Why the most demanding companies entrust their AI to OCI & NVIDA

Blackwell GPU Architecture Announcements (NVIDIA GTC)

• Training performance is approximately four times faster than the previous Hopper generation.
• Inference performance is 30 times greater than the current generation while consuming 25% less energy.
• Introduction of the GB200 NVL72 rack-scale system, interconnecting 72 chips to function as a single unit.
• The Blackwell architecture and GB200 rack-scale systems are scheduled to arrive on Oracle Cloud Infrastructure (OCI) next year.
• The L40S GPU, arriving by June, is positioned as a versatile "do-it-all" chip for graphics, compute, and AI inference.
• L40S pricing is positioned between the A100 and H100, offering A100-level compute performance with a Transformer engine for LLMs.

Strategic Cloud Model Benefits

• Cloud deployment allows customers to avoid hardware obsolescence risk inherent in on-prem GPU investments.
• Oracle Cloud Infrastructure (OCI) IaaS growth driven by GPU demand has exceeded 50% on a quarterly basis in the US.
• OCI plans to replicate its successful US growth model across European markets.
• A major capacity increase of NVIDIA GPUs is scheduled for European data centers by this summer.

OCI Technology Architecture and Scalability

• OCI offers bare metal access at no extra cost compared to virtualized instances, ensuring direct GPU power.
• The platform utilizes Rocky RDMA over Ethernet version 2 (based on NVIDIA acquisition of Mellanox technology).
• This interconnect architecture provides InfiniBand-level performance at Ethernet pricing, offering a 20-40% cost advantage when scaling.
• The architecture supports clusters up to 32,000 nodes, facilitating massive scale for training and inference.
• Microsoft conducts Bing search inference workloads on OCI infrastructure.

Data Sovereignty and Security

• OCI operates two sovereign data centers in Europe (Frankfurt and Madrid) managed by European teams.
• Security protocols include collaboration with Thales to ensure state-of-the-art encryption.
• The "Sovereign Cloud" model addresses customer concerns regarding data privacy and sovereignty.

Software and AI Infrastructure Stack

• NVIDIA Inference Microservices (NIM) package CUDA libraries, Triton, and other frameworks into a single pre-optimized container.
• NIM reduces deployment complexity from managing multiple containers and specialists to deploying a single container.
• Oracle Database 23c includes built-in vector database capabilities integrated with the standard database engine.
• The combination of OCI, NIM, and Oracle Database provides an end-to-end pipeline from data preparation to RAG-enabled inference.

Future Outlook and Market Positioning

• Enterprise focus is shifting from training (often startup-led) to inference and AI service delivery.
• NVIDIA and Oracle are expanding their partnership to accelerate database processing alongside AI models.
• The joint offering targets full-stack AI deployment for both startups requiring massive scalability and enterprises focusing on secure, sovereign infrastructure.