Oracle is expanding its artificial-intelligence infrastructure agreement with Nvidia to what the companies describe as multi-gigawatt scale, centered on servers using the forthcoming Vera Rubin platform. At the same time, Oracle Cloud Infrastructure will add Google tensor processing units and inference systems from Cerebras and SambaNova, turning a larger Nvidia commitment into a broader multivendor strategy rather than an exclusive supplier relationship.
The gigawatt language measures planned electrical capacity for computing infrastructure; it is not a disclosed purchase price, recognized revenue or completed installation. Nvidia chief executive Jensen Huang has previously associated one gigawatt of Nvidia systems with as much as US$40 billion in potential revenue. Multiplying that illustrative figure by an unspecified number of gigawatts would not produce a reliable value for Oracle’s agreement because configuration, delivery, pricing and customer funding remain unknown.
The Nvidia portion is still strategically central. Vera Rubin is the company’s next architecture for large-scale model training and inference, and Oracle expects to deploy servers built around it. Yet those systems will omit some Nvidia networking components in favor of Oracle’s Acceleron technology, showing that the cloud provider intends to control more of the data path rather than buy every layer as a bundled Nvidia stack.
Google’s inclusion is notable because Oracle will offer tensor processing units to customers for the first time. TPUs provide an alternative accelerator architecture already used extensively in Google’s own AI services. Availability inside Oracle’s cloud could let customers choose hardware according to model, software framework, performance and cost without moving their entire application estate to a different provider.
Cerebras and SambaNova address another part of the market. Nvidia remains dominant in training, where models absorb enormous datasets and computing clusters, but inference—the repeated execution of trained models for real users—is more open to specialized designs. Oracle plans to offer both companies’ inference chips, and Cerebras says the arrangement expands its access to large customers through installations inside another operator’s data centers.
A diverse catalog can reduce dependence on one supplier and give Oracle leverage when matching scarce chips to workloads. It also raises integration demands: customers need compatible software, predictable networking, security, support and pricing across architectures. Accelerators that look strong in a benchmark still have to deliver utilization and economics within complete clusters, not only peak performance on an isolated task.
Financing is another unresolved layer. The companies did not disclose the agreement’s value or payment structure, and large cloud customers increasingly help finance the chips and facilities dedicated to their projects. Oracle kept its fiscal-2027 capital-expenditure forecast at US$90 billion to US$95 billion, a range that signals extraordinary construction but does not identify how much belongs to Nvidia, Google, Cerebras, SambaNova or customer-backed capacity.
Execution will be measured through signed customer commitments, delivery schedules, energized data-center capacity, software availability and returns on invested capital. The announcement shows Oracle trying to scale its relationship with the market leader while preventing that relationship from defining its entire AI platform. It does not yet demonstrate deployed multi-gigawatt capacity, a quantified contract value or profitable demand for every architecture in the new catalog.
