Google Cloud A5X Puts 960,000 Rubin GPUs on the Roadmap

Google Cloud A5X instances powered by NVIDIA Vera Rubin NVL72 are scheduled to become generally available in November 2026, combining rack-scale NVIDIA systems with Google's networking and liquid-cooling infrastructure. The headline number is bigger than the launch: Google Cloud says it could deploy as many as 960,000 Rubin GPUs across multiple sites, making A5X a serious bid to own the infrastructure layer for reinforcement learning and agentic AI.

The timing requires precision. NVIDIA's October 21, 2026 announcement describes general availability as coming the following month, not as something already delivered. European expansion is also planned, but NVIDIA and Google Cloud did not provide specific regions, deployment dates, pricing, quotas or customer capacity.

Google Cloud A5X combines Rubin with Google's infrastructure

A5X is Google Cloud's instance family built around NVIDIA's Vera Rubin platform. Its core building block is Vera Rubin NVL72, a rack-scale system that connects 72 Rubin GPUs with NVIDIA Vera CPUs using high-bandwidth interconnects, allowing the rack to operate more like one large accelerator than a collection of isolated servers.

The surrounding infrastructure matters almost as much as the silicon. A5X combines Rubin GPUs and Vera CPUs with NVIDIA ConnectX-9 SuperNICs, the Google Cloud Virgo Network and Google's fourth-generation advanced liquid-cooling technology. Together, those components address three recurring bottlenecks in frontier AI systems: moving data between accelerators, coordinating work across clusters and removing the heat produced by dense racks.

That integrated design supports training, inference and reinforcement learning on the same platform. Training builds or updates a model; inference runs that model for users or software agents; reinforcement learning improves behavior through feedback generated from environments, evaluators or task outcomes. At large scale, all three increasingly blur into a continuous loop rather than separate stages.

The 960,000 Rubin GPU plan is the real signal

Google Cloud says its Vera Rubin infrastructure could scale to as many as 960,000 Rubin GPUs across multiple sites. That figure is a deployment ceiling or plan, not evidence that nearly one million GPUs have been installed, sold or reserved. NVIDIA's announcement does not specify a completion date, capital budget or how much capacity will be available to ordinary cloud customers.

Even with those caveats, the number is unusually large. At 72 GPUs per NVL72 system, 960,000 GPUs would be equivalent to more than 13,000 fully populated 72-GPU racks before accounting for networking, storage, power distribution and cooling. The operational challenge is therefore not merely buying accelerators; it is turning multiple data centers into a reliable, schedulable AI computer.

Multi-site deployment also raises a crucial distinction. A tightly connected cluster inside one facility can support synchronized training, while capacity distributed across regions may be better suited to inference, data generation or independent reinforcement-learning experiments. Google Cloud has not explained how much of the proposed fleet will operate as unified clusters versus geographically distributed pools.

Reinforcement learning becomes a cloud workload

Google Cloud and NVIDIA are positioning A5X for demanding agentic inference and large-scale reinforcement learning. Agentic inference refers to models performing multi-step tasks by repeatedly reasoning, calling tools, checking results and trying again. It consumes far more compute than a single chatbot response because one user request can trigger dozens or thousands of model operations.

Reinforcement learning compounds that demand. Advanced systems can generate candidate actions, execute them in simulated or real environments, score the outcomes and train on the resulting experience. The process creates a flywheel in which inference produces training data and training produces a stronger inference system. That is why networking, cluster scheduling and sustained utilization now matter as much as peak benchmark performance.

Ineffable Intelligence, a London technology company founded by AlphaGo creator David Silver, was the first named customer to access A5X. According to NVIDIA, the company is using the platform to develop continuously learning systems based on large-scale reinforcement learning. Its early access demonstrates that the hardware is operating for at least one customer, but it should not be confused with broad commercial availability.

Why Google Cloud wants both NVIDIA GPUs and TPUs

Google has spent years developing its own Tensor Processing Units, or TPUs, yet it continues to invest heavily in NVIDIA infrastructure. That is not a contradiction. TPUs give Google vertical control and potential cost advantages, while NVIDIA compatibility attracts customers whose models, libraries and engineering workflows already depend on the CUDA software ecosystem.

A5X therefore serves two strategic goals. It lets Google Cloud compete for portable NVIDIA workloads against Amazon Web Services, Microsoft Azure and Oracle Cloud Infrastructure, while Virgo networking and proprietary cooling give Google a way to differentiate what might otherwise look like the same NVIDIA rack sold by every major cloud provider.

The competitive unit is shifting from the individual accelerator to the complete AI factory: chips, CPUs, memory, network fabrics, cooling, orchestration and available power. Cloud providers can all announce access to the same GPU architecture. Far fewer can deploy it across multiple sites, keep it fed with data and offer customers enough reliable capacity to run months-long learning loops.

Availability will matter more than the announcement

For AI builders, the immediate questions are practical: what A5X costs, how much capacity customers can reserve, which software stacks are supported and when European regions come online. None of those details is settled by a maximum GPU target. A giant roadmap number is useful, but rentable clusters are the product.

Still, Google Cloud's plan identifies where the market is heading. Frontier AI infrastructure is being designed not just for periodic model training, but for systems that learn and act continuously. The Rubin race will be won by the cloud that turns ambitious GPU totals into dependable compute—not the one with the largest press-release number.