Google Cloud's Vera Rubin A5X goes generally available next month, with a 960,000 GPU ceiling

Google Cloud and Nvidia announced on October 21, 2026 that Nvidia Vera Rubin infrastructure is generally available on Google Cloud, with the A5X instances reaching general availability next month and a planned expansion to Europe. The headline number is 960,000 Rubin GPUs. That figure is a multi-site ceiling Google plans to scale toward, not a single cluster and not capacity available today.

What is actually available

According to Nvidia's announcement, A5X instances are accelerated by Nvidia Vera Rubin NVL72 systems, with Nvidia ConnectX-9 SuperNICs, Google Cloud's Virgo Network and its fourth-generation liquid cooling. They sit inside Google Cloud's AI Hypercomputer. The announcement says the instances become generally available next month, so as of October 22 the news is a date, not yet a product you can rent.

Google Cloud also lists confidential computing options in the same release: rack-scale Nvidia Confidential Computing on bare-metal A4X Max instances powered by GB300 NVL72, and confidential computing generally available on G4 VMs with RTX PRO 6000 Blackwell Server Edition GPUs.

The 960,000 figure, read carefully

The announcement says Google Cloud plans to scale its Vera Rubin-based infrastructure to as many as 960,000 Rubin GPUs across multiple sites. Earlier coverage of the April 22, 2026 launch at Google Cloud Next, reported by Let's Data Science from Nvidia's blog, gave the breakdown: single-site clusters reach 80,000 Rubin GPUs, and federated multi-site deployments reach 960,000. The larger number combines multiple sites. That is a twelvefold difference between what one training job could use inside one site and the number in the headline.

The words matter too. "Plans to scale to as many as" is a ceiling and an intention. Neither announcement gives a delivery schedule for the ceiling, and neither says how many of those GPUs are already ordered, installed or powered. Pricing, region availability and service-level terms were not fully disclosed in the April coverage, and the October post does not add them.

The efficiency claims

The April announcement made vendor claims of up to 10 times higher token throughput per megawatt and up to 10 times lower inference cost per token compared with the previous generation. Let's Data Science presented these as vendor claims and noted it had no supporting benchmarks, and the article cited one source: Nvidia's blog. We saw no independent benchmark in the October release either. Treat "up to 10x" as a design target until a third party measures it on real workloads.

The October post adds a power angle. Google Cloud and Nvidia say they are exploring Nvidia's DSX MaxLPS for Vera Rubin infrastructure to improve performance per watt and expand capacity within existing data-center power budgets. That is a direction of work, not a result. Nvidia's DSX blog the same day puts the underlying problem plainly: "Power is now the biggest bottleneck in infrastructure."

Who is using it

Two customers are named in the release, which is a small base for a platform pitched at near-million-GPU scale.

The first named customer is Ineffable Intelligence of London, the first to access A5X, which uses it for large-scale reinforcement learning. Its founder and CEO, David Silver, said: "We're building superlearners that continually improve through their own experience". WPP, also London-based, demonstrated a creative agent reference architecture on A5X using Nvidia Nemotron models and Nvidia OpenShell. Workloads that typically needed a multi-node cluster ran on a single A5X instance in initial tests, which is a vendor-reported result from a demo, not a benchmark.

Why this matters for the race

Frontier labs do not buy compute from one place. Fox News reported on October 21 that Google Cloud now offers OpenAI models alongside Gemini and Anthropic's Claude, as a private preview, making it the only major cloud with all three. The same cloud supplying Rubin racks and hosting rival models is a sign of how interlinked the competitors' compute has become.

The physical limits remain. Nvidia said the same day that it is expanding substrate, packaging, testing and optical capacity with six suppliers, a build that runs to 2027 and 2028 in its own schedule. For a rival rack-level power claim, see how Microsoft framed its own in our report on the Maia 300 accelerator rack.

What a buyer should ask

For a team planning a large training run, the practical questions are narrow. Which region gets the first A5X capacity? What is the minimum commitment, and is reserved capacity priced differently from on-demand? Does the cluster size you need fit inside one site, given the 80,000-GPU single-site limit, or would the job have to span sites? None of these are answered in the October announcement. They are the difference between a press release and a purchasing decision.

The hype check

The claim is a near-million-GPU Google Cloud footprint on Vera Rubin. The evidence is a statement of plan from two vendors, with one confirmed early customer and one demo. Nothing independent confirms the ceiling, the efficiency multiples or the delivery date. The availability date is the only part a customer can hold Google to.

Our take

Do not read 960,000 as a number of GPUs anyone will have next year. Read it as the size of the order Google and Nvidia want customers to imagine. Three checks matter: whether A5X is actually generally available in November as stated, whether a European region is named with a date, and whether any customer publishes a measured cost per token against the prior generation. If the first two happen on time, the compute race moves on schedule. If the third never appears, the 10x stays a marketing number.