Google Cloud A5X reaches general availability with Rubin GPUs
Google Cloud has made its A5X instances, built on NVIDIA Vera Rubin NVL72, generally available. The change is dated November 6 in Google’s infrastructure announcement, originally published October 21. Its weekly news roundup confirms the availability update.
Google directs customers to their account team for capacity reservations and setup instructions. General availability does not guarantee a particular project’s start date.
What an A5X instance includes
Google’s machine documentation lists a5x-highgpu-4g-metal with four Rubin GPUs and 1,152 GB of HBM4 GPU memory. It also includes 176 physical Vera CPU cores, 1,440 GB of instance memory and 48,000 GiB of local SSD storage. The service is bare metal.
An NVL72 domain connects up to 18 instances, or 72 GPUs. Buyers should distinguish a single instance from the larger connected GPU domain when sizing deployments.
Where customers can access it
At this check on November 8, Google’s GPU location table lists A5X in us-west1-b, in The Dalles, Oregon. The Europe rows do not list A5X. Google’s October announcement discusses expansion into EMEA regions, but the public table does not yet establish an available European A5X location.
A listed zone identifies where the product is offered. It does not show unallocated inventory or guarantee that a requested cluster size is ready. Teams with residency constraints should confirm the actual deployment zone before arranging data transfers or committing to a launch schedule.
Supported consumption options include future reservations for capacity blocks and resource based commitments. Spot and Flex-start are not supported. Google bills Hyperdisk storage and network usage separately. The documentation also explicitly excludes A5X from the Compute Engine Service Level Agreement.
These details should be part of the procurement discussion alongside the hourly or committed price. Ask what service commitments apply, how capacity is reserved and what happens during maintenance. The GA label does not answer those operational questions.
Check integration availability separately
The October announcement describes Managed Lustre support for NVIDIA GPUDirect Storage and advanced KV caching in future tense. The short November 6 GA notice does not separately confirm that every announced integration has shipped. Teams relying on either capability should verify its current product documentation and availability.
Google also describes Virgo networking scaling to 80,000 Rubin GPUs at one site and 960,000 across sites. Those are architecture scaling claims, not a count of installed machines or a promise that a customer can reserve that many GPUs.
What the launch changes
A5X adds another commercial route to Rubin infrastructure, following Oracle’s separate Vera Rubin availability announcement. The two services should be compared using the intended workload, available locations, capacity terms and total operating cost rather than the GPU name alone.
For a useful evaluation, run a representative model with the context length, concurrency and storage traffic expected in production. Measure completion time and cost for the whole job. Hardware specifications can help narrow a shortlist, but they do not establish an application’s speed or price advantage without that workload evidence.
