Google Cloud Launches A4X VMs with NVIDIA GB200 NVL72 for Hyperscale AI

Google Cloud's October 2025 launch of A4X virtual machines marks a significant leap in hyperscale AI infrastructure, built on NVIDIA's GB200 NVL72 rack-scale systems. Each A4X node ships with 72 Blackwell B200 GPUs tightly coupled via NVLink, delivering approximately 13.5TB of HBM3e memory per node and a peak FP8 throughput exceeding 1.4 exaFLOPS at the rack level. On-demand pricing is set at approximately $340 per hour for a full NVL72 node, with 1-year Committed Use Discounts (CUDs) bringing that figure down to roughly $221 per hour — a 35% reduction that meaningfully improves economics for sustained training runs. Google is initially rolling out A4X in us-central1 and europe-west4, with asia-northeast1 availability targeted for Q1 2026. Alongside A4X, Google simultaneously announced general availability of TPU v7, giving customers a first-party alternative at comparable price points for transformer-heavy workloads.

Placing A4X in competitive context is critical for procurement decisions. AWS's equivalent P6e instances, powered by NVIDIA GB200 and available in clusters of up to 64 GPUs, are priced at approximately $310–$330 per hour per node on-demand, with 1-year Reserved Instance pricing near $210 per hour — making AWS marginally cheaper on list price but with fewer NVLink-connected GPUs per node. Azure's NDv5 H200 instances (ND H200 v5 series, 8 H200 GPUs per VM) run at roughly $98 per hour on-demand per 8-GPU VM, meaning a comparable 72-GPU cluster on Azure costs approximately $1,200 per hour — substantially more expensive at scale. Oracle Cloud Infrastructure's BM.GPU.B4.8 (H100-based) remains a generation behind at around $32 per hour per 8-GPU bare-metal node. On a price-per-petaflop basis for FP8 inference and training workloads, GCP A4X delivers a compelling ratio, particularly when CUDs are applied, and the NVL72 unified memory fabric reduces inter-GPU communication overhead that fragmented multi-VM setups on competing clouds cannot match.

The launch matters most for three customer segments: large language model pre-training teams running multi-trillion-parameter models, enterprise AI labs scaling reinforcement learning from human feedback (RLHF) pipelines, and hyperscale inference providers serving sub-50ms latency SLAs at massive concurrency. The 13.5TB unified HBM3e pool per node is a game-changer for model parallelism, allowing organizations to fit 405B+ parameter models entirely within a single NVL72 node without tensor sharding across hosts. Competitive pressure is now acute on AWS and Azure to accelerate their own GB200 NVL72 deployments and pricing; expect both providers to respond with aggressive CUD/RI discounts within 60–90 days. Key caveats include limited regional availability at launch, potential for GPU allocation queuing given early supply constraints, and meaningful Terraform and Kubernetes reconfiguration costs when migrating existing A100 or H100 workloads to the new NVLink fabric topology. Vendor lock-in risk is also elevated, as TPU v7 co-availability creates an incentive for Google to bundle pricing in ways that make cross-cloud portability harder to justify financially.

For customers with active large-scale AI training or inference workloads, the immediate action is to model your current per-GPU-hour spend against A4X CUD pricing. Organizations currently running more than 256 A100 or H100 GPUs continuously on any cloud should request A4X quota in us-central1 now, as early allocation windows tend to fill within 30–60 days of a major launch. For teams spending more than $500,000 per month on GPU compute, the 1-year CUD at $221 per hour versus on-demand $340 per hour represents a potential saving of over $1 million annually per NVL72 node at full utilization. Before committing, benchmark your specific model architecture on A4X — particularly attention to FP8 vs BF16 throughput trade-offs — since not all workloads will achieve the theoretical peak FLOP counts in practice. Teams using Kubernetes-based ML platforms should also validate their NCCL and NVLink topology-aware scheduling configurations before signing a CUD, as misconfigured collective communication can negate the hardware advantages entirely.

At TCOIQ, we view the A4X launch as exactly the kind of inflection point where a structured TCO analysis prevents costly over-commitment or missed savings. The TCOIQ TCO Calculator at tcoiq.com/tco.html can model A4X on-demand versus 1-year CUD versus AWS P6e Reserved versus Azure NDv5 side-by-side, incorporating your actual GPU utilization rates, storage egress, and networking costs to produce a true total cost of ownership rather than a headline hourly rate comparison. The Inventory Builder at tcoiq.com/inventory.html allows you to tag and categorize your existing GPU fleet so you can identify which workloads are genuine migration candidates for A4X and which are better served by staying on H100 or moving to TPU v7. Our AI Migration Assessment surfaces hidden costs like data transfer, model re-optimization time, and cluster reconfiguration that vendors never include in their pitch decks. Start by running your current GPU inventory through the TCOIQ Inventory Builder to produce a migration-readiness score before making any CUD commitment.