Oracle unveils new hyperscale AI compute cluster

Oracle has unveiled a new flagship for hyperscale AI compute: the Oracle Cloud Infrastructure (OCI) Zettascale10 cluster, underpinned by its Acceleron RoCE networking architecture.

The company described Zettascale10 as the largest AI supercomputer in the cloud, designed to deliver multi‑gigawatt capacity and up to 16 zettaFLOPS of peak performance by interconnecting hundreds of thousands of NVIDIA GPUs across multiple sites.

Revealed at Oracle AI World in Las Vegas, Zettascale10 is positioned as the next generation of OCI’s large‑scale AI clusters.

Oracle said the system links vast numbers of NVIDIA accelerators to create unified, production‑grade training and inference environments for frontier models. The architecture has also been tied directly to OpenAI’s Stargate programme, with the first supercluster based in Abilene, Texas.

Re-engineering the network fabric

At the heart of Zettascale10, Oracle’s Acceleron RoCE networking has introduced a multiplanar, Ethernet‑based design that uses switching capabilities in modern GPU NICs to connect to multiple isolated planes.

Oracle has said this approach has improved scale, reliability and power efficiency versus traditional three‑tier networks, while delivering more uniform GPU‑to‑GPU latency for large training jobs.

Additionally, Zettascale10 clusters will be housed on gigawatt‑class campuses, with data centres densely sited within a roughly two‑kilometre radius to reduce latency and network tiers. Oracle added this as a way to accelerate deployments and cut costs while preserving predictable performance at extreme scale.

Zettascale10 availability

Oracle has reportedly begun taking orders and has said Zettascale10 will be available in the second half of next calendar year, with configurations scaling to as many as 800,000 NVIDIA GPUs.

The Abilene site—developed with OpenAI as part of the Stargate initiative—forms the flagship Zettascale deployment, with OpenAI and Oracle having announced plans for several additional U.S. sites and 4.5 GW of added capacity.

Portions of the Abilene campus have already come online, and the wider Stargate build‑out is projected to reach multi‑gigawatt scale.