Vera Rubin Live on Oracle Cloud With 150,000 CPUs

NVIDIA’s newest AI systems, Vera Rubin NVL72, and its Vera CPUs are now generally available on Oracle Cloud Infrastructure (OCI), the two companies said on 27 October. They also announced plans to grow Oracle’s Vera Rubin clusters to “multi-gigawatt scale”.

The joint announcement, made at Oracle AI World in Las Vegas and published on NVIDIA’s newsroom and Oracle’s, gives no dollar value, no gigawatt figure and no timeline for that expansion.

What is available now

Vera Rubin NVL72 is a rack-scale system: each rack links 72 GPUs through NVIDIA’s NVLink. The companies say the systems run inside OCI Superclusters, whose network scales “across hundreds of thousands of GPUs”. The new systems build on the GB200 and GB300 NVL72 machines Oracle already offers.

In a blog post on the launch, Oracle says Vera Rubin delivers “up to 30 times higher inference throughput per watt and a 4x leap in training performance” compared with NVIDIA’s previous Blackwell generation. Those are vendor figures, not independent tests.

The same blog says Vera CPUs are launching with “an initial deployment of 150,000 CPUs”, and that “Oracle is the first hyperscaler to bring Vera CPUs to market.”

Oracle also plans a smaller option, HGX Rubin NVL8: eight Rubin GPUs in an x86 server with two 64-core Intel Xeon 6 processors, 4TB of DDR5 memory and 6.4 Tbps of networking. No launch date was given.

Why a cloud wants CPUs for AI agents

The companies make a point about AI agents. Agents spend part of their time on GPU work, running the model, and part on CPU work between model calls: running tools, testing code and fetching data. If that CPU work is slow, expensive GPUs sit idle.

Vera CPUs are meant for that second job. The companies say customers can scale CPU and GPU capacity separately, and that Oracle’s Acceleron networking gives “hardware-based isolation for workloads running untrusted, AI-generated code”. TSN reported earlier that Modal plans to run its code sandboxes on Vera CPUs after early tests.

OpenAI is the named early user. “We are bringing Vera into production on OCI to support these workloads at scale,” said Uday Ruddarraju, OpenAI’s chief technology officer of compute.

The release also says OpenAI “has trained and served GPT-6 Astra on an OCI Supercluster with more than 100,000 NVIDIA Blackwell GPUs” at Abilene, Texas. That claim comes from NVIDIA and Oracle’s release.

What the missing numbers mean

“Multi-gigawatt” is the headline phrase, but it is a plan, not a figure. Neither company says how many gigawatts, when the capacity will arrive, where it will be built or what it will cost. A gigawatt is roughly the output of a large power station, so even one would be a major build. Without a number, readers cannot compare this with other AI data-centre deals or judge how much of it is already contracted.

What is concrete is narrower: the systems are on sale now, Oracle names an initial 150,000 Vera CPUs, and OpenAI says it is moving Vera into production. Those are real signs of demand, especially from OpenAI, which already runs very large workloads on Oracle.

The CPU angle is the more interesting shift. For the past few years, AI cloud capacity has been measured almost only in GPUs. As agents do more work between model calls, cloud providers are starting to sell CPU capacity as part of the AI story too. Ian Buck, NVIDIA’s vice president of hyperscale and high-performance computing, said: “AI factories must be productive, durable and fungible.”

The deal also fits Oracle’s wider push to sell AI to businesses, from opening its Fusion apps to ChatGPT to a $1 billion commitment to its partner network. For investors and rivals, the key detail still to come is the size of the expansion.

What happens next

Oracle says HGX Rubin NVL8 is planned, and the multi-gigawatt build-out is described as a plan. Any figures on capacity, sites or timing would come from later Oracle or NVIDIA updates.