Oracle links its Rubin launch to a plan for more than 3GW of NVIDIA capacity

Oracle and NVIDIA announced general availability of Vera Rubin NVL72 systems and standalone Vera CPUs on Oracle Cloud Infrastructure on 27 October. Oracle also says it plans more than 3GW of NVIDIA GPU capacity by the end of 2027. That broader capacity target spans NVIDIA hardware; it does not describe an already completed, all-Rubin installation.[1][2]

The two developments work on different clocks. A cloud product can enter service while the larger estate of buildings, power systems and cooling continues to grow. For customers choosing where to train or run a model, both matter: the configuration available for a job now, and the operator’s ability to expand it later.

Three offers, with different delivery stages

Oracle’s CPU notice describes an initial 150,000-processor rollout across its regions. Although headed as a general-availability announcement, it ends by saying the CPUs will be available in selected regions this quarter. The practical reading is a launch with a staged regional rollout, for which a customer still needs a location-specific offer.[4]

The proposed NVL8 option gives the expansion another scale. It combines eight Rubin GPUs with Intel Xeon 6 processors and high-speed networking, preserving an x86 server configuration for customers whose applications use it. That is a different deployment choice from the much larger NVL72 rack.[1]

A campus designed to survive its first generation of chips

Oracle describes gigawatt-class campuses built for density within a two-kilometre radius, with power delivery and liquid cooling serving systems across multiple racks. The company says these sites can accommodate successive NVIDIA platforms and shift capacity between frontier-model training and production inference.[3]

That design has an economic purpose even before attaching a price to a GPU hour. A building dedicated too narrowly to its first workload could become difficult to use when demand changes. Infrastructure that can accept another generation of equipment, or move from training a model to serving it, gives the operator more ways to keep the site productive.

Communication is part of that flexibility. Oracle’s described network spreads traffic over separate planes, allowing it to move away from congestion or a failing path. The intended effect is to keep a large job doing useful work while parts of the system encounter trouble.[3]

The engineering claim is consequently about continuity as much as size: making a growing installation behave as a usable computing service. The 3GW plan supplies the scale of the ambition. Regional availability, delivered capacity and sustained operation will show how much of that ambition customers can actually use.

Report an error or suggest a correction