Google Cloud GTC 2026: Vera Rubin NVL72 Systems Guide

Google Cloud GTC 2026 Announcement Overview

Google Cloud's GTC 2026 announcement confirmed a deepened infrastructure partnership with NVIDIA that brings Vera Rubin NVL72 rack-scale systems into Google Cloud's AI Hypercomputer offering alongside existing TPU v5e and v5p capacity. The announcement was made during Jensen Huang's keynote, where Google Cloud CEO Thomas Kurian appeared on stage to outline the availability timeline and positioning.

The core commitment is to provide NVL72 configurations as a first-class offering on Google Cloud, not as a limited research program. Google Cloud is launching the A4 Ultra instance family specifically to expose NVL72 capacity to enterprise customers, with the same SLA, networking, and operational tooling available for existing A3 instances built on H100 hardware.

The announcement positions Google Cloud's AI Hypercomputer as a heterogeneous platform rather than a TPU-exclusive offering. This is a meaningful strategic shift. For several years, Google Cloud differentiated on proprietary TPU architecture while ceding large- scale NVIDIA-native training workloads to competitors. The NVL72 partnership signals that Google Cloud is prioritizing breadth of workload support over proprietary differentiation. For context on the financial scale behind this hardware cycle, see the analysis of NVIDIA's $1 trillion order pipeline revealed at the same GTC keynote.

Vera Rubin NVL72 Architecture and Specifications

Vera Rubin is NVIDIA's successor to the Blackwell architecture, named after the astronomer who provided foundational evidence for dark matter. The NVL72 configuration places 72 Vera Rubin GPUs in a single rack, connected by NVLink 6 rather than PCIe, creating a unified memory fabric that spans the entire rack. From the perspective of a training job, an NVL72 rack behaves more like a single very large accelerator than a cluster of individual GPUs.

The key architectural advances over Blackwell NVL72 configurations include higher per-GPU memory bandwidth, increased HBM4 memory capacity per GPU, and the NVLink 6 interconnect that doubles bandwidth per link over the NVLink 4 found in H100 systems. The FP8 training performance of 1.4 exaFLOPs per rack represents a significant step up from the roughly 720 petaFLOPs per rack achievable with H100 NVL8 configurations in dense cluster arrangements.

The NVL72 form factor matters because it changes the programming model for distributed training. In H100 cluster configurations, training jobs typically use tensor parallelism within a node of 8 GPUs connected via NVLink, then pipeline or data parallelism across nodes using InfiniBand. With NVL72, the NVLink domain extends to 72 GPUs, allowing more of the model to reside in the high-bandwidth interconnect fabric and reducing reliance on slower inter-rack communication for common training architectures.

A4 Ultra Instance Family on Google Cloud

Google Cloud is exposing NVL72 capacity through a new A4 Ultra instance family that fits into the existing accelerated compute naming scheme alongside A3 Mega instances (H100) and A2 Ultra instances (A100). The A4 Ultra family is designed for workloads that need the full NVL72 rack as a single compute unit, rather than slicing individual GPUs into smaller instances.

Instance configurations will be offered at the rack level for large training jobs and at sub-rack configurations for teams that need NVL72 performance characteristics without committing a full rack. The sub-rack options allow research teams and smaller AI companies to evaluate the architecture without the cost of reserving an entire NVL72 system.

NVLink 6 Interconnect Advantages

NVLink 6 is the interconnect technology that makes the NVL72 configuration coherent rather than simply a cluster of GPUs. The doubled per-link bandwidth compared to NVLink 4 (found in H100 systems) reduces the communication bottleneck that limits scaling efficiency in large distributed training jobs. When gradient synchronization can complete faster, training throughput approaches linear scaling with GPU count rather than the sub-linear scaling typical of PCIe-connected clusters.

For transformer-based models with billions of parameters, the backward pass involves extensive all-reduce operations across all participating GPUs. In H100 NVL8 configurations, eight GPUs share NVLink bandwidth before traffic crosses to InfiniBand for inter-node communication. With NVL72, 72 GPUs share NVLink 6 bandwidth, meaning most of the gradient synchronization for common model sizes can complete within the NVLink fabric without touching inter-rack networking.

