Summary
On October 27, 2026, Oracle and NVIDIA announced that Vera Rubin NVL72 and Vera CPUs are now generally available on Oracle Cloud Infrastructure, with OCI Superclusters expanding to multi-gigawatt scale; OpenAI trains and serves GPT-6 Astra on an OCI Supercluster powered by 100,000-plus NVIDIA Blackwell GPUs.
Event Overview
- Date: October 27, 2026 (local time), announced jointly in Las Vegas.
- Parties: Oracle and NVIDIA, with statements from OpenAI engineering leadership and its compute CTO.
- Event: Per the joint press release, the NVIDIA Vera Rubin NVL72 and NVIDIA Vera CPU are now generally available on OCI; OCI Superclusters will scale with Vera Rubin NVL72 to multi-gigawatt capacity.
Key Information
| Item | Detail |
|---|---|
| Announcement date | October 27, 2026, Las Vegas |
| GA products | NVIDIA Vera Rubin NVL72 and NVIDIA Vera CPU (generally available on OCI) |
| Deployment scale | OCI Superclusters expanding to multi-gigawatt scale with Vera Rubin NVL72 |
| NVL72 rack | 72 GPUs connected via NVLink per rack, paired with Spectrum-X Ethernet |
| Networking | Spectrum-6 switches plus BlueField-4 DPUs |
| Planned system | HGX Rubin NVL8: one x86 server with 8 Rubin GPUs |
| NVL8 specs | 2x 64-core Intel Xeon 6 CPUs, 4TB DDR5 memory, 6.4 Tbps RDMA networking |
| OpenAI today | GPT-6 Astra trained and served on 100,000-plus Blackwell GPUs in Abilene, Texas |
Background and Impact
- Per the two companies: Each NVL72 rack connects 72 GPUs via NVLink with Spectrum-X Ethernet (Spectrum-6 switches plus BlueField-4 DPUs); OCI also plans to add the HGX Rubin NVL8 system — a single x86 server with 8 Rubin GPUs, 2x 64-core Intel Xeon 6 CPUs, 4TB DDR5 memory and 6.4 Tbps RDMA networking.
- OpenAI today: The release quotes OpenAI engineering lead Sreekanth Pothanis, who said OpenAI trains and serves GPT-6 Astra on an OCI Supercluster in Abilene, Texas, powered by more than 100,000 NVIDIA Blackwell GPUs.
- Role of Vera CPU: OpenAI compute CTO Uday Ruddarraju said Vera CPUs are being brought into OCI production to support agentic sandbox workloads.