDIA DSX MaxLPS’ into the Vera Rubin-based infrastructure to maximize performance per watt and extend processing capacity within the limited power budgets of data centers.
- Core Infrastructure Components: Vera Rubin NVL72, ConnectX-9 SuperNIC, Virgo network, 4th Gen liquid cooling system
- Initial Application Areas: Ineffable Intelligence (Reinforcement Learning), WPP (Creative AI Agents)
- Enhanced Security: Rack-scale confidential computing provided via A4X Max and G4 VMs
- Commercial Launch Timeline: General availability on public cloud in November 2026 with expansion into European regions
According to data released by the two companies, the A5X instances will begin full-scale general availability for public cloud customers in November 2026, with plans for continuous supply expansion to European data center hubs.
Official Source:
– NVIDIA Official Blog Announcement (October 21, 2026): https://blogs.nvidia.com/blog/google-cloud-vera-rubin-infrastructure/