In-depth analysis · About 5 min read
Google Cloud outlines next-month A5X rollout and Vera Rubin expansion plans
NVIDIA outlines A5X infrastructure, early customer uses and confidential-computing stages. The planned GPU expansion is not installed capacity or a public price quote.
- Published here
- Source published
- Source checked
NVIDIA’s October 21 Google Cloud announcement says A5X instances based on Vera Rubin are due for general availability next month, with a planned expansion to Europe. Despite the headline’s availability wording, the body describes a future general rollout and separate customer examples; it does not establish open access in every region today.
Listen to the news
Read this article aloud or keep listening to the news.
Preparing the reader…
Device voice
Speech is handled by your browser or device. AISubHub does not send the text to an audio API. Voice availability and playback with a locked screen vary by device.
Resuming may repeat a short phrase.
At a glance
- A5X combines Vera Rubin NVL72, ConnectX-9 SuperNICs, Google Cloud Virgo Network and fourth-generation liquid cooling. Google Cloud plans up to 960,000 Rubin GPUs across multiple sites; this is a plan, not verified installed capacity.
- Ineffable Intelligence is described as the first A5X customer. WPP reports an early creative-agent demonstration with human approval, while G4 confidential VMs and planned A4X Max bare-metal protection have different availability stages.
The infrastructure package and the availability clock are separate
The announcement describes A5X as an integrated system inside Google Cloud’s AI Hypercomputer, combining compute, networking and storage for training, inference and large-scale reinforcement learning. Its named ingredients are Vera Rubin NVL72, ConnectX-9 SuperNICs, Virgo networking and Google Cloud’s fourth-generation liquid cooling. These are the platform components, not a reported application benchmark. The body says general availability is coming next month, which means November relative to its October 21 publication, without a specific day. European expansion is planned, and the article does not give a complete region-by-instance schedule. Google Cloud also plans to scale to as many as 960,000 Rubin GPUs across multiple sites. Our reading keeps a maximum planned fleet distinct from capacity installed, allocated or immediately bookable by a customer. A procurement decision still needs a confirmed configuration and access date.
The first customer illustrates an experience-driven learning loop
Ineffable Intelligence, founded by David Silver, is identified as the first customer to access A5X. The company is using the platform for large-scale reinforcement learning and describes an ambition for systems that keep learning from experience. NVIDIA explains the workload as repeated experience generation, outcome evaluation and model updates, rather than a single training job. Silver highlights close connections between generating experience and training, including bandwidth between Vera CPUs and Rubin GPUs and within the rack. This describes why the architecture may matter to that research. It does not independently demonstrate a new capability level, a finished general learner or a measured advantage over every other platform. Our analysis is to evaluate the entire loop: a faster component only helps if generation, evaluation and updates together improve the task under a controlled comparison.
WPP’s creative-agent demonstration has explicit boundaries
WPP demonstrated a reference architecture on Vera Rubin-based A5X using Nemotron models and OpenShell. Conversational requests become multistep creative workflows such as generating, upscaling and restyling video, with human approval before execution. OpenShell supplies governed sandboxes to help control access to brand assets and creative tools. NVIDIA reports that WPP’s initial tests ran workloads on a single A5X instance that typically needed a multi-node cluster. We retain that as an early partner observation, not a universal consolidation ratio or proof of lower cost. The article does not supply the full comparison hardware, workload settings or numerical quality criteria. Our interpretation is that execution approval and asset access are part of the system, not optional details erased by faster hardware. Readers should preserve those controls when evaluating whether a similar workflow is suitable.
Confidential computing and energy optimization have their own stages
The same article separates several security offerings. Google Cloud is making Confidential Computing generally available on G4 VMs with RTX PRO 6000 Blackwell Server Edition GPUs. By contrast, it will be among the first providers offering rack-scale confidential computing on bare-metal instances such as GB300 NVL72-based A4X Max. Neither statement should be silently substituted for A5X availability. Named customer uses of Confidential G4 concern privacy, verifiable execution and policy-governed intelligence; they do not prove every workload or region has identical protection and access. Google Cloud and NVIDIA are also exploring DSX MaxLPS for performance per watt and capacity within power budgets, not reporting a universal measured energy saving. Our suggested deployment review separates product, region, access stage, approvals, sensitive-data controls and total cost. Those are checks we recommend, not tests performed here. The source provides no subscription quote or complete allocation guarantee.
Why it matters
Our analysis: tightly connected compute and networking can support a complete reinforcement-learning loop, but adoption depends on actual instance access, workload fit, operating controls and a full cost comparison. Early access and a roadmap should not be read as a universal delivery commitment.
Limits and availability
This is NVIDIA’s account of Google Cloud and partner work. AISubHub has not accessed the instances or reproduced the demonstrations. Exact regional rollout, capacity allocation, customer prices and comparable end-to-end performance are not established by this article.
Sources and evidence
Source published: · Source checked:
AI-assisted editorial articles checked against linked primary sources. Vendor claims are attributed; inclusion is not endorsement.