Oracle unveils Zettascale10 AI supercomputer, claims will be largest in cloud (www.datacenterdynamics.com)

Oracle announced the OCI Zettascale10 AI supercomputer, a multi–data-center fabric built from hundreds of thousands of Nvidia GPUs that the company says will be the largest AI supercomputer in the cloud. Zettascale10 is designed as multi‑gigawatt clusters that Oracle claims can reach up to 16 zettaflops of peak AI performance (measured in low‑precision FP8 or FP16), and it will be offered to customers starting in the second half of 2026, with initial deployments targeting about 800,000 GPUs. Technically, Zettascale10 pairs Nvidia’s AI hardware with Oracle’s Acceleron RoCE custom networking to scale fabric‑wide performance while concentrating power on compute — the same networking and cluster fabric used in the OpenAI “Stargate” Abilene site, according to Oracle and OpenAI. For the AI/ML community this signals greater access to true zettascale-class training and inference capacity on demand, enabling much larger model training, more aggressive parallelism and faster iteration for industrialized AI. It also underscores the growing importance of hyperscale interconnect and power/cooling engineering as core enablers of next‑generation model scaling; Oracle’s previous plan for a 131k‑GPU Blackwell cluster (announced Sept 2024) is being superseded by this far larger cloud offering.