Oracle is deepening its bet on Nvidia’s newest AI hardware while also opening its cloud to Nvidia’s rivals — a two-track move announced Tuesday at Oracle AI World in Las Vegas. Oracle Cloud Infrastructure (OCI) Superclusters built on Nvidia Vera Rubin NVL72 systems will expand to multi-gigawatt scale, the companies said, and Oracle will offer Google’s tensor processing units (TPUs) on its cloud for the first time.
The expansion covers the infrastructure behind AI training, inference, and agentic applications. The joint announcement says Vera Rubin NVL72 systems and Vera CPUs are now generally available on OCI, and that the planned multi-gigawatt buildout extends OCI’s existing Nvidia-powered systems, which already include GB200 and GB300 NVL72 hardware. Each NVL72 rack connects 72 GPUs through Nvidia’s NVLink networking, and Oracle says its Superclusters combine those racks with its own Acceleron networking and gigawatt-class power and cooling, scaling across hundreds of thousands of GPUs.
There is a second half to the announcement that matters as much as the headline scale. AI agents do more than generate text: between model calls they run tools, execute code, and retrieve data, all CPU-intensive work that can leave expensive GPUs waiting. Oracle says customers can now scale CPU and GPU resources independently on OCI, with Vera CPUs handling the in-between work and OCI Kubernetes Engine managing the split. The goal, in the companies’ words, is to keep more infrastructure productive instead of idle.
OpenAI is the named proof point. According to the release, OpenAI trained and served its GPT-6 Astra model on an OCI Supercluster with more than 100,000 Nvidia Blackwell GPUs in Abilene, Texas. Sreekanth Pothanis, an engineering leader for compute infrastructure at OpenAI, is quoted saying Vera Rubin on OCI provides “the computing performance and scale to accelerate model training and serve increasingly demanding AI workloads.” Uday Ruddarraju, OpenAI’s chief technology officer of compute, said the company is bringing Vera CPUs into production on OCI for its agentic sandbox workloads. Both quotes come from the companies’ release, not from independent interviews.
The hedge: Google, Cerebras, and SambaNova join the menu
Reuters reports that Oracle’s diversification goes further. Alongside Google TPUs, Oracle plans to offer chips from Cerebras Systems and Intel-backed SambaNova Systems, both focused on inference — the work of answering user requests, a market where Nvidia’s dominance is thinner than in training. Even the Nvidia-based systems will skip some Nvidia networking chips in favor of Oracle’s in-house Acceleron networking, Reuters reports.
Mahesh Thiagarajan, president of Oracle Cloud Infrastructure, put the logic plainly in a statement cited by Reuters: “Different AI workloads have different compute needs.” Cerebras chief executive Andrew Feldman told Reuters the Oracle deal is “a massive expansion of our customer base of tier-one customers,” and Oracle said Cerebras Inference is planned to become available on OCI during calendar year 2027.
What the announcement does not include is a price tag. The companies did not disclose the value of the expanded deal. For context, Reuters notes that Nvidia chief executive Jensen Huang has previously said Vera Rubin systems can generate up to $40 billion in revenue per gigawatt for Nvidia — a general statement about the hardware line, not a figure for this Oracle agreement, and it should not be read as one.
Oracle also plans a smaller on-ramp: the HGX Rubin NVL8, an eight-GPU configuration in an x86-based server with two 64-core Intel Xeon 6 CPUs, 4TB of DDR5 memory, and 6.4 Tbps of RDMA networking, aimed at customers that want Rubin hardware without a Supercluster-scale commitment. And the expansion itself is framed as plans, not completed construction — the release carries Oracle’s standard forward-looking disclaimer that timing and features may change.
Why now
The timing is no accident. Cloud AI capacity has become the industry’s binding constraint: AWS capacity is largely sold out through 2027, and the scramble for GPUs has pushed other players into heavy borrowing, as SpaceX’s $40 billion GPU debt shows. Oracle’s answer is to sell both scale and choice at once — more Nvidia hardware than before, plus rival chips for workloads where they fit better — while rivals fund their own buildouts, including Meta’s planned $95 billion stock offering for AI infrastructure.
That makes the Oracle announcement easy to misread. It is not Oracle walking away from Nvidia: the Vera Rubin expansion deepens the partnership to a scale neither company quantifies in gigawatts beyond “multi.” Nor is it a pure Nvidia story: the first-time Google TPU offering and the Cerebras and SambaNova plans mean OCI customers will soon choose between rival accelerators inside a single cloud. Oracle is betting that in a capacity-constrained market, the winning pitch is not loyalty to one chipmaker — it is having every credible chip available, with the networking and CPU plumbing to keep them all busy.
Sources: NVIDIA/Oracle joint release, Reuters.
