Oracle has announced its OCI Zettascale10, a cloud-based AI supercomputer it claims offers 16 zettaFLOPS of peak performance. The system, utilizing 800,000 Nvidia GPUs, is designed to support large-scale AI workloads developed by partners including OpenAI.
The company asserts that the system can achieve a peak performance of 16 zettaFLOPS distributed across its 800,000 Nvidia GPUs. This level of output, when calculated on a per-GPU basis, equates to approximately 20 petaflops for each unit. This individual performance metric is comparable to the output of the Grace-Blackwell GB300 Ultra chip, a component used in high-end desktop systems specifically designed for artificial intelligence tasks. The total figure positions the Zettascale10 as a significant entry in large-scale computational infrastructure.
Oracle has identified the platform as the foundational infrastructure for OpenAI’s Stargate cluster, which is located in Abilene, Texas. This facility is being constructed to manage some of the most demanding AI workloads currently emerging from both research initiatives and commercial applications. Peter Hoeschele, vice-president of Infrastructure and Industrial Compute at OpenAI, stated, “The highly scalable custom RoCE design maximizes fabric-wide performance at gigawatt scale while keeping most of the power focused on compute.”
Central to the Zettascale10 system is the Oracle Acceleron RoCE networking architecture, which has been engineered to enhance scalability and reliability for data-heavy AI operations. This design employs network interface cards that function as miniature switches, creating direct links between GPUs across several isolated network planes. This configuration is intended to reduce latency in communication between GPUs. It also provides redundancy, allowing computational jobs to continue processing without interruption even if one of the network paths experiences a failure.
Nvidia’s role in the system was highlighted by Ian Buck, vice-president of Hyperscale at the company. “Featuring Nvidia full-stack AI infrastructure, OCI Zettascale10 provides the compute fabric needed to advance state-of-the-art AI research and help organizations everywhere move from experimentation to industrialised AI,” Buck said. Oracle also claims its network structure can lower costs by simplifying the tiers within the network fabric while delivering consistent performance across all nodes. The system introduces Linear-Pluggable and Receiver Optics technologies, aimed at reducing both energy consumption and cooling requirements without sacrificing bandwidth.
The 16 zettaFLOPS performance claim from Oracle has not been independently verified. Performance metrics for cloud systems can differ based on the methodology used for calculation, and the company’s figure might be based on theoretical peak performance rather than sustained operational rates. Since the system’s advertised total output equals the sum of its 800,000 GPUs operating at their maximum potential, its real-world efficiency will depend significantly on factors like network design and software optimization. Analysts are expected to wait to see if the configuration delivers performance comparable to established AI clusters from other major cloud providers.
The Zettascale10 system is designed to allow customers to train and deploy large AI models across Oracle’s distributed cloud environment, which includes data sovereignty measures. Mahesh Thiagarajan, executive vice-president at Oracle Cloud Infrastructure, commented, “With OCI Zettascale10, we’re fusing OCI’s Oracle Acceleron RoCE network architecture with next-generation Nvidia AI infrastructure to deliver multi-gigawatt AI capacity at unmatched scale.” He added that customers can build and train models using less power and operate with “strong data and AI sovereignty.” The system also offers operational flexibility through independent plane-level maintenance, which permits updates with reduced downtime.
Observers have noted that other major cloud providers are concurrently building their own large-scale GPU clusters and developing advanced cloud storage systems, which could narrow any competitive advantage held by Oracle. The Zettascale10 system is scheduled for a rollout next year. Its capacity to meet growing demand for scalable, efficient, and reliable AI computation will be evaluated following its deployment.