Model FLOP utilization (MFU) is the practical metric that determines how efficiently a training cluster converts its theoretical peak performance into useful computation. NVLink 6's bandwidth improvement directly raises achievable MFU for large models by keeping more computation in the fast path. For organizations evaluating cloud training costs, higher MFU means fewer GPU-hours and lower cost to reach a given loss threshold.

Availability Timeline and Regional Rollout

Google Cloud has committed to preview availability for A4 Ultra instances in Q2 2026, with initial capacity in us-central1 (Iowa) and europe-west4 (Netherlands). These regions were chosen based on existing data center infrastructure, power availability, and the geographic distribution of expected early adopters in North America and Europe.

Regional expansion beyond the two preview regions will follow Google Cloud's standard capacity rollout process. Based on historical patterns for A3 and previous accelerated instance families, additional US regions such as us-east4 and asia- southeast1 can be expected within the first year following general availability. Google Cloud has not specified a GA date, but preview programs for previous instance families typically ran for three to six months.

Workload Guidance and Target Use Cases

Not every AI workload benefits equally from NVL72 infrastructure. Google Cloud's own guidance distinguishes between workloads that are compute-bound and bandwidth-bound, directing each to the appropriate infrastructure. Understanding this distinction is essential for making cost-effective infrastructure decisions.

Google Cloud's AI Hypercomputer orchestration layer manages job placement across TPU and GPU infrastructure, theoretically allowing training pipelines to use NVL72 for pretraining passes and then shift to TPU v5e for distillation and inference serving without separate infrastructure management. This unified control plane is a competitive advantage Google Cloud holds over pure-play NVIDIA cloud providers.

How This Positions Google Cloud vs AWS and Azure

Prior to the NVL72 announcement, AWS and Azure held an advantage in high-density NVIDIA compute for enterprises that needed NVIDIA-native infrastructure. AWS P5 instances (H100 NVL8) and Azure ND H100 v5 instances were the primary options for large NVIDIA-based training, while Google Cloud offered A3 Mega (H100) as a more recent entrant. The Vera Rubin NVL72 on A4 Ultra leapfrogs both competitors' current NVIDIA offerings in raw compute density.

AWS and Azure are expected to announce Vera Rubin availability on their own timelines, and the competitive advantage Google Cloud gains from NVL72 preview access is likely to be temporary. However, for enterprises evaluating multi-year cloud commitments for AI infrastructure, being among the first to offer NVL72 strengthens Google Cloud's credibility as a serious AI infrastructure provider alongside its proprietary TPU heritage.

TPU v5 and NVL72 Coexistence Strategy

Google Cloud's strategy is emphatically not to replace TPU infrastructure with NVIDIA hardware. TPU v5e and v5p remain the recommended path for inference serving, fine-tuning of JAX- based models, and workloads that benefit from Google's custom optimizations. The NVL72 partnership adds CUDA-ecosystem support for workloads that need it, without requiring customers to abandon existing TPU investments.

In practice, large AI organizations often run mixed infrastructure. Pretraining happens on GPU clusters with PyTorch and CUDA, while inference serving uses TPUs or purpose-built inference accelerators. Google Cloud's ability to serve both phases within a single cloud reduces the friction of multi-cloud management that has led some enterprises to split training and inference workloads across providers.

Business Impact and AI Strategy Implications

For enterprise technology buyers, the Google Cloud NVL72 announcement has immediate and medium-term implications. In the immediate term, organizations with upcoming large model training requirements should include Google Cloud A4 Ultra in their infrastructure evaluation alongside AWS and Azure NVIDIA offerings. The preview access program is worth entering even if GA timelines are uncertain, as preview benchmarks will inform procurement decisions.

In the medium term, the availability of Vera Rubin-class compute through Google Cloud accelerates the development timeline for foundation models and downstream AI products. Faster training runs at lower cost per token trained means the gap between frontier models and enterprise-accessible models continues to shrink. For businesses building on top of foundation models via API, this translates to more capable base models and more competitive pricing from providers who can train and serve more efficiently. For deeper context on how this hardware cycle fits into broader digital transformation trends, see our AI and digital transformation services.

The strategic takeaway for businesses that are not training their own foundation models is that cloud infrastructure competition benefits end users through better models and lower inference costs. Google Cloud's NVL72 adoption, combined with its TPU v5 inference infrastructure, creates a vertically integrated AI platform that can compete for enterprise AI workloads that were previously distributed across specialized providers.