Google Launches A5X Instances for Faster AI Compute
The new bare-metal instances use NVIDIA Vera Rubin hardware to increase efficiency by up to 10 times.
Updated on Oct. 21, 2026 in Data Centers
Live Poll
Do you trust that massive corporate data center expansions are ultimately good for your local community?
Google Cloud has launched A5X bare-metal instances, which provide customers with direct access to physical NVIDIA Vera Rubin NVL72 hardware. The architecture eliminates the standard virtualization layer to support high-performance computing requirements.
Why it matters
By removing the virtualization overhead, Google aims to significantly improve energy efficiency and reduce inference costs for large-scale AI workloads. This shift reflects a broader industry move toward custom infrastructure designed specifically for large-scale training and reinforcement learning.
Each A5X rack houses 72 Rubin GPUs and 36 Vera CPUs interconnected via NVLink 6. These systems achieve their efficiency gains by pairing the Virgo networking fabric with NVIDIA ConnectX-9 SuperNICs, a specialized network interface card that accelerates data transfer between compute nodes.
The players
Google Cloud
A division of Alphabet that provides global cloud infrastructure, including specialized hardware and AI services for enterprise and research.
NVIDIA
A developer of GPU architectures and networking hardware that dominates the market for high-performance AI training and inference.
Ineffable Intelligence
An early adopter of specialized high-performance computing resources focused on reinforcement learning research.
The details
The A5X instances operate as bare-metal systems, providing users with direct access to physical hardware without a virtualization layer—software that typically simulates hardware to run multiple operating systems on one machine. By bypassing this layer, the system reduces latency and resource contention. The architecture utilizes the Virgo networking fabric to bridge massive GPU clusters, enabling the scaling of infrastructure to a stated capacity ceiling of 960,000 GPUs.
Timeline
January 5, 2026: NVIDIA launches the Vera Rubin platform.
April 2026: Google announces the A5X instances.
October 21, 2026: The first A5X instances go live.
November 2026: General availability for A5X instances is planned.
The Tech Race
Google is positioning its A5X infrastructure to lead the transition to the Vera Rubin generation of hardware. This deployment represents a significant scaling milestone in the race to provide, at maximum capacity, up to 960,000 GPUs across multisite data center environments.
Engineering teams and research labs focusing on reinforcement learning will gain access to direct hardware control starting with this initial live phase. Those looking to scale across European deployments can expect wider availability as Google moves toward future planned infrastructure expansions.
The takeaway
This launch marks a significant shift toward bare-metal access for large-scale GPU clusters. Observers should track the November 2026 rollout for broader availability and potential expansion into European data centers.
What happens next
General availability for all A5X customers is currently scheduled for November 2026.
Further reading
For more on the latest trends in infrastructure, see the Data Centers section.
Live Poll
Do you trust that massive corporate data center expansions are ultimately good for your local community?
