Oracle plans Google TPU and Cerebras cloud services for 2027

Oracle said on October 27 that it plans to offer Google TPUs on its cloud in 2027. Cerebras inference is also planned for Oracle Cloud Infrastructure, or OCI, that year. Customers would gain more choices for training models and generating responses within one cloud.

The timing differs from Nvidia products announced the same day. Oracle and Nvidia said Vera Rubin NVL72 systems and Vera CPUs are generally available on OCI. Expansion of Vera Rubin infrastructure to multi-gigawatt scale remains a plan.

TPUs are Google’s purpose-built accelerators for AI training and inference. The companies say OCI will become the first hyperscaler outside Google to offer them. Oracle will supply the power, cooling and data-center networking needed to deploy TPUs within its own infrastructure, connecting the compute offering to the facilities that operate it.

In their Google Cloud blog, the companies say TPUs inside OCI would let customers fine-tune models and run inference near existing databases, reducing cross-cloud transfers. Oracle’s network would connect storage and databases with TPU clusters.

Oracle also announced networking changes for distributed AI workloads. New communications software helps jobs use healthier paths between servers. Automatic path selection, failover and congestion control are part of the front-end network expansion. Oracle describes the goal as keeping expensive accelerators productive while data moves through the system.

For software portability, Google points to TorchTPU, its open-source stack for using PyTorch on TPUs. The company says teams can train and serve models with tools from the existing ecosystem. It identifies productive compute time as a key training metric, with latency and cost per token central to inference economics.

Cerebras integrates compute and memory on a wafer-scale processor to reduce inter-chip data movement. A wafer is the substrate on which semiconductor circuits are fabricated.

Cerebras targets interactive agents whose repeated model calls add waiting time. The planned service covers AI labs, enterprises and developers; OCI speed measurements and prices are unannounced.

Agents also need CPUs to retrieve data, execute code and call tools between model interactions. Oracle and Nvidia say customers can scale CPU and GPU resources independently on OCI. That separates demand for model computation from the surrounding software work. Oracle also provides hardware-based isolation for workloads running untrusted AI-generated code.

Oracle’s separate Vera CPU post describes bare-metal and flexible virtual-machine offerings and an initial supply of 150,000 CPUs. The disclosed hardware configuration has 176 CPU cores and 1 TB of system memory. It says availability will cover selected regions this quarter. The advertised performance gain of up to 80% against leading x86 CPUs comes from Nvidia platform benchmarks, rather than OCI instance testing. Results vary with workload and configuration.

TPU deployment on OCI is scheduled to begin in 2027 and ramp over several years. Oracle’s broader infrastructure post sets a gigawatt-scale capacity goal. Initial regions, customer allocations and OCI prices for the Google TPU and Cerebras services are absent from these announcements.