Oracle and Google announced that Google Tensor Processing Units will be offered through Oracle Cloud Infrastructure, extending Google's custom AI accelerators to a rival hyperscale cloud for the first time. Both companies say deployment will begin in calendar 2027 and expand over multiple years.
A new distribution channel for Google's custom silicon
Google designed TPUs as vertically integrated systems spanning chips, optical interconnects, cooling and software. Oracle says OCI will provide the data-center power, cooling and networking needed to operate those systems at scale, placing TPUs alongside the GPUs and other accelerators already available through its cloud.
The initial portfolio is expected to include Google's eighth-generation TPU 8t for training and TPU 8i for inference. Google reports that 8t delivers nearly three times the compute performance per pod of the previous generation, while 8i targets low-latency inference and agent workloads; those comparative figures are vendor-reported.
Competition is shifting from chips to access and placement
Offering TPUs in OCI gives customers another way to locate training and inference near Oracle databases and enterprise data without moving workloads to Google Cloud. It also turns Google's accelerator business into a broader infrastructure platform rather than a capability tied exclusively to its own cloud.
No current capacity was announced. The practical effect depends on deployment regions, pricing, quotas and software compatibility that the partners have not yet detailed. The agreement is material because it opens a new hyperscale distribution path, but customers cannot treat the planned 2027 rollout as generally available service today.