Oracle announced a sweeping expansion of its cloud AI chip lineup on Tuesday, October 27, 2026, at Oracle AI World in Las Vegas, deepening its partnership with Nvidia while adding rival hardware from Google, Cerebras and SambaNova to Oracle Cloud Infrastructure. The headline move takes Oracle Cloud Infrastructure Superclusters powered by Nvidia Vera Rubin NVL72 systems to multi-gigawatt scale, and makes Vera Rubin NVL72 systems and Vera CPUs generally available on the platform. At the same time, Oracle said it will offer Google’s tensor processing units for the first time, starting in 2027, making it the first major cloud provider outside Google itself to do so.
Vera Rubin Goes Live On Oracle Cloud
The Nvidia expansion is both a product launch and a capacity plan. Each Vera Rubin NVL72 rack connects 72 graphics processors through Nvidia’s NVLink networking, and Oracle pairs those racks with its own Acceleron networking, Nvidia Spectrum-X Ethernet, BlueField-4 data processing units, and gigawatt-class power and cooling. The superclusters are designed to scale across hundreds of thousands of graphics processors, a scale aimed at training and serving frontier models that keep getting larger.
Oracle also plans HGX Rubin NVL8, a smaller eight-GPU system in an x86 server. The other half of the announcement is about the work agents do between model calls. Vera CPUs are now generally available, aimed at CPU-heavy tasks such as calling tools, executing code and retrieving data. Nvidia’s partnerships keep expanding across the AI stack, and Oracle says customers can scale CPU and GPU resources independently, so a slow tool call does not leave an expensive graphics processor idle. OpenAI is an early example, having trained GPT-6 Astra on an Oracle supercluster with more than 100,000 Blackwell graphics processors.
Google TPUs Arrive In 2027
The most surprising part of the announcement is the Google deal. Oracle will bring Google’s custom tensor processing units to its cloud starting in 2027, with the TPU 8t for training and the TPU 8i for inference, in a phased rollout expected to scale over several years. Until now, customers who wanted Google’s chips largely had to go to Google Cloud. Oracle is betting that choice itself is the selling point: different workloads have different needs for performance, throughput and cost, and a customer should be able to pick the right chip inside one cloud rather than stitching several providers together.
Inference Gets More Competition
Google is not the only new option. Oracle also plans to offer chips from Cerebras Systems and SambaNova Systems, both focused on inference. Cerebras chief executive Andrew Feldman called the deal a massive expansion of his company’s customer base of tier-one customers. Oracle is also tightening its grip on networking: even the Nvidia-based systems will skip some Nvidia networking chips in favour of Oracle’s in-house Acceleron networking in some configurations, which can reroute traffic when links or network cards fail. In a system with hundreds of thousands of processors, that resilience matters, since a single failed link can stall a costly training run.
The Money Question Oracle Did Not Answer
What Oracle did not disclose is just as important as what it did. The companies gave no value for the expanded Nvidia deal and no firm completion date or exact gigawatt figure for the multi-gigawatt expansion. Nvidia chief executive Jensen Huang has previously said Vera Rubin systems can generate up to $40 billion in revenue per gigawatt for the chipmaker, which gives a sense of scale, but it is not a price tag for Oracle’s purchase. Oracle also did not say how it plans to finance the chip purchases, ahead of an investor day this week. During an earnings call last month, Oracle maintained a capital spending target of about $90 billion to $95 billion for the 2027 fiscal year. Oracle has already been spending heavily to build its AI ecosystem, and investors will now want to know whether customer commitments are carrying the cost.
Why It Matters
Oracle’s announcement matters because it signals a shift in how cloud AI is sold. For years, choosing a cloud meant largely choosing a chip supplier, and for most customers that meant Nvidia. Oracle is now trying to sell choice itself: Nvidia for the biggest training jobs, Google’s TPUs for workloads tuned to Google’s architecture, and Cerebras and SambaNova for fast, specialised inference, all inside one platform. If that works, competitive pressure moves from who has the fastest chip to who can offer the best mix at the best price. The risk is equally clear: multi-gigawatt buildouts are expensive and unforgiving, and a broad chip menu only helps if Oracle can deliver capacity on time.
Background
Oracle Cloud Infrastructure has grown from a challenger cloud into one of the main hosts for frontier AI training, built on successive generations of Nvidia systems and large customer commitments. The Vera Rubin generation follows the Blackwell systems that powered the previous wave of model training, and Oracle’s decision to pair it with rival chips marks the most open version of its cloud strategy yet. Google’s own AI ambitions keep expanding in parallel, from world models to robotics, which makes its willingness to sell TPUs through Oracle a notable change in how the largest AI companies treat each other as both rivals and suppliers.
Source: Reuters Oracle chip report and Nvidia Oracle announcement
0 Comments