Oracle brings Nvidia GB200 NVL72 systems to OCI, targeting large-scale LLM training and inference

Oracle said Tuesday it has made Nvidia GB200 NVL72 systems generally available on Oracle Cloud Infrastructure in Phoenix, Ashburn, and Frankfurt, offering customers reserved capacity for training and serving large language models. The rollout includes support for Nvidia AI Enterprise and NIM microservices on OCI, with cluster provisioning through OCI Supercluster. Oracle framed the launch as a way to meet demand for high-bandwidth GPU fabrics and European data locality.

Oracle is leaning on Nvidia’s Blackwell performance claims to attract enterprise AI workloads. Nvidia stated at GTC 2024 that Blackwell brings significantly higher inference throughput for trillion parameter models and lower total cost of ownership compared with H100 systems, with NVLink 5 and NVSwitch 5 enabling rack-scale coherence across 72 GPUs. OCI Supercluster exposes clustered RDMA networking and parallel file systems that enterprises already use for H100 training runs.

Pricing will follow OCI’s existing GPU model of on demand with volume discounts and multiyear reservations, Oracle said, without disclosing hourly rates. Nvidia AI Enterprise is available via the OCI Marketplace as an add-on license, and NIM containers are supported for accelerated deployment of common model endpoints. Oracle said customers can burst or pin capacity, with quotas allocable per tenancy to manage cost and governance.

The company positioned Frankfurt availability as relevant for European customers preparing for AI Act high risk system rules that begin phasing in from 2026, citing customer requests for regional processing and customer managed keys. OCI’s Supercluster previously advertised 1.6 Tbps RDMA per node, and Oracle says Blackwell clusters inherit the same fabric and storage topology used by its H100 estates. Oracle also referenced its expanded 2024 infrastructure partnership with Microsoft for Azure interconnect, which the companies said would support AI capacity for shared customers.

The launch puts Oracle into direct competition with AWS, Microsoft, and Google as the industry begins deploying Blackwell at scale. Synergy Research Group estimated Oracle’s share of global cloud infrastructure services at about 3 percent in 2024, far behind AWS and Microsoft, so early Blackwell capacity could differentiate OCI for training reservations and managed inference tied to Nvidia’s NIM stack. For enterprises, the availability broadens options for GPU supply, cluster scale, and regional controls as AI programs move from pilots to multi-region production.