Companies
Google Cloud to Expand Vera Rubin Infrastructure to 960,000 GPUs
Google Cloud has announced plans to expand its NVIDIA Vera Rubin infrastructure to accommodate as many as 960,000 GPUs. This ambitious growth will be facilitated by a new family of A5X instances built on NVIDIA’s Vera Rubin NVL72 systems.
The A5X instances are bare-metal, allowing customers direct access to the physical hardware without any virtualization layer. Each instance is powered by an NVL72 rack, which integrates 72 Rubin GPUs and 36 Vera CPUs, all interconnected via NVLink 6. To scale beyond a single rack, Google is utilizing its Virgo networking fabric and NVIDIA ConnectX-9 SuperNICs to connect systems across multiple sites.
Performance claims for the A5X instances are significant, reportedly offering up to 10 times lower inference costs per token compared to previous generations, along with up to 10 times higher throughput per megawatt.
The Vera Rubin platform was launched by NVIDIA on January 5, 2026, with Google announcing the A5X instances at its Google Cloud Next event in April 2026. The first A5X instances went live on October 21, 2026, with general availability expected in November 2026. The initial customer for these instances is Ineffable Intelligence, a London-based firm utilizing A5X for complex reinforcement learning systems.
Google has also indicated plans for a European expansion of this infrastructure, although specific sites and timelines have not been disclosed. It is important to note that the 960,000 GPU figure represents a capacity ceiling across multisite deployments rather than the number of chips currently operational. Future developments to watch include the November 2026 general availability and details regarding the European rollout.
New Insights on GPU Lifespan and Demand
- Nvidia CEO Jensen Huang estimates that Blackwell and Vera Rubin GPUs could remain useful for seven to eight years, contrasting with the typical four-year lifespan used in financial models.
- Nvidia projects a combined revenue of $1 trillion from Blackwell and Vera Rubin through 2027, doubling its previous projection of $500 billion.
- Blackwell entered volume production in 2025, while Vera Rubin is expected to ramp up in the second half of 2026.
- Operators are increasingly extending GPU depreciation schedules to five to six years, as older models retain economic value despite the introduction of newer models.
- CoreWeave has extended contracts on Nvidia’s A100 GPUs, which were launched in 2020, to run until 2029, indicating strong demand for older GPUs.
FAQ
What is the purpose of Google's expansion of the Vera Rubin infrastructure?
Google Cloud is expanding its NVIDIA Vera Rubin infrastructure to accommodate up to 960,000 GPUs to enhance performance and scalability for customers utilizing advanced computing tasks.
What are A5X instances and how do they differ from previous instances?
A5X instances are bare-metal servers built on NVIDIA's Vera Rubin NVL72 systems, providing customers direct access to physical hardware without virtualization, which allows for improved performance and efficiency.
When will the A5X instances be generally available?
The A5X instances are expected to be generally available in November 2026, following their initial launch on October 21, 2026.
Who is the first customer for the A5X instances?
The first customer for the A5X instances is Ineffable Intelligence, a London-based firm that is utilizing the instances for complex reinforcement learning systems.
Will Google Cloud expand the Vera Rubin infrastructure to Europe?
Yes, Google has indicated plans for a European expansion of the Vera Rubin infrastructure, although specific sites and timelines have not yet been disclosed.