Oracle partners with Google for TPU hosting

Oracle has agreed to host Google's TPU chips on its Oracle Cloud Infrastructure (OCI) beginning in 2027. This marks the first time a major cloud provider outside of Google will offer access to Google's TPU technology to its customers. The announcement was made during Oracle AI World, a conference currently taking place in Las Vegas.

Infrastructure setup for TPU hosting

Oracle will provide the necessary infrastructure to host the TPU chips in its data centers. This includes electrical power, cooling systems, and network connectivity. OCI customers will be able to access Google's TPU accelerators without needing to use Google Cloud services.

TPU capabilities for AI workloads

According to both companies, Google's TPU chips are designed to handle the full lifecycle of artificial intelligence tasks. This includes training large AI models, post-training processes, large-scale inference, and processing related to AI agents. Oracle emphasizes the scalability of these workloads, describing it as nearly linear.

Two TPU generations targeted

The agreement focuses on two generations of Google's TPU chips. The eighth generation includes two distinct architectures: TPU 8t, optimized for training AI models, and TPU 8i, designed for inference tasks such as reasoning applications, AI agents, and processing that requires long context inputs.

Oracle's multi-architecture approach

Oracle is positioning this offering as a way to provide multiple AI computing architectures within a single cloud environment. This would allow customers to choose between different infrastructure options based on their specific needs for performance, data throughput, and cost efficiency.

Google's expanded distribution strategy

For Google, the partnership represents a shift in its strategy for distributing TPU technology. Previously, Google's TPU chips were primarily used within its own cloud platform. This agreement allows for broader access to the technology through another major cloud provider.

Details still pending

Specifics regarding the pricing of TPU instances, the OCI regions where the chips will be deployed, the initial number of chips, and the first customers using the service are not yet available. These details are expected to be announced in the future.