Nvidia GTC Berlin: Vera Rubin in production, OpenShell, and Nemotron 3.5 Other

At GTC Berlin 2026, Nvidia announced that its Vera Rubin architecture has reached volume production and introduced OpenShell, an open reference architecture for running secure autonomous agents. The company also published open model releases including Nemotron 3.5 Super VL, the Nemotron Retriever collection, and Cosmos 3 world models.

Nvidia highlighted nearly billion in European investments over the past year across 14 startup investments and 12,000 regional partners. CEO Jensen Huang framed modern data centers as AI factories that manufacture intelligence as an industrial output.

The hardware announcements centered on modular deployment speed. Vera Rubin accelerators deliver 20 petaflops each, while NVLink switches connect 72 accelerators in an NVL72 rack to produce 1.44 exaflops of liquid-cooled compute. Node assembly takes under five minutes compared to two hours on Grace Blackwell. The accompanying DSX POD packages seven liquid-cooled racks into a prefabricated shipping container that deploys on site in two weeks instead of four to five months.

For agentic workloads, Nvidia introduced OpenShell alongside Nvidia Sentry, a policy-monitoring layer running on BlueField-4 data processing units that operates in a separate trust domain to isolate misbehaving agents. Mistral AI is partnering with Nvidia to use CUDA-X and Omniverse for industrial simulation, generating synthetic data for engineering foundation models.

What it is

• Vera Rubin production. 20 petaflops per accelerator in NVL72 racks delivering 1.44 exaflops with complete liquid cooling.
• DSX POD AI factories. Containerized data center units with 17 km of Spectrum-X cabling, shipping in early 2027 through Dell, HPE, Gigabyte, Delta, and QCT.
• OpenShell and Nvidia Sentry. Open sandboxed runtime for autonomous agents with hardware-isolated policy enforcement on BlueField-4 DPUs.
• Open weights releases. Nemotron 3.5 Super VL (120B), Nemotron 3.5 Bolt (74B), Nemotron Retriever for search and embeddings, and Cosmos 3 world models.
• Autonomous machines. Agile Robots unveiled the Agile ONE humanoid built on Isaac GR00T and Jetson Thor, and Nvidia demonstrated Drive Hyperion for Level 4 robotaxis.

Sources

• Nvidia DSX POD announcement
• Nvidia OpenShell reference architecture
• Nvidia Sentry agent security
• Nemotron 3.5 on Hugging Face
• Mistral AI and Nvidia engineering collaboration