A wide variety of AI chips will be available on Oracle Cloud Infrastructure (OCI). In addition to variants of Nvidia’s Vera Rubin architecture, Google TPUs will be available for the first time starting next year; the same applies to the large Cerebras chips. AMD, Intel, SambaNova, and OpenAI Jalapeño are expected to follow, though it is not yet known when.
Oracle unveiled a long list of hardware announcements at AI World in Las Vegas. OCI is adding or expanding seven compute platforms, and its own Acceleron network is also getting an upgrade. However, anyone looking to get started with AI computing power today still has relatively few options. Only the Nvidia Vera Rubin NVL72 and standalone Vera CPUs are generally available immediately.
The rest is planned, mostly for 2027. It’s possible that this date won’t be met, though the succession of new chip generations is fast enough that some urgency is needed to keep the hardware relevant and attractive.
The most notable development is the arrival of Google TPUs, which until now have been exclusive to Google Cloud. OCI will thus become the first hyperscaler other than Google itself to offer these chips. This is the eighth generation, consisting of two chips: the TPU 8t for training (which, according to Google, delivers nearly three times as much compute per pod as the previous generation) and the TPU 8i for inference. Oracle provides all infrastructure outside of the chip itself, including power, cooling, and networking. The rollout will begin in 2027 and is expected to grow over several years into a gigawatt-scale deployment.
Nvidia remains the foundation
Meanwhile, Nvidia remains the foundation behind OCI’s large-scale AI infrastructure. The Vera Rubin NVL72 racks connect 72 GPUs via NVLink and run on Oracle Acceleron in combination with Spectrum-X Ethernet, Spectrum-6 switches, and BlueField-4 DPUs. Oracle plans to scale these Superclusters up to multiple gigawatts. The Vera CPU instances are intended for the work that agents perform between model calls, such as tool calls, code execution, and retrieval. This is an increasingly important component of agentic workloads as they become more long-running in nature. In addition, OKE scales the CPU and GPU pools independently of one another. A smaller HGX Rubin NVL8 server with eight GPUs is still in the works.
Oracle cites OpenAI as an example customer. OpenAI trained GPT-6 Astra on a Supercluster with more than 100,000 Blackwell GPUs in Abilene, at OpenAI’s Stargate campus.
AMD, Intel, and a host of inference specialists
AMD is supplying the E7 instances based on sixth-generation EPYC Venice CPUs, with up to 256 cores per socket. In addition, Helios racks will be introduced, featuring 72 Instinct MI455X GPUs and up to 31 TB of HBM4. Intel and Oracle also announced a multi-year partnership, including OCI X14 on Diamond Rapids with AMX support for FP8. No dates have been specified for either.
The range of options for inference beyond GPUs and CPUs is also expanding. Cerebras will bring its Wafer-Scale Engine to OCI in 2027, designed for latency-sensitive agents. SambaNova SN50 will follow as well. Notable are OpenAI’s Jalapeño accelerators, developed in collaboration with Broadcom. These appear to primarily provide capacity for OpenAI itself. Additionally, access is already available to Quantinuum’s quantum computer, which was previously brought to OCI.
The heterogeneous setup makes Oracle’s offering more versatile. These choices revolve around more than just price: Nvidia is said to be leading the way in training but may not be the most attractive option for inference.
Acceleron must hold it all together
To ensure all these chips receive computations in a timely manner, Oracle is also expanding Acceleron. The High-Performance Fabric brings RDMA to front-end compute, and NVMe block storage gains more interconnect options with up to three times higher performance. The Oracle Collective Communications Library distributes traffic across the multiplanar backend network and automatically switches paths if links, ports, or NICs fail. Oracle has not yet provided a release date for these networking features either, but as the scale of deployments grows, the benefits will become increasingly apparent.