Summary
- Google TPUs will be introduced on Oracle Cloud Infrastructure from 2027.
- OCI will become the first hyperscale cloud provider outside Google to offer the processors.
- Regional availability, installed processor quantities, and initial customer capacity have not been established.
Oracle and Google have agreed to bring Google’s Tensor Processing Units to Oracle Cloud Infrastructure from 2027, allowing customers to access the specialised AI processors through a hyperscale cloud provider outside Google’s own platform.
The companies intend to begin deploying TPU capacity during calendar year 2027, with a broader expansion planned over several years.
OCI will become the first hyperscale cloud provider outside Google to offer TPU infrastructure. Neither company has specified the initial number of processors, the quantity of power allocated to the first installations, or the OCI regions that will receive the systems.
Google developed TPUs as specialised processors for machine learning applications. Their design combines computing hardware with networking, memory, cooling, and software intended to support training and inference workloads.
Large AI systems require more than raw arithmetic capability. Processor performance depends on the rate at which data can be supplied from memory, communicated between devices, and transferred into and out of the wider computing environment.
Google’s approach has involved developing these elements together, allowing hardware and software to be optimised for the characteristics of supported models and workloads.
The new agreement will make this architecture available within Oracle’s existing cloud infrastructure portfolio, adding another option alongside the processors and accelerator systems already offered by OCI.
Different hardware architectures can suit different workloads. Training a large model involves repeated processing of extensive datasets and coordination across many devices, while production inference may place greater emphasis on response latency, sustained throughput, and operating cost.
Specialised processors can improve efficiency for supported applications, although the realised benefit depends on the model, numerical precision, memory requirements, and software environment.
Oracle’s additional hardware offering will require appropriate integration with its computing, networking, storage, and operational systems. A TPU installation also needs suitable electricity distribution and thermal infrastructure, particularly where large quantities of accelerators operate together.
The physical arrangement of processors influences network performance and how efficiently data can move between systems. Fibre links, switching equipment, memory systems, and supporting software all contribute to the resulting computing capability.
Power and cooling requirements will be relevant to the locations selected for deployment. Higher-density accelerator systems can require facility infrastructure designed around substantial and concentrated electrical loads.
The agreement also introduces software considerations. Customers using TPUs need access to supported model frameworks, compilers, and development tools, and workloads developed for another accelerator architecture may require adaptation.
That integration work can influence the practical economics of adopting a new computing platform, alongside the equipment’s throughput and electricity requirements.
Oracle is seeking to offer a broader range of processing technologies within OCI so that customers can select infrastructure suited to particular training and inference applications.
That approach may improve flexibility but also increases the number of hardware types that must be maintained, monitored, supported, and scheduled within the cloud environment.
For European customers, data residency and processing location will depend on the OCI regions in which TPU capacity becomes available. The agreement does not establish that the first installations will be located in Europe.
The companies have said the planned programme could ramp to gigawatt scale over multiple years. This describes the anticipated expansion trajectory, not a contracted electricity connection or confirmed first-year capacity at any OCI region.
Google’s eighth-generation TPU portfolio divides training and inference responsibilities between TPU 8t and TPU 8i. Google claims nearly three times the compute performance per pod for TPU 8t over the preceding generation, a comparison that should not be interpreted as the measured performance of a future Oracle installation.
The joint development also involves optical networking, power delivery, liquid cooling and compiler software, rather than only transferring processor hardware into Oracle facilities. Those supporting systems will influence which campuses can host the equipment and how effectively the installed processors can work together on distributed models.
Oracle and Google have not provided a European capacity allocation or deployment timetable. The proposed gigawatt-scale trajectory therefore cannot be presented as new electricity demand at a named European facility or as a new European grid connection agreement.
The initial deployment therefore remains a future infrastructure commitment. Further information about the installed capacity, geographical distribution, pricing, and supported services will be required before customers can assess the availability of TPU resources for particular projects.
