October 2026: NVIDIA and Oracle Expand Vera Rubin AI Deployment to Multi-Gigawatt Scale
At the Oracle AI World conference on October 27, 2026, Oracle Corporation and NVIDIA announced a major expansion of their enterprise infrastructure partnership. Oracle Cloud Infrastructure (OCI) achieved general availability for NVIDIA Vera Rubin NVL72 rack-scale systems and standalone NVIDIA Vera CPUs, while outlining plans to scale OCI Superclusters to a multi-gigawatt footprint.
This deployment marks one of the earliest public cloud rollouts of NVIDIA’s Rubin platform, moving the next-generation architecture directly into production cloud data centers. More broadly, the agreement reflects a structural shift in how foundational compute is measured across the technology sector. As leading artificial intelligence research laboratories, including OpenAI, run production workloads on OCI, the primary benchmark for cloud infrastructure has evolved from counting individual processors to securing multi-gigawatt electrical capacity, high-density liquid cooling, and low-latency optical interconnect fabrics.
Building on existing NVIDIA GB200 and GB300 NVL72 deployments, the expanded collaboration targets scalable OCI Superclusters capable of linking more than 131,000 GPUs within a single network fabric. Evaluating the implications of this rollout requires examining the physical specifications of the Vera Rubin NVL72 system, the operational rationale behind decoupling standalone Vera CPUs for agentic workloads, the power-efficiency gains achieved over the Blackwell generation, and the capital expenditure models supporting Oracle’s $90 billion to $95 billion fiscal 2027 infrastructure budget.
Hardware Architecture Analysis: Vera Rubin NVL72, Standalone Vera CPUs, and HGX NVL8
The technical foundation of the expanded OCI deployment centers on three distinct hardware configurations engineered to handle distributed frontier model training, multi-step agentic workflows, and enterprise fine-tuning.
Vera Rubin NVL72 Rack Specifications
The flagship system deployed across OCI Superclusters is the liquid-cooled NVIDIA Vera Rubin NVL72. Operating as a unified rack-scale compute domain, each cabinet integrates 72 Rubin GPUs and 36 Vera CPUs interconnected via sixth-generation NVIDIA NVLink (NVLink 6), which delivers 3.0 TB/s of bidirectional bandwidth per GPU and 216 TB/s across the entire rack.
Each Rubin GPU is equipped with 288 GB of next-generation High Bandwidth Memory 4 (HBM4), providing up to 19.2 TB/s of memory bandwidth per chip alongside BlueField-4 Data Processing Units (DPUs) and ConnectX-9 SuperNICs for storage, security, and network offload. The transition to HBM4 directly addresses the memory-wall bottlenecks commonly encountered when serving reasoning-focused models and long-context agent environments. Whereas earlier accelerators often required trillion-parameter models to be sharded across multiple nodes simply to hold large key-value (KV) caches, the expanded memory capacity and bandwidth of the Rubin architecture allow massive context windows to remain in local high-speed memory, improving compute utilization and lowering per-token inference costs.
Decoupling the NVIDIA Vera CPU on OCI
A notable differentiator in OCI’s rollout is its position as the first hyperscale cloud provider to offer the NVIDIA Vera CPU as a standalone, independently scalable compute instance, with an initial deployment of 150,000 Vera CPUs planned across Oracle cloud regions. Built on 88 custom Armv9.2-compatible "Olympus" cores supporting 176 threads via spatial multi-threading and paired with up to 1.5 TB of high-bandwidth LPDDR5X memory, the Vera CPU is purpose-built for the non-GPU stages of modern AI workflows.
In earlier generations such as Grace Hopper and Grace Blackwell, host CPUs were primarily deployed on tightly coupled compute boards alongside GPUs. While highly effective for dense pre-training, that fixed ratio can create resource mismatches for agentic applications. Autonomous AI agents spend substantial time between model inference calls executing auxiliary tools, compiling and testing code in secure sandboxes, querying databases, and orchestrating retrieval-augmented generation (RAG) pipelines.
