Google Cloud’s first A5X instances, built on NVIDIA’s Vera Rubin NVL72 systems, are now live with an initial customer, and general availability is planned for November 2026. Why it matters: the next generation of AI chips is moving from launch announcements into rentable cloud capacity, which is what most AI builders actually buy.
October 21, 2026
What happened
NVIDIA and Google Cloud announced the milestone at GTC Berlin on Wednesday, October 21. The first live A5X instances are running with an initial customer, general availability follows next month, and expanded availability in Europe is also planned. Google first announced A5X at Google Cloud Next in April 2026, so this is the step from unveiling to working machines.
Google Cloud said it plans to scale its Vera Rubin-based infrastructure to as many as 960,000 Rubin GPUs across multiple sites. That is a stated plan for the build-out, not a count of GPUs already installed.
What A5X is
Each A5X instance combines NVIDIA Vera Rubin NVL72 rack-scale systems and NVIDIA ConnectX-9 SuperNICs with Google Cloud’s Virgo Network and Google’s fourth-generation advanced liquid-cooling technology. Inside Google Cloud’s AI Hypercomputer, A5X integrates compute, networking and storage as a single system.
The instances are designed for large-scale agentic inference and reinforcement learning — workloads where a model repeatedly generates experience, evaluates the outcome and updates itself. That generate-evaluate-update loop needs tight, high-bandwidth connections between the Vera CPUs and Rubin GPUs and among the GPUs in the rack.
Who is using it first
The first customer is Ineffable Intelligence, the London company founded by AlphaGo creator David Silver, which is building systems that learn from experience rather than only from human data. NVIDIA quoted Silver saying that work needs a tight connection between generating experience and training on it, and crediting Vera Rubin’s compute and interconnects with letting the company push its research further.
Why it matters
Chip launches do not change what AI teams can do until the hardware shows up in a cloud they can rent by the hour. A5X reaching general availability in November puts NVIDIA’s newest rack-scale systems into that market, with Google competing on the full system — networking, cooling and storage — rather than raw GPUs alone. Read the 960,000-GPU figure carefully: it is a scaling plan across multiple sites, and actual deployments will depend on capacity, power and demand. Still, the direction is plain: the biggest clouds are racing to lock in next-generation AI capacity before their rivals do.
Frequently asked questions
When will Google Cloud A5X instances be generally available?
General availability is planned for November 2026, with expanded availability in Europe also planned. The first live instances are already running with an initial customer.
What hardware powers A5X?
A5X is built on NVIDIA Vera Rubin NVL72 rack-scale systems with ConnectX-9 SuperNICs, combined with Google Cloud’s Virgo Network and fourth-generation liquid cooling.
What is A5X designed for?
It is aimed at demanding AI workloads, especially large-scale agentic inference and reinforcement learning, inside Google Cloud’s AI Hypercomputer.
How many Rubin GPUs will Google Cloud deploy?
Google Cloud says it plans to scale its Vera Rubin-based infrastructure to as many as 960,000 Rubin GPUs across multiple sites. That is a plan, not installed capacity.
Sources: NVIDIA announcement via Unite.AI, WebProNews, Securities.io