Oracle is expanding its relationship with Nvidia while adding competing AI processors to its cloud infrastructure. The move reflects a broader change in artificial intelligence: businesses increasingly need to match different computing architectures to the economics of individual workloads.
Oracle has announced a major expansion of its artificial intelligence infrastructure strategy, combining a new multigigawatt arrangement with Nvidia and plans to introduce processors developed by Google, Cerebras and SambaNova into its cloud platform.
The announcements, made during Oracle AI World in Las Vegas on 27 October, reflect the company’s efforts to position Oracle Cloud Infrastructure (OCI) as a platform capable of supporting several competing AI computing architectures.
Rather than selecting a single chip supplier, Oracle intends to offer customers different technologies for training, deploying and operating artificial intelligence models.
The development could have significant consequences for cloud computing costs, enterprise AI investment and competition among semiconductor manufacturers.
According to Oracle’s infrastructure announcement, the company is expanding its partnerships with Nvidia, AMD and Intel while adding new specialised AI processors.
Nvidia remains central to the strategy
Oracle and Nvidia have announced plans to expand OCI Superclusters using Nvidia’s Vera Rubin architecture to a multigigawatt scale.
The companies also announced the general availability of Vera Rubin NVL72 systems and Vera CPUs within Oracle’s cloud infrastructure.
These systems are designed to support demanding AI workloads, including large-scale model training and inference.
Training involves processing substantial quantities of data to develop or improve a model.
Inference refers to the computational work performed when an already-trained model responds to user requests or carries out tasks.
Both require considerable computing power, but their technical requirements are not identical.
Some applications prioritise high memory capacity and fast communication between processors. Others depend more heavily on low latency, energy efficiency or the ability to process large numbers of requests economically.
Nvidia remains the dominant supplier of advanced accelerators used for many demanding AI workloads.
Oracle’s new arrangements do not represent a withdrawal from Nvidia technology.
On the contrary, they involve a substantial expansion of its deployment.
However, Oracle is simultaneously preparing to offer customers alternatives.
Google enters another cloud platform
The most consequential change concerns Google’s Tensor Processing Units.
Under a partnership announced by Google and Oracle, Google’s purpose-built AI processors will become available through OCI beginning in 2027.
Oracle will become the first hyperscale cloud provider outside Google to offer TPUs.
The processors are designed for artificial intelligence workloads and have been developed through close integration of chip design, software and supporting infrastructure.
Google’s eighth-generation TPU family includes architectures intended for both model training and inference.
For Oracle customers, access to these processors could provide additional options when evaluating performance, energy consumption and operating costs.
For Google, offering TPUs through another cloud platform expands the commercial reach of its processor technology beyond its own infrastructure.
The arrangement also illustrates how competition in AI infrastructure differs from competition in conventional software markets.
Google and Oracle can compete for cloud customers while simultaneously cooperating over the technologies those customers use.
Cerebras and SambaNova target inference
Oracle is also expanding its planned use of specialised inference technologies.
Cerebras produces processors built around a wafer-scale architecture, designed to reduce the communication and memory-access constraints associated with certain AI workloads.
Under the new Cerebras agreement, Oracle intends to offer Cerebras inference capabilities through OCI from 2027.
SambaNova, another supplier of specialised AI processors, is also expected to join Oracle’s infrastructure portfolio.
These technologies are particularly relevant to applications involving AI agents.
An agent may need to make several model requests, retrieve information, execute software tools and evaluate intermediate results before completing a task.
Where such processes require repeated inference, delays can accumulate.
Improving inference speed and cost efficiency can therefore materially affect the usability and economics of enterprise AI applications.
However, specialised processors do not necessarily outperform conventional GPUs across every workload.
Actual performance depends on model architecture, software compatibility, batch sizes and deployment requirements.
A more complicated market for enterprise buyers
Oracle’s strategy introduces additional choice, but also creates more complex procurement decisions.
Businesses considering AI deployment will increasingly need to compare systems based on the total cost of running a workload rather than simply the advertised performance of an individual chip.
That includes hardware rental charges, memory requirements, networking, electricity consumption, software compatibility and engineering costs.
Moving between processor architectures may also require changes to applications and development workflows.
Consequently, hardware diversity does not automatically eliminate vendor dependence.
Some workloads may remain closely tied to the software environments and tools associated with particular suppliers.
Oracle’s Acceleron networking platform is intended to help connect different computing architectures to its broader cloud infrastructure.
The company argues that this approach will give customers greater flexibility as models and applications evolve.
The financial question remains
The expansion comes with substantial investment requirements.
Oracle has maintained plans for approximately $90 billion to $95 billion in capital expenditure during its 2027 financial year, according to reporting on the announcements.
The company has not disclosed the full commercial value of its latest Nvidia arrangements.
Much will depend on whether customer demand for AI services can support the costs of acquiring processors, constructing data centres and financing their operation.
As QuestComms recently examined in its reporting on the economics of enterprise AI, rapid growth in AI adoption does not remove the importance of infrastructure costs and sustainable margins.
Oracle’s announcement signals a shift towards cloud platforms offering multiple AI processor architectures.
The decisive question for enterprise customers will be whether that choice produces measurable improvements in cost, performance and operational flexibility.
