Google Cloud Adds Blackwell

Google Cloud is widening access to Nvidia’s Blackwell chips, adding newer Blackwell systems now and lining up Vera Rubin machines next. (blogs.nvidia.com) At Google Cloud Next in Las Vegas this week, Nvidia said Google will offer A5X bare-metal instances built on Vera Rubin NVL72 rack systems. Nvidia said those clusters can scale to 80,000 Rubin graphics processing units in one site and 960,000 across multiple sites. (blogs.nvidia.com) Google and Nvidia were already selling Blackwell-based cloud systems before this week. Google’s A4X virtual machines, announced in February 2025 and made generally available on May 29, 2025, package 72 Blackwell graphics processors and 36 Grace central processors into one GB200 NVL72 system. (cloud.google.com) A cloud “instance” is rented computing capacity, and a bare-metal instance gives customers direct access to the whole machine instead of a slice managed by a hypervisor. Google said its AI Hypercomputer bundles chips, networking and software into one service aimed at low-latency, high-throughput AI work. (cloud.google.com) That matters for companies building larger models and “agentic” systems, software that can take multi-step actions instead of just answering prompts. Google said the infrastructure is designed for training, fine-tuning and inference, the stage where a model generates answers for users. (cloud.google.com) Nvidia is also pushing Blackwell deeper into ordinary enterprise data centers, not just giant cloud campuses. On April 22, Nvidia introduced the RTX PRO 4500 Blackwell Server Edition and vGPU 20 software for virtualized servers. (developer.nvidia.com) The RTX PRO 4500 is a smaller 165-watt, single-slot server card with 32 gigabytes of GDDR7 memory. Nvidia said it supports up to two Multi-Instance GPU partitions, which split one physical chip into separate slices so multiple virtual machines can run at once. (nvidia.com, developer.nvidia.com) Nvidia said vGPU 20 adds fixed-share scheduling, an AI Virtual Workstation Toolkit and broader cloud compatibility, including Google Cloud. In Nvidia’s tests, the new software and card delivered nearly 1.9 times the graphics acceleration of earlier Nvidia architectures in virtualized environments. (developer.nvidia.com) Google is still backing its own Tensor Processing Units alongside Nvidia hardware. But this week’s announcements show the company is expanding its AI cloud with both homegrown chips and Nvidia’s newest systems, from fractional virtual machines to rack-scale Rubin clusters. (cloud.google.com, blogs.nvidia.com)