Oracle Corporation has revealed its latest cloud infrastructure innovation: the OCI Zettascale10 supercomputer, described as the most powerful AI training compute fabric offered via a public cloud. According to the company announcement, the system links “hundreds of thousands” of NVIDIA Corporation GPUs across multiple data centres and is capable of reaching up to 16 zettaFLOPS of peak performance.
The architecture targets large-scale generative AI and high-performance computing workloads by combining GPUs, a custom network fabric called “Oracle Acceleron RoCE”, and hyper-dense data-centre campuses built within a two-kilometre radius of each other to minimise latency. One flagship deployment is in partnership with OpenAI under the “Stargate” supercluster in Abilene, Texas, where the firm says it has already begun implementing the Zettascale10 infrastructure.
Oracle expects initial customer deployments to reach up to around 800,000 NVIDIA GPUs, with multi-gigawatt power usage, and will extend through 2026. The firm emphasises its aim to offer customers the ability to “build, train and deploy their largest AI models into production using less power per unit of performance,” according to an executive quote.
In its previous announcements, Oracle discussed its earlier Zettascale cluster which supported up to 131,072 NVIDIA Blackwell GPUs and delivered roughly 2.4 zettaFLOPS. The jump to Zettascale10 therefore represents a significant leap in compute scale and ambition.
Industry analysts view Oracle’s move as a direct response to mounting demand for generative-AI infrastructure, where training ever-larger models is pushing cloud providers to deploy hyperscale GPU clusters. By offering “zettascale” capability in the cloud, Oracle aims to position itself as a major competitor to other hyperscalers and supercomputing providers. The low latency networking fabric — key in synchronising thousands of GPUs — is a critical differentiator in Oracle’s pitch.
Yet the announcement comes with caveats. While Oracle publicises the 16 zettaFLOPS figure, this is labelled as "peak performance" and will depend on full-scale deployment. Actual customer availability timelines show the system is scheduled for “second half 2026” deployment for large-scale public clusters. The effective performance for real-world AI model workloads may vary and depends on how the infrastructure is used — for example, model size, I/O demands, and inter-GPU communication overhead.
The collaboration behind Zettascale10 is notable. NVIDIA provides the full-stack AI infrastructure, while Oracle supplies the cloud services and network architecture. As NVIDIA’s vice-president of Hyperscale computing noted, “OCI Zettascale10 provides the compute fabric needed to advance state-of-the-art AI research and help organisations everywhere move from experimentation to industrialised AI.” Meanwhile, Oracle’s EVP for cloud infrastructure stressed the importance of combining the Acceleron network with next-generation NVIDIA hardware to deliver “multi-gigawatt AI capacity at unmatched scale.”
Another contributing factor is the growing interest in sovereignty and distributed-cloud models. Oracle emphasises that customers will have “the freedom to operate across Oracle’s distributed cloud with strong data and AI sovereignty controls,” which may appeal to enterprise and governmental users with data-residency or compliance constraints. As the AI ecosystem expands, such capabilities may become more important.
However, some uncertainties remain. The scale of the infrastructure demands massive power and cooling support, with design and operational complexity increasing with size. While the promotional material references gigawatt-scale campuses and clusters sited within two-kilometre radius for latency benefits, managing such facilities at scale under cloud economics is a challenge. Furthermore, the competitive landscape remains intense: other cloud providers and chip-makers are also pushing hard in AI infrastructure, so the window for differentiation may narrow.
Enterprises will also need to evaluate whether they can absorb and operationalise such immense compute capacity in a cost-effective way — simply offering the capability does not guarantee favourable ROI. The time-to-deployment, integration with AI software stacks, and real world utilisation will play major roles.
Topics
Technology