Executive Summary
- According to an NVIDIA blog, A5X instances with Vera Rubin NVL72 will be generally available starting next month, with planned expansion to Europe.
- A5X integrates Rubin GPUs, ConnectX-9 SuperNICs, the Google Virgo Network and fourth-generation liquid cooling for agent inference and reinforcement learning.
- Google Cloud plans to scale the Vera Rubin infrastructure to up to 960,000 Rubin GPUs across multiple locations.
- Confidential Computing will be offered at rack scale and on G4 VMs; early users of A5X and Confidential G4 are named.
What happened?
According to the blog posts from NVIDIA and Google Cloud, both companies announced general availability for A5X instances based on NVIDIA Vera Rubin. The announcement states availability begins next month with planned expansion to Europe. A5X instances combine NVIDIA Vera Rubin NVL72 GPUs with NVIDIA ConnectX-9 SuperNICs, the Google Cloud Virgo Network and Google’s fourth-generation liquid cooling.
The posts describe A5X as part of Google Cloud’s AI Hypercomputer—an integrated system that brings compute, networking and storage together for experience-driven learning. Google Cloud plans to scale this Vera Rubin–based infrastructure to up to 960,000 Rubin GPUs across multiple sites.
Early users include Ineffable Intelligence (the company around David Silver), which is using A5X for large-scale reinforcement learning and «superlearner» systems, and WPP, which tested creative agent architectures on A5X with NVIDIA Nemotron and OpenShell. WPP reports that in initial tests workloads that otherwise required a multi-node cluster ran on a single A5X instance.
Google Cloud and NVIDIA also announce plans to offer NVIDIA Confidential Computing at rack scale (for example on A4X Max bare-metal instances with GB300 NVL72) and to make Confidential Computing generally available on G4 VMs with NVIDIA RTX PRO 6000 Blackwell Server Edition GPUs. Users of Confidential G4 named include Vertebrae, Polaris Cloud and Confidential Core AI. Finally, Google and NVIDIA are evaluating DSX MaxLPS capabilities to optimize performance per watt.
Why does it matter?
This announcement effectively expands the availability of cloud infrastructure tailored for agent inference and large-scale reinforcement learning. The combination of specialized GPU hardware, high intra-rack network bandwidth and integrated cooling is, according to the announcement, targeted at workloads that require tight coupling between data generation, training and inference.
From an infrastructure perspective, the option to consolidate previously distributed multi-node workloads onto single A5X instances (as reported in initial WPP tests) can potentially reduce orchestration complexity and shorten iteration cycles. The planned scaling to up to 960,000 Rubin GPUs indicates the providers intend to supply dedicated capacity for very large projects.
The availability of Confidential Computing both at rack scale and on G4 VMs addresses protection of proprietary models and sensitive data when running on cloud infrastructure. That is relevant for organizations that want to secure models and data even in third-party managed environments.
What does this mean for mid-sized companies?
Factually: the announcement expands cloud infrastructure options for demanding AI projects, particularly for agent systems and reinforcement learning. It also indicates planned European availability and options for confidential processing.
Context for decision-makers in mid-sized companies:
- Strategic relevance: Companies needing agentic AI, continuous reinforcement learning or very compute-intensive models can now access specialized cloud resources; the infrastructure is designed for high-scale use.
- Data protection and compliance: The announced Confidential Computing options provide technical means to protect models and sensitive data on cloud resources; this can make using external infrastructure for confidential workloads easier without requiring large new local data center investments.
- Operational impact: The ability to consolidate previously distributed workloads onto single A5X instances may reduce complexity and latency in development and test cycles; concrete effects depend on individual workloads and integration effort.
Related: AI Agent Generator
Recommendation from Future AI Solutions
Fact-based steps, separated from assessment:
- Assess needs: Evaluate whether your AI initiatives include agent inference, large-scale reinforcement learning or data-intensive model training that could benefit from specialized Rubin hardware.
- Pilot on confidential VMs: Use available Confidential Computing options (e.g., G4 VMs) for pilot lighthouse projects with sensitive data before planning wide-scale migrations.
- Cost, compliance and integration check: Clarify cost models, data and model ownership, and integration effort (network, storage, cooling) with your cloud provider, especially regarding planned availability in Europe.
- Partner contact and roadmap monitoring: Contact your cloud partners or NVIDIA resellers to clarify timelines, availabilities and concrete SLA/support capabilities for A5X and Confidential options.
In the short term, Future AI Solutions recommends running targeted feasibility and data-protection tests; if results are positive, companies can gradually move pilot workloads to the offered instances while evaluating operational requirements.