Google Breaks TPU Exclusivity, Bringing AI Chips to Oracle
Google will offer its custom Tensor Processing Units outside its own ecosystem for the first time through a partnership with Oracle Cloud Infrastructure. Beginning in 2027, Oracle enterprise customers will be able to run machine learning workloads on Google's TPU 8t and TPU 8i hardware.
Breaking a decade of hardware lock-in
For the first time since their inception, Google's proprietary AI chips will live in a competitor's data center.
Google has jealously guarded its Tensor Processing Units for a decade, using the proprietary silicon as the primary moat to draw machine learning developers into Google Cloud. That walled-garden era is officially ending. Announced at Oracle's annual AI World conference, Oracle Cloud Infrastructure (OCI) will make Google's newest accelerators directly accessible to its own enterprise base.
The agreement marks the first time any rival hyperscaler has been licensed to sell TPU capacity. Rather than forcing clients into a single cloud environment, both companies are embracing a multi-cloud reality where compute capacity often dictates where models get trained.
Google TPU 8th-Gen on OCI: availability and silicon roles
| Spec | Detail |
|---|---|
| Hardware models | TPU 8t (training), TPU 8i (inference) |
| First third-party provider | Oracle Cloud Infrastructure (OCI) |
| Launch year on OCI | 2027 |
| Primary architecture target | AI model training and inference |
TPU 8t and 8i arrive on OCI in 2027
The partnership focuses on Google's eighth-generation silicon unveiled earlier this year: the TPU 8t, optimized for complex model training, and the TPU 8i, tailored for efficient real-time inference. Unlike general-purpose graphics cards, Google designed these chips from the gate level solely for neural network operations.
OCI President Mahesh Thiagarajan pointed out that varying enterprise workloads require different computing architectures. By adding TPUs, Oracle can offer its customers an immediate counterweight to Nvidia's hardware dominance. Even as enterprise AI spending faces scrutiny over real-world payoff, demand for non-GPU compute clusters that lower the cost of serving large models remains acute.
Google selling its crown jewels to Larry Ellison would have sounded absurd two years ago. Yet it makes tactical sense: Nvidia's massive margins and supply constraints have opened a window. If Google wants TPUs to become an industry-wide software target rather than an internal convenience, it has to put them everywhere developers deploy.
Availability and timeline
Access to Google's TPU 8t and 8i on Oracle Cloud Infrastructure will roll out globally across OCI regions starting in 2027. Neither company has released tier pricing or reservation terms yet, though OCI typically prices specialized instances under sustained-use commitments and bare-metal configurations.
Frequently asked questions
When can developers rent Google TPUs on Oracle Cloud?
Oracle plans to make both the TPU 8t and TPU 8i available to OCI customers starting in 2027.
What is the difference between Google TPU 8t and TPU 8i?
The TPU 8t is optimized specifically for large-scale model training, while the TPU 8i is designed for low-latency AI inference workloads.
Sources
- 'The next generation of AI will require a diverse range of compute': Oracle deal means Google TPUs are coming to a rival cloud for the first time | TechRadar. techradar.com, 2026. https://www.techradar.com/pro/the-next-generation-of-ai-will-require-a-diverse-range-of-compute-oracle-deal-means-google-tpus-are-coming-to-a-rival-cloud-for-the-first-time