Oracle goes multi-gigawatt on Nvidia Vera Rubin, then adds Google TPUs, Cerebras and SambaNova

Oracle used its AI World event on October 27, 2026 to make two promises that pull in different directions. It will scale Nvidia's newest Vera Rubin systems to multi-gigawatt size, and it will also sell customers Google's TPUs and chips from Cerebras and SambaNova. The honest reading is that Oracle is buying itself options. The announcements describe a roadmap, and the gigawatt figure everyone wants is missing.

What Nvidia says is live today

Nvidia's release is the firmest document. It says Vera Rubin NVL72 systems and Vera CPUs are now generally available on Oracle Cloud Infrastructure, and that Oracle plans to expand OCI Superclusters with NVL72 systems to multi-gigawatt scale. Each NVL72 rack connects 72 GPUs over NVLink, and OCI's network scales across hundreds of thousands of GPUs, according to the release. Oracle also plans to add HGX Rubin NVL8 systems, each with eight Rubin GPUs, two 64-core Intel Xeon 6 processors, 4TB of DDR5 memory and 6.4 Tbps of RDMA networking.

The release leans on one customer reference. Nvidia says OpenAI trained and served GPT-6 Astra on an OCI Supercluster in Abilene with more than 100,000 Nvidia Blackwell GPUs. Uday Ruddarraju, OpenAI's chief technology officer of compute, is quoted: "We are bringing Vera into production on OCI to support these workloads at scale." Nvidia's Ian Buck adds that "AI factories must be productive, durable and fungible."

What the release does not contain is the number that defines "multi-gigawatt." It uses the phrase and "gigawatt-class power and cooling" without a capacity target, a date or a deal value. Our earlier coverage of Google Cloud's Vera Rubin rollout shows what a named ceiling looks like. This one has none.

What Oracle added around Nvidia

The more unusual news is the list of chips Nvidia does not make. According to a news summary of Oracle's announcement, Oracle will offer Google tensor processing units on OCI and bring Cerebras fast inference to the cloud, both planned for calendar 2027, and SambaNova will deploy inference chips in Oracle data centers with no timing given. Next Platform describes the TPU offer as the first time Google accelerators would be sold in a hyperscale environment other than Google's own. Oracle's own press release was not available to us when we ran this story, so these details rest on secondary coverage.

Next Platform lays out the intended division of labor. Google TPUs are pitched for model training, post-training and inference. Cerebras's CS-4 system, built on a chip it calls WSE-3 Turbo, targets fast inference, and Cerebras claims up to 30 times the per-user tokens per second of GPUs. SambaNova's SN50 pairs with GPUs: the GPUs handle prefill, the first stage of reading a prompt, and SambaNova's reconfigurable dataflow units handle decode, the stage that generates the answer. That split is the most concrete technical idea in the package. We have not seen an independent benchmark for any of these claims.

Why Oracle wants options

Oracle's numbers explain the urge to hedge. Next Platform reports sales of $67.4 billion in fiscal 2026, up from $57.4 billion, and OCI revenue of $7.4 billion in the first quarter of fiscal 2027, up 121% year over year. Remaining performance obligations, the contracted backlog, hit $664 billion, up $209 billion. Capital expenditure was $55.7 billion in fiscal 2026 and, per the article, could exceed $90 billion this fiscal year.

A backlog that size ties Oracle's fortunes to a few AI customers and to Nvidia's supply. Nathan Thomas, a senior vice president for OCI, framed the multi-vendor move as customer choice: "We don't want to put ourselves in a position where we're dictating any of that stuff." Choice is also insurance. If one supplier slips, a TPU or a wafer-scale system is a second source for the same inference demand.

The Neuron's October 28 digest also notes a report, which it attributes to Reuters citing The Information, that OpenAI plans to buy Google TPUs. Neither company had confirmed it, and we could not open the Reuters piece, so we treat it as unverified. If it is true, it would make Oracle's TPU offer less odd: its biggest customer would already be a TPU buyer.

The networking and the people

The Nvidia release lists more than GPUs. Alongside Vera Rubin it names Oracle's Acceleron networking, Nvidia's Spectrum-X Ethernet with Spectrum-6 switches and BlueField-4 data processing units, and OCI Kubernetes Engine as part of the stack. Mahesh Thiagarajan, president of Oracle Cloud Infrastructure, is quoted: "Oracle Cloud Infrastructure gives customers the performance, security and scale to advance frontier AI." Sreekanth Pothanis, an engineering leader for compute infrastructure at OpenAI, says: "As our frontier models advance, we need infrastructure that can keep pace." Both are vendor statements, and neither is a number.

What this does not tell us

Three things remain unknown. The size of the Nvidia commitment, in gigawatts or dollars. The price of TPU access on OCI, and whether any customer has signed for it. And the power: gigawatt-class sites need grid connections and generation that no press release supplies. Our behind-the-meter power explainer covers why that last item is the usual bottleneck.

What we would watch

Look for gigawatts under contract in Oracle's next filing, because a plan without a capacity number is a press release. Watch whether the TPU and Cerebras offers carry named customers in 2027. And watch whether Nvidia's willingness to sit beside rivals on a shelf is a sign of confidence or of a supply chain customers no longer trust to be single-sourced.