Google Cloud will make Nvidia’s Vera Rubin infrastructure generally available next month, and the scale it describes is hard to ignore. The company plans to grow to as many as 960,000 Rubin GPUs across multiple sites. European customers must wait, since the hardware reaches them later.
Customers will reach Vera Rubin through A5X instances, which pair Nvidia Vera Rubin NVL72 racks with ConnectX-9 SuperNICs and Google’s Virgo Network. Google cools them with its fourth-generation liquid cooling, and the instances also sit inside its AI Hypercomputer. Each NVL72 rack holds 36 CPUs and 72 GPUs. Compared with earlier NVL72 systems, this generation uses full liquid cooling and cable-free modular trays, which Nvidia says cut installation from two hours to five minutes.
The platform itself spans six chips: the Vera CPU, Rubin GPU, NVLink 6 Switch, ConnectX-9 SuperNIC, BlueField-4 DPU and Spectrum-6 Ethernet Switch. Vera succeeds Grace, and Rubin succeeds Blackwell. Nvidia claims Rubin delivers 5x the inference performance and 3.5x the training performance of Blackwell. Those are vendor figures, so customers should test their own workloads before planning around them.
One early customer shows what the hardware is for. Ineffable Intelligence, a London-based company, is using Vera Rubin for reinforcement learning research. Founder and CEO David Silver says its “superlearners” need heavy compute plus a tight link between producing experience and training on it. High-bandwidth connections between the CPU and GPUs, and between GPUs inside the rack, let that loop run faster.
Security gets an upgrade too. Google Cloud says it will be among the first providers to offer Nvidia Confidential Computing at rack scale on bare metal instances such as A4X Max, powered by GB300 NVL72. Confidential Computing is also generally available now on G4 VMs with RTX PRO 6000 Blackwell Server Edition GPUs.
Competition is already close. CoreWeave claims it was first to get a Vera Rubin system running, so Google is joining a race, not leading one.
Availability will matter more than announcements. Readers should watch which regions switch on first, how quickly capacity grows toward that 960,000 figure, and when European users finally gain access.
