Oracle is dramatically broadening the range of artificial-intelligence chips available through its cloud platform, expanding its partnership with Nvidia while adding Google TPUs and Cerebras processors. The strategy reflects a shift in AI infrastructure from a single dominant architecture toward a more diverse market for training and inference.
Oracle and Nvidia said OCI will expand to multi-gigawatt scale using Nvidia’s newest Vera Rubin systems.
Vera Rubin remains the flagship
Nvidia’s Vera Rubin platform is designed for large-scale AI training and inference. Oracle announced general availability of Vera Rubin NVL72 systems and Vera CPUs on OCI, while also introducing new Rubin configurations.
The scale of the deployment could make Oracle one of Nvidia’s largest cloud customers.
Google TPUs are coming to OCI
Oracle and Google announced that Tensor Processing Units will be offered through Oracle Cloud Infrastructure in 2027. OCI will become the first hyperscale cloud outside Google to offer the chips directly.
That gives customers an alternative to GPUs for workloads where TPUs offer better economics or performance.
Oracle is also adding Cerebras
Oracle plans to offer inference systems from Cerebras, a company known for wafer-scale processors designed for very large AI models.
The agreement could help Cerebras reach customers that want its technology without building dedicated infrastructure.
Networking is another point of differentiation
Oracle said some Nvidia-based systems will use its own Acceleron networking technology rather than relying entirely on Nvidia networking products.
AI clusters require enormous bandwidth between chips, so networking can materially affect performance and cost.
Training and inference are separating
Nvidia remains strongest in frontier-model training, where software compatibility and cluster scale are critical. Inference is becoming more fragmented because customers are more willing to use specialized chips optimized for particular workloads.
That creates opportunities for Google, Cerebras and other accelerator vendors.
Customers want flexibility
Large AI customers increasingly operate across multiple chip architectures and cloud providers. They want to match each workload with the hardware that offers the best cost, speed and availability.
Oracle is positioning OCI as a marketplace where customers can access several leading architectures without moving between clouds.
The financing question remains large
Oracle has maintained a fiscal 2027 capital-spending target of roughly $90 billion to $95 billion. Building multi-gigawatt AI infrastructure requires massive spending on chips, power and data centers.
Investors are therefore watching whether customer contracts and prepaid capacity can fund that expansion without putting too much pressure on cash flow.
AI infrastructure competition is intensifying
Amazon, Microsoft and Google have developed their own chips, while Nvidia continues to launch new generations at a rapid pace. Startups are also competing in inference and specialized computing.
Oracle’s strategy is to avoid betting on only one supplier and instead make its cloud compatible with several.
2027 will be a major deployment year
Google TPUs on OCI are expected to ramp through 2027, while Vera Rubin systems are already entering commercial deployment. The timing means customers will have more choices just as AI workloads become even larger.
For Oracle, the challenge is turning that hardware breadth into sustainable cloud growth while managing one of the largest capital-investment programs in the technology industry.
Sources
- Oracle press release, Oct. 27, 2026 — primary source
- Oracle Cloud Infrastructure announcement — primary source
- Reuters, Oct. 27-28, 2026
Customers may mix architectures inside one application
Future AI systems may train on one chip architecture and run inference on another. Some workloads could use Nvidia for frontier training, Google TPUs for specific large-scale services and specialized accelerators for low-latency inference.
Oracle’s strategy is designed around that reality. Rather than forcing customers into one hardware stack, OCI is trying to make different accelerators available behind a common cloud environment.
Software compatibility will decide adoption
Hardware diversity only works if developers can move models between systems without major rewrites. Nvidia’s CUDA ecosystem remains a major advantage. Google, Cerebras and Oracle therefore need strong compilers, frameworks and model-serving tools to make alternative chips easy to use.
Power availability may be the ultimate constraint
Oracle can buy more accelerators, but large AI clusters also require enormous amounts of electricity and cooling. In several markets, grid connections now take years to secure. That means the company’s ability to deploy Vera Rubin, TPUs and Cerebras systems at planned scale will depend as much on power contracts and data-center construction as on chip supply.
