Running giant training and reinforcement-learning workloads in the cloud is about to feel more like ordering a single, prebuilt machine.
Google Cloud is launching A5X instances built on NVIDIA's Vera Rubin NVL72 systems. The A5X package combines compute, networking, storage and software inside Google Cloud's AI Hypercomputer, and it's pitched for model training, inference and large-scale reinforcement learning. General availability is planned for November 2026.
Google also described platform pieces that matter in practice: its Virgo Network can link multiple sites into very large Rubin GPU clusters, and Managed Lustre will support GPUDirect Storage so GPUs can pull data straight from parallel storage into GPU memory without extra copying.
Why this matters: until now teams had to assemble GPUs, high-speed networking and tuned storage themselves. That coordination is time consuming and error prone. A pre-integrated instance removes much of that friction and shortens the path from experiment to full-scale training runs.
How it works, at a glance: think of A5X as a prewired supercomputer. The parts are chosen and connected to move data directly to the GPUs, so a single job can spread across many machines without getting choked by slow storage or messy networking. Google says Virgo can stitch together nearly one million Rubin GPUs across sites, which is the scale signal here.
What changes now: teams that already spend on cloud compute will find it easier to run very large experiments and agentic workloads. Startups and research groups may be able to prototype at higher scale sooner. Practical reality check: this does not make huge clusters cheap or simple overnight. You still need serious budget and engineering to use them.
What's next: the real question is whether prebuilt, hyper-scale instances like A5X will decentralize frontier AI by letting more organizations run their own giant jobs, or whether they will instead deepen the advantage of whoever already controls the biggest cloud contracts. We'll get clearer answers as GA rolls out.
