Oracle Adds Google TPUs, Cerebras Wafer Chips, and NVIDIA Vera Rubin to Cloud Infrastructure
Oracle becomes the first external hyperscaler to offer Google 8th-gen TPUs on OCI, alongside multi-gigawatt NVIDIA Vera Rubin NVL72 and Cerebras wafer-scale chips.
At Oracle AI World on October 28, 2026, Oracle Cloud Infrastructure (OCI) announced a multi-silicon expansion, becoming the first third-party hyperscaler to offer Googles 8th-generation TPUs (TPU 8t and 8i). The expansion also deploys multi-gigawatt NVIDIA Vera Rubin NVL72 clusters, general availability of NVIDIA Vera CPUs, and high-speed Cerebras Wafer-Scale Engine inference.
Google TPUs Break Cloud Exclusivity on Oracle OCI
In a landmark cross-hyperscaler agreement announced October 28, 2026, Oracle became the first external cloud provider to offer Googles custom Tensor Processing Units. Starting in 2027, OCI will deploy Googles 8th-generation TPU portfolio, including the TPU 8t for large-scale foundation model pretraining and the TPU 8i optimized for multi-agent inference. The partnership allows enterprise organizations to run TPU workloads directly alongside relational data in Oracle Database and Autonomous Database environments without inter-cloud egress penalties.
Multi-Gigawatt NVIDIA Vera Rubin Clusters and Vera CPUs
Simultaneously, Oracle expanded its deep infrastructure pact with NVIDIA, deploying flagship Vera Rubin NVL72 liquid-cooled superclusters at multi-gigawatt scale alongside a new 8-GPU HGX Rubin NVL8 form factor. Oracle confirmed that NVIDIA Vera standalone CPUs are now generally available on OCI, allowing developers to decouple CPU and GPU scaling for agentic reasoning, code execution, and database indexing while integrating NVIDIA Spectrum-X Ethernet and BlueField-4 DPUs through Oracles proprietary Acceleron networking fabric.
Cerebras Wafer-Scale and SambaNova Inference Scaling
Targeting ultra-low latency requirements for real-time agentic workflows, OCI will also deploy Cerebras Systems Wafer-Scale Engine (WSE), which integrates compute and memory onto a single giant silicon wafer to eliminate inter-chip communication delays. Complemented by plans to roll out SambaNova Systems SambaRack SN50 inference accelerators, Oracles multi-architecture strategy provides enterprises with specialized hardware tailored specifically across training, agent orchestration, and sub-second inference.