Oracle and NVIDIA unleash a 16-zettaflop AI monster cluster

Running cutting-edge AI models doesn't mean much if your enterprise data is still trapped inside legacy databases and messy spreadsheets.

At Oracle AI World, NVIDIA and Oracle announced a sweeping expansion of their partnership, embedding accelerated computing and enterprise AI software directly into Oracle Cloud Infrastructure.

Alongside a massive leap in peak computing capacity, the collaboration targets the real-world bottlenecks in enterprise data ingestion, vector indexing, and model deployment.

Multi-gigawatt clusters meet GPU-accelerated databases

The headline hardware announcement is the OCI Zettascale10 computing cluster. Built for massive training and inference, the cluster is powered by NVIDIA's next-generation GPUs and delivers up to 16 zettaflops of peak AI performance. Powered by NVIDIA Spectrum-X Ethernet, the architecture is designed to interconnect millions of GPUs with multi-gigawatt power capacity.

Yet for everyday enterprise workloads, the software updates inside the database stack might be the real game changer.

Oracle Database 26ai now integrates with NVIDIA NeMo Retriever APIs, enabling developers to run embedding models and full retrieval-augmented generation pipelines using NVIDIA NIM microservices. NVIDIA covers the entire RAG pipeline from multimodal data ingestion to reranking and LLM response generation.

Oracle is also bringing GPU acceleration to the Oracle Private AI Services Container. Vector index build times frequently turn into major bottlenecks as dataset volumes explode.

By leveraging the open-source NVIDIA cuVS library on GPUs, users can offload heavy approximate nearest neighbor search index creation from CPUs to GPUs, quickly formatting the results for querying inside Oracle AI Database 26ai.

Zero-code Spark acceleration and native enterprise suites

The integrations extend further into data analytics. The new Oracle AI Data Platform features the NVIDIA RAPIDS Accelerator for Apache Spark plug-in. By coupling cuDF with Spark's distributed architecture, organizations can accelerate ETL, data transformation, and machine learning pipelines on GPUs without altering existing code.

Specialized workloads also get tailored models. Oracle Media and Entertainment is tapping NVIDIA NeMo Curator and a Nemotron vision language model to automate video preprocessing, segmentation, and captioning. Meanwhile, Oracle Fusion Document Intelligence adopted NeMo Retriever Parse to handle complex, structured enterprise documents.

To tie it all together, NVIDIA AI Enterprise is now natively embedded in the OCI Console rather than requiring a separate order from the cloud marketplace. Customers can turn it on during GPU instance provisioning across public, sovereign, and dedicated regions.

Why it matters

For enterprise teams managing petabytes of legacy data on Oracle, this deal removes massive engineering friction. Building RAG applications previously meant extracting data, standing up external vector databases, and stitching together fragile custom pipelines. Doing index builds and embedding generation right where the data lives saves months of plumbing.

The zero-code Spark acceleration is equally pragmatic, letting existing analytics workloads tap GPU efficiency without forcing teams to rewrite business logic from scratch.

From an industry perspective, Oracle is defending its enterprise stronghold while NVIDIA secures another deep layer of software lock-in. Instead of fighting cloud hyperscalers on raw commodity compute alone, Oracle is positioning itself as the premium highway for high-throughput enterprise AI.

We think NVIDIA's strategy here is exceptionally sharp. By solving tedious, unglamorous bottlenecks like document parsing and vector indexing, they make their software ecosystem indispensable. The obvious catch, however, is vendor lock-in. Once an enterprise bakes NeMo Retriever, cuVS, and NIM microservices into its core database workflows, moving to an alternative cloud or chip provider looks increasingly difficult.