Through OCI Compute and the OCI Kubernetes Engine (OKE), enterprise customers can now scale Vera CPU instances and Rubin GPU clusters independently. OpenAI, for example, is already utilizing OCI’s standalone Vera CPU instances in production to power its agentic coding sandboxes, while Oracle Acceleron provides hardware-enforced isolation for running AI-generated code. Decoupling the CPU layer allows engineering teams to scale host processing dynamically without leaving expensive GPU silicon sitting idle.
Planned HGX Rubin NVL8 for Enterprise x86 Compatibility
To support modular enterprise deployments and existing x86 application ecosystems, OCI also announced plans to introduce the NVIDIA HGX Rubin NVL8 platform. Configured with eight NVLink-connected Rubin GPUs, dual 64-core Intel Xeon 6 processors, 4 TB of DDR5 system memory, and 6.4 Tbps of high-speed RDMA networking, the NVL8 server provides a smaller eight-GPU building block for organizations conducting domain-specific training, fine-tuning, and inference without requiring full 72-GPU liquid-cooled racks.
Generational Power Efficiency vs. Utility Grid Bottlenecks
The push toward multi-gigawatt AI infrastructure is heavily shaped by physical energy constraints. With utility interconnection queues in major data center hubs often taking several years to deliver new high-voltage capacity, cloud operators cannot expand compute supply solely by building more facilities. Scaling requires extracting significantly more compute throughput from every megawatt of available power.
Maximizing Compute Yield Under Fixed Substation Limits
Delivering up to 30 times higher inference throughput per watt fundamentally alters cloud unit economics. When a data center operator secures a multi-hundred-megawatt utility power allocation, that electrical ceiling remains fixed over the near term. By deploying Vera Rubin racks that generate substantially more tokens per watt, while sharing a common CUDA software foundation with existing GB200 and GB300 infrastructure, OCI can dynamically shift capacity between frontier model training and high-volume production inference, maximizing revenue generation per megawatt of contracted grid power.
Network Topologies and Physical Infrastructure: Connecting 131,000+ GPUs
Scaling an AI supercluster across hundreds of thousands of processors introduces severe networking challenges. When thousands of GPUs synchronize model gradients simultaneously, network bottlenecks can leave processors waiting on data transfers rather than executing computations.
Multipath Reliable Connection (MRC) and Eight-Plane Networking
OCI addresses scale-out networking bottlenecks by combining its Oracle Acceleron architecture with NVIDIA Spectrum-X Ethernet, utilizing 102.4 Tb/s Spectrum-6 switches, ConnectX-9 SuperNICs, and BlueField-4 DPUs. Rather than relying on traditional three-tier network topologies that add hop latency and switch costs, OCI combines multiplanar networking with SRv6 source routing to build flat, two-tier networks capable of supporting more than 131,000 GPUs.
At the transport layer, Oracle Acceleron utilizes its Multipath Reliable Connection (MRC) protocol across an eight-plane network architecture, where every server connects to eight independent network planes simultaneously. Instead of pinning a GPU-to-GPU data transfer to a single physical fiber path, which often causes congestion hotspots when large data streams collide, MRC sprays packets from a single connection across hundreds of available network paths at once. By dynamically balancing traffic across the multiplanar fabric, the system minimizes tail latency and keeps GPU clusters operating at predictable utilization.
The Two-Kilometer Campus Design
Physical distance also constrains distributed superclusters, as optical fiber introduces roughly 5 microseconds of propagation delay per kilometer. If a multi-gigawatt deployment is scattered across distant facilities, speed-of-light latency degrades synchronous training performance. Consequently, OCI Superclusters are engineered inside gigawatt-class campuses optimized for high density within a two-kilometer radius, integrating multi-rack power delivery and closed-loop liquid cooling within a compact physical footprint.
The $40B Per Gigawatt Benchmark and Oracle’s Multi-Vendor Strategy
The multi-gigawatt expansion between NVIDIA and Oracle underscores the massive capital commitments shaping the AI infrastructure sector. While the two companies did not disclose a specific dollar value for the October 2026 agreement, NVIDIA management has previously cited roughly $40 billion per gigawatt as a benchmark for the total data center infrastructure opportunity across compute, networking, and systems.
