Oracle Cloud widens AI chip lineup with first Google TPUs, plus Cerebras and SambaNova

Oracle used its Oracle AI World event to widen the menu of AI accelerators available on Oracle Cloud Infrastructure, announcing new agreements with Google and Cerebras Systems while expanding its long-running partnership with Nvidia.

Under the plan, Oracle will deploy Nvidia's Vera Rubin systems across OCI Superclusters at multi-gigawatt scale, introduce Google's tensor processing units for the first time, and offer Cerebras' specialised inference technology, according to Tech Wire Asia. Oracle is also collaborating with Intel-backed SambaNova Systems on deploying SambaRack SN50 inference chips in OCI data centres.

The Google agreement makes OCI the first hyperscale cloud platform outside Google Cloud to offer Google's TPUs. Google's eighth-generation TPU portfolio includes two architectures built for different jobs: TPU 8t for training large AI models, and TPU 8i for inference, including agent-style applications that make several model calls before finishing a task.

Oracle said the TPUs will support model training, fine-tuning and inference, with Oracle supplying the data centre infrastructure, power, cooling and networking required to run them. The processors are expected to become available in 2027.

On the inference side, Cerebras brings its Wafer-Scale Engine, which combines compute and memory on a single wafer-scale processor to reduce data movement between chips. Cerebras says the design supports faster token generation for applications that involve multiple processing steps; the service is scheduled to arrive on OCI in 2027.

Oracle also plans to deploy SambaNova's SambaRack SN50 chips for inference, giving model developers another computing option. "Different AI workloads have different compute needs," said Mahesh Thiagarajan, president of Oracle Cloud Infrastructure, adding that access to multiple processor architectures gives customers more flexibility on performance, throughput and cost.

The Nvidia relationship is being scaled rather than replaced. Nvidia Vera Rubin NVL72 systems and Vera CPUs are already generally available on OCI, with an HGX Rubin NVL8 configuration using eight Rubin GPUs planned as a smaller deployment option. The Vera CPUs take the partnership beyond GPUs, and Oracle says customers can scale CPU and GPU capacity independently. On networking, OCI Superclusters pair Oracle's Acceleron technology with Nvidia's Spectrum-X Ethernet, Spectrum-6 switches and BlueField-4 data processing units, and Oracle plans to use Acceleron to connect Google TPU clusters to its databases and storage services. Neither company disclosed the financial value of the expansion.

The expansion rests on heavy spending. In its first quarter of fiscal 2027, ended August 31, 2026, Oracle reported cloud infrastructure revenue of $7.4 billion, up 121% year over year, with capital spending of $28.5 billion versus $8.5 billion a year earlier, remaining performance obligations of $664 billion, and 850 megawatts of added data centre capacity. It kept its fiscal 2027 capex forecast at $90 billion to $95 billion.

Because the Google TPU and Cerebras services only arrive in 2027, the multi-architecture strategy will take time to show up in revenue, and near-term OCI capacity still leans on Nvidia hardware that is already shipping. The open questions are how much third-party TPU capacity Oracle can actually secure and sell, and how it explains the pace of that capex to investors.