Google Cloud has made its first NVIDIA Vera Rubin-based Compute Engine machine generally available, turning the newest rack-scale accelerator platform into reservable cloud capacity rather than a future roadmap item.
What changed
The new A5X machine series is exposed as the bare-metal a5x-highgpu-4g-metal shape. Google documents two 88-core NVIDIA Vera Arm CPUs, 1,440 GB of system memory, and four Rubin GPUs with 288 GB of HBM4 each, for 1,152 GB of GPU memory in one instance. The platform also bundles 48,000 GiB of local SSD capacity, supports up to 7,200 Gbps of network bandwidth, and connects the GPUs through an NVLink full mesh.
Google positions A5X for frontier-model training, mixture-of-experts inference, and high-performance computing. It can also be orchestrated with GKE, including topology-aware scheduling and RDMA networking features.
Why it matters
Teams that cannot procure and operate an NVL72 rack can now target Rubin through Google Cloud’s reservation system. The large coherent CPU-GPU memory footprint and faster HBM and interconnect are most relevant to models that otherwise spill across nodes or spend substantial time moving parameters and key-value cache data.
Important limits
This is not ordinary on-demand capacity. Google lists A5X only for future reservations through capacity blocks; on-demand, Spot, flex-start, and conventional reservations are unavailable. Customers must discuss requests with an account or sales team, and Google says the Compute Engine service-level agreement does not apply to A5X. The Arm CPU and bare-metal environment can also require software and deployment changes.
The throughput, memory-bandwidth, and scaling figures are Google and NVIDIA specifications or vendor claims. Google has not published an independently reproduced application benchmark for this launch, and DrComps has not tested A5X hardware. Pricing is not presented as a simple public on-demand rate, so practical cost-per-training-run remains unverified.