Funding a $90B–$95B Capital Expenditure Roadmap
During its October 2026 investor updates, Oracle reaffirmed its fiscal year 2027 capital expenditure target of approximately $90 billion to $95 billion as it works toward deploying multi-gigawatt data center capacity. Because funding multiple gigawatts of infrastructure solely through operating cash flow would strain any single corporate balance sheet, Oracle noted that it is pacing its capital deployment through a combination of long-term customer prepayments, equity financing, and customer-supplied hardware arrangements—where major AI clients cover a growing share of chip procurement costs while OCI provides the energized data center shells, liquid cooling, and Acceleron networking fabric.
Alongside the hardware rollout, Oracle also expanded its enterprise software collaboration with OpenAI to cover Oracle AI Workspace and an AI Security Mesh, embedding frontier model capabilities deeper into enterprise workflows.
Silicon Diversification: Google TPUs, Cerebras, and SambaNova
Even as Oracle deepens its flagship partnership with NVIDIA, the company is pursuing a multi-vendor chip strategy to give cloud customers hardware flexibility across diverse workloads. Alongside the Vera Rubin announcement, Oracle unveiled agreements to integrate Google Cloud Tensor Processing Units (TPUs) as well as specialized AI chips from wafer-scale developer Cerebras Systems and SambaNova into its cloud ecosystem. This multi-architecture approach allows enterprises to route specialized inference or domain-specific tasks to alternative accelerators while relying on NVIDIA Vera Rubin superclusters for massive-scale training and agentic systems.
Conclusion
The expansion of NVIDIA Vera Rubin infrastructure to multi-gigawatt scale on Oracle Cloud Infrastructure highlights the industrial transformation of cloud computing. By combining liquid-cooled NVL72 racks, up to 30x higher inference efficiency per watt, and an eight-plane Acceleron network capable of linking over 131,000 GPUs within a two-kilometer radius, OCI and NVIDIA are engineering around the physical limits of power delivery and network latency. Paired with decoupled Vera CPUs for agentic workloads and a diversified chip portfolio that includes Google TPUs, Cerebras, and SambaNova, Oracle’s $90 billion to $95 billion capital roadmap positions gigawatt-scale energy and compute orchestration at the center of enterprise AI competition.
FAQs
What did NVIDIA and Oracle announce in October 2026?
On October 27, 2026, at Oracle AI World, NVIDIA and Oracle announced plans to expand OCI Superclusters to multi-gigawatt scale using NVIDIA Vera Rubin NVL72 rack-scale systems. The companies also announced the immediate general availability of Vera Rubin NVL72 systems and standalone NVIDIA Vera CPUs on OCI, alongside plans to introduce eight-GPU HGX Rubin NVL8 servers.
How does the NVIDIA Vera Rubin platform compare to Blackwell?
According to Oracle and NVIDIA, the Vera Rubin platform delivers up to 30 times higher inference throughput per watt and up to a 4x training performance improvement over the previous Blackwell generation. Each Rubin GPU features 288 GB of HBM4 memory with 19.2 TB/s of bandwidth and connects via sixth-generation NVLink at 3.0 TB/s bidirectional bandwidth per GPU.
Why is Oracle offering standalone NVIDIA Vera CPUs?
Modern AI agents spend significant processing time outside of GPUs—executing code in sandboxes, calling external tools, and retrieving data. By offering 150,000 standalone 88-core NVIDIA Vera CPUs across its regions, OCI allows customers like OpenAI to scale CPU-intensive agent workloads independently from GPU clusters, reducing idle hardware capacity.
How is Oracle funding its multi-gigawatt AI infrastructure buildout?
Oracle maintained its fiscal 2027 capital expenditure target of $90 billion to $95 billion. To manage balance-sheet demands while expanding to multi-gigawatt scale, Oracle utilizes a mix of operating cash flow, customer prepayments, equity issuance, and arrangements where large customers help cover the cost of chip procurement.
Disclaimer
This general market information, possibly from third-party, commercial, or sponsored sources, is provided for reference and educational purposes only and does not constitute financial or investment advice, an offer, solicitation, or guarantee. We disclaim liability for its accuracy, completeness, reliability, and any resulting losses. Investments/trading are risky; past performance doesn’t guarantee future results. Users should research, judge prudently, and take full responsibility. For further details, please consult KuCoin’s Terms of Use and Risk Disclosure.