Oracle AI World: OCI Zettascale10, an AI supercomputer of up to 800,000 GPUs

2025-10-14 · Publisher : Oracle · Domain : Cloud & DevOps

At Oracle AI World, Oracle announces OCI Zettascale10, a cluster linking hundreds of thousands of NVIDIA GPUs across several data centers, up to 16 zettaflops peak. Orders are open, for availability in the second half of 2026.

### Background On 14 October 2025, at Oracle AI World (Las Vegas), Oracle announces Oracle Cloud Infrastructure (OCI) Zettascale10. Oracle presents it as the largest AI supercomputer in the cloud: hundreds of thousands of NVIDIA GPUs linked across several data centers to form multi-gigawatt clusters, with up to 16 zettaflops of peak performance (vendor figure, expressed in low precision). It is the evolution of the first Zettascale cluster introduced in September 2024. ### Key facts - Oracle Acceleron RoCE networking: a technique that uses the switching built into GPU network cards to connect them to several switches on separate network planes, to limit latency and failures. - Clusters are housed in gigawatt-scale data center campuses within a two-kilometer radius, to reduce latency between GPUs. - This architecture is being deployed with OpenAI at the Stargate site in Abilene (Texas), where it was developed first. - First configurations target up to 800,000 NVIDIA GPUs; orders open, availability in the second half of 2026. - OpenAI and NVIDIA are among the contributors quoted in the press release. ### Why it matters The announcement confirms that Oracle is positioning itself in training very large models, alongside AWS, Google Cloud and Microsoft. For customers outside AI labs, it is a reminder that the announced capacities are reserved for a few players and that peak performance figures are theoretical maxima, not to be confused with performance on a real workload. ### Takeaway Oracle moves up the AI infrastructure range: an 800,000-GPU cluster is offered for order.

Tags : Oracle, NVIDIA, Hyperscaler, Datacenter, Conférence