Oracle made several infrastructure announcements at Oracle AI World in Las Vegas on October 27, 2026. Taken together, they show one cloud offering several different accelerator architectures, each aimed at a different part of an AI workload.
Google TPUs outside Google
Oracle and Google announced that Google's Tensor Processing Units will be available through Oracle Cloud Infrastructure, "marking the first time Google TPUs will be offered on a hyperscale cloud platform outside Google." Oracle will provide "the power, cooling, and datacenter networking." The release names Google's eighth-generation parts, the TPU 8t for training and the TPU 8i for inference. It repeats Google's claim that the TPU 8t delivers "nearly 3x the compute performance per pod over the previous generation."
The timing is not immediate. Google TPUs on OCI "will be available in calendar year 2027, ramping over a multi-year period." The release gives no pricing, no regions and no named customers.
Vera Rubin and Vera CPU, available now
A separate NVIDIA and Oracle release says Vera Rubin NVL72 systems and NVIDIA Vera CPUs are generally available on OCI. Oracle plans to grow its Superclusters to multi-gigawatt scale and to add HGX Rubin NVL8, an eight-GPU x86 server with two 64-core Intel Xeon 6 CPUs, 4 TB of DDR5 memory and 6.4 Tbps of RDMA networking. The release also says OpenAI trained and served GPT-6 Astra in Abilene on an OCI Supercluster with more than 100,000 Blackwell GPUs.
The Vera CPU part is aimed directly at agents. The release says agents "call tools, execute code and retrieve information," and that slow execution of those CPU-heavy steps can "leave GPUs waiting between model calls." Customers can scale CPU and GPU resources independently. This matches a pattern this digest covered on October 20, when Modal said it would run agent sandboxes on Vera CPUs.
What the mix means for engineering teams
Offering several accelerators in one cloud gives buyers choice, but it does not make code portable. CUDA kernels, custom Triton kernels and vendor-specific quantization formats do not run on TPUs as written. TPU workloads are normally written for the XLA compiler through JAX or PyTorch/XLA. A team that wants to move between architectures needs model code that stays at the framework level, a test suite that checks numerical results on each backend, and cost figures measured on its own workload.
The CPU announcement points to a cost that is easy to miss. In an agent loop, the GPU often sits idle while a tool runs, a test executes or a document is fetched. If that wait is long, buying faster GPUs does not help much. Measuring how long each step waits on the CPU tells you whether the bottleneck is the model or the tools.
What remains uncertain. The TPU offering is more than a year away and has no published price. Oracle's performance claims are vendor claims. The releases do not say whether TPUs on OCI will connect to the same networking fabric as the NVIDIA systems, or how TPU capacity will be scheduled for customers.
Analysis: For a small team the practical step is to keep options open without paying for them now. If Belgavi's Book Store ran its own return-classification model, Tanishka could keep the training and serving code in plain PyTorch or JAX, record which custom kernels it depends on, and log the GPU idle time during each agent step. Those three facts decide whether a second accelerator or a faster CPU would actually lower cost.
Sources and evidence
Oracle said on October 27 that Google TPUs (TPU 8t and 8i) will be available on OCI in calendar 2027, the first hyperscale cloud outside Google to offer them, with no pricing yet; NVIDIA Vera Rubin NVL72 and Vera CPUs are generally available on OCI now, with Vera aimed at the CPU-heavy tool and code steps that leave GPUs idle in agent loops. Several accelerator types in one cloud adds choice, not code portability.