Oracle Cloud Adds NVIDIA GB200 Clusters for AI

Oracle Cloud Infrastructure (OCI) is expanding its AI compute arsenal with dedicated NVIDIA GB200 NVL72 clusters designed specifically for large-scale AI training workloads. The move positions Oracle as a serious contender against AWS, Microsoft Azure, and Google Cloud in the increasingly competitive race to offer the most powerful GPU infrastructure for enterprise AI development.

The new offering gives customers access to full rack-scale GB200 systems — NVIDIA's most powerful AI training platform to date — through OCI's global data center network, removing the need for enterprises to build and manage their own supercomputing infrastructure.

Key Takeaways at a Glance

• NVIDIA GB200 NVL72 clusters are now available as dedicated infrastructure on Oracle Cloud
• Each GB200 NVL72 rack contains 72 Blackwell GPUs and 36 Grace CPUs connected via NVLink
• The clusters deliver up to 30x inference performance improvements over the previous Hopper generation
• Oracle targets enterprise customers training foundation models and running massive AI workloads
• Pricing follows OCI's consumption-based model with dedicated bare-metal access
• The deployment complements Oracle's existing NVIDIA H100 and H200 GPU offerings

NVIDIA GB200 NVL72 Brings Unprecedented AI Training Power

NVIDIA's Blackwell architecture represents a generational leap in AI compute capability. The GB200 NVL72 configuration — sometimes referred to as an 'NVL72 rack' — packs 72 Blackwell GPUs into a single liquid-cooled rack, interconnected through NVIDIA's proprietary NVLink fabric.

This architecture enables all 72 GPUs to operate as a single, unified accelerator with a combined 13.5 terabytes of high-bandwidth memory. Compared to the previous-generation H100 clusters, the GB200 NVL72 delivers dramatically higher throughput for training trillion-parameter models.

The system supports FP4 precision for inference workloads, a feature absent in earlier Hopper-based systems. This allows enterprises to run both training and inference on the same infrastructure, maximizing utilization and reducing total cost of ownership.

Oracle Challenges AWS and Azure in the GPU Cloud Wars

The cloud GPU market has become one of the most fiercely contested segments in enterprise technology. Amazon Web Services launched its own NVIDIA GB200 instances earlier this year, while Microsoft Azure has been aggressively expanding its GPU fleet to support OpenAI's training demands.

Oracle's strategy differs from its hyperscaler rivals in several important ways:

• Bare-metal access: OCI provides direct hardware access without virtualization overhead, delivering near-on-premises performance
• Network architecture: Oracle's non-oversubscribed RDMA network fabric is purpose-built for distributed AI training
• Pricing transparency: OCI's GPU pricing has historically been 30-50% lower than comparable AWS and Azure offerings
• Cluster scale: Oracle has committed to building some of the largest contiguous GPU clusters in the industry, with configurations exceeding 65,000 GPUs

Larry Ellison, Oracle's co-founder and CTO, has repeatedly emphasized that OCI's cloud network architecture gives it a structural advantage for AI workloads. Unlike traditional cloud networks designed for web applications, OCI's backend fabric was engineered from the ground up for low-latency, high-bandwidth communication between GPU nodes — a critical requirement for distributed training.

Technical Specifications That Matter for AI Teams

For engineering teams evaluating GPU cloud options, the technical details of Oracle's GB200 deployment are significant. Each NVL72 rack delivers approximately 1.4 exaflops of AI training performance in FP8 precision.

The liquid-cooling infrastructure required for Blackwell GPUs represents a major capital investment that most enterprises cannot justify building independently. By offering these systems through OCI, Oracle absorbs the complexity of deploying and maintaining liquid-cooled data centers at scale.

Key technical specifications include:

• 72 Blackwell GPUs per NVL72 rack with 13.5 TB combined HBM3e memory
• NVLink Switch System providing 130 TB/s of GPU-to-GPU bandwidth
• Liquid cooling infrastructure managed entirely by Oracle
• OCI Supercluster networking for scaling across multiple racks
• NVIDIA AI Enterprise software stack pre-configured for common training frameworks
• Support for PyTorch, JAX, and NVIDIA NeMo out of the box

The bandwidth between GPUs is particularly noteworthy. NVLink's 130 TB/s interconnect bandwidth dwarfs what traditional InfiniBand networks can deliver, making it possible to train models with tens of trillions of parameters without encountering communication bottlenecks.

Enterprise AI Training Enters a New Phase

The availability of GB200 clusters on OCI signals a broader shift in how enterprises approach AI development. Training large foundation models was once the exclusive domain of a handful of well-funded AI labs — OpenAI, Google DeepMind, Anthropic, and Meta.

Today, an increasing number of enterprises in finance, healthcare, manufacturing, and defense are investing in custom model training. These organizations need access to massive GPU clusters but lack the expertise or desire to build data center infrastructure from scratch.

Oracle's managed GB200 offering addresses this gap directly. Customers get dedicated hardware — not shared or virtualized — with the full performance characteristics of an on-premises supercomputer, but without the 18-24 month lead time typically required to procure and deploy NVIDIA's latest GPUs independently.

This trend toward 'cloud supercomputing' is accelerating as model architectures grow more complex. The latest generation of mixture-of-experts models, multimodal systems, and reasoning-focused architectures all demand significantly more compute than their predecessors from just 12 months ago.

What This Means for Developers and Businesses

For AI teams considering OCI's GB200 clusters, the practical implications are substantial. Organizations training models in the 100-billion to 1-trillion parameter range will see the most immediate benefits.

The dedicated nature of OCI's offering eliminates the 'noisy neighbor' problem that plagues shared GPU clouds. When training runs can cost $5-50 million in compute alone, even small performance degradations from shared infrastructure translate to significant financial losses.

Startups and mid-size companies also stand to benefit from Oracle's pricing model. OCI has consistently undercut AWS and Azure on GPU instance pricing, and early indications suggest the GB200 offerings will follow the same pattern. For a startup training a competitive large language model, choosing OCI over AWS could save hundreds of thousands of dollars on a single training run.

Additionally, Oracle's deep integration with its enterprise database and application stack creates opportunities for companies already in the Oracle ecosystem. Training AI models on proprietary enterprise data stored in Oracle Autonomous Database or Oracle Cloud Applications becomes significantly more streamlined when the compute and data reside on the same cloud platform.

Looking Ahead: The Race for AI Infrastructure Dominance

Oracle's GB200 deployment is part of a larger $80+ billion capital expenditure plan the company has outlined for cloud infrastructure expansion. The company is building new data centers across the United States, Europe, and Asia specifically designed to house liquid-cooled GPU clusters.

The competitive landscape will intensify throughout 2025 and into 2026. NVIDIA's next-generation Rubin architecture, expected in late 2026, promises another major performance leap. Cloud providers that establish strong GPU infrastructure partnerships now will be best positioned to offer Rubin-based systems when they become available.

For Oracle, the stakes extend beyond just GPU hosting. Successfully capturing a meaningful share of the AI training market validates the company's broader cloud strategy and proves that OCI can compete with hyperscalers on the most demanding workloads in computing.

The AI infrastructure market is projected to exceed $200 billion annually by 2028, according to multiple industry analysts. Oracle's aggressive investment in NVIDIA's latest technology suggests the company is betting heavily that enterprise AI training will be a defining workload of the next decade — and that OCI will be where a significant portion of that training happens.