Google Cloud’s latest NVIDIA system is easy to summarise as a very large GPU announcement. The more important engineering claim sits between the chips.
A5X instances based on NVIDIA’s Vera Rubin NVL72 platform are scheduled for general availability in November 2026. Google says its Virgo Network can support clusters of up to 80,000 Rubin GPUs inside one site and 960,000 across multiple sites. Those numbers describe the scale the architecture is designed to support; they are not a count of installed, commissioned or customer-available GPUs.
A rack is only the starting unit
NVIDIA’s NVL72 design combines 72 Rubin GPUs with Vera CPUs and high-bandwidth links inside a rack-scale system. Google’s A5X wraps that hardware in its AI Hypercomputer architecture, adding networking, storage, cooling and orchestration.
The 21 October Google Cloud announcement says A5X will initially be offered as bare-metal GPU-accelerated instances, with additional European regions intended to support local data-residency and sovereignty requirements. NVIDIA’s own announcement says availability is due the following month.
Virgo is the headline behind the headline
Training and reinforcement-learning jobs divide work across many accelerators. If those chips cannot exchange data quickly and predictably, expensive hardware waits idle. Virgo uses a non-blocking mesh fabric across and between buildings, according to Google. “Non-blocking” means the network is designed so simultaneous traffic can proceed without one path becoming a structural choke point.
Scaling across sites is harder than adding racks in one hall. Distance adds latency, failures become more common, and software must coordinate work across multiple fault domains. The 960,000-GPU figure therefore says more about an intended network topology than about a single giant computer ready for use.
Storage and memory also join the system
A5X will pair Google Cloud Managed Lustre with NVIDIA GPUDirect Storage. That gives data a remote-direct-memory-access path from the parallel file system into GPU memory, bypassing the CPU and system memory. Reducing those copies can improve throughput when models repeatedly read large datasets or checkpoints.
Google is also adding multi-tier key-value caching. During inference, a model stores representations of the prompt and prior tokens in a KV cache. Reusing that state can avoid recalculating the same context, especially in long conversations and multi-step agent workloads. Moving some cache beyond scarce high-bandwidth GPU memory trades raw speed for greater effective context capacity and lower redundant computation.
Google Kubernetes Engine will handle orchestration, including topology-aware scheduling and management of RDMA network interfaces. Those software layers determine whether developers can use the hardware efficiently rather than merely reserve it.
Availability is not the same as utilisation
A general-availability date is a commercial announcement, not an independent benchmark. Google has not published customer pricing in the announcement, a full deployment count or evidence that a 960,000-GPU workload has run. The company says the first customer to access A5X is Ineffable Intelligence, founded by AlphaGo creator David Silver, but early access does not validate every scale claim.
The practical test will be delivered capacity, job efficiency, reliability, power use and cost per useful result. Large clusters also face power and cooling limits; Google says A5X uses its fourth-generation advanced liquid cooling, while Google and NVIDIA are exploring designs that optimise performance within data-centre power budgets.
A5X’s significance is therefore architectural. It treats the GPU, network, storage, cache, scheduler and cooling plant as one computer. The 960,000 number may attract attention, but the system’s value will be decided by how little of that expensive machine spends time waiting—and whether customers can obtain capacity without turning a technical ceiling into an assumption about near-term supply.
Leave a Reply