NVIDIA AI Updates: October 22, 2026

1. NVIDIA Publishes the OpenShell Reference Architecture for Safe AI Agents, With BlueField-4 Sentry Monitoring

NVIDIA. At GTC Berlin, NVIDIA introduced the OpenShell Reference Architecture, built on OpenShell, its open source runtime that treats agent workloads as untrusted. OpenShell is split into a gateway (control plane for sandbox lifecycles and policies), a supervisor that checks outbound requests against policy and mediates credential access from outside the sandbox, the sandbox itself, and a policy prover that uses formal verification to confirm a policy stays under an organization-defined permission ceiling. An optional hardware layer runs NVIDIA Sentry on BlueField-4 DPUs, monitoring requests, network and file activity in an isolated trust domain and quarantining an agent within milliseconds. CoreWeave reported more than 3x faster sandbox startup on Vera CPUs than on x86 and over 11,000 concurrent agent environments per rack. Source

2. Agent Toolkit Adds Agent Knowledge Foundation and the Open Kumo Models for Structured Data

NVIDIA. The open source NVIDIA Agent Toolkit now includes Agent Knowledge Foundation, which bundles CUDA-X libraries (cuDF, cuGraph, cuML, cuOpt, cuVS), a new Auto-Ontology library that maps raw company data to business concepts like customers and orders, Nemotron Retriever, and the new Kumo Structured models (Tabular, Relational, Time Series, Forecast), plus a new NemoClaw blueprint for agent optimization. Kumo Relational predicts directly on relational databases via in-context learning over up to 10,000 historical examples, with tasks written in Predictive Query Language. NVIDIA reports it beats gradient-boosted trees on expert-engineered features by nearly 4 AUROC points on RelBench V1 classification without task-specific training; code is Apache 2.0 and weights use the OpenMDW 1.1 license. Source Source

3. Google Cloud Brings Vera Rubin NVL72 to A5X Instances, With GA Next Month

NVIDIA. Google Cloud and NVIDIA announced A5X instances combining Vera Rubin NVL72, ConnectX-9 SuperNICs, Google’s Virgo network, and fourth-generation liquid cooling, aimed at agentic inference and large-scale reinforcement learning. Despite the headline, the post states general availability is coming next month, with expansion to Europe planned; Google Cloud plans to scale Vera Rubin infrastructure to as many as 960,000 Rubin GPUs across multiple sites. David Silver’s Ineffable Intelligence was the first A5X customer. Google Cloud also made Confidential Computing generally available on G4 VMs with RTX PRO 6000 Blackwell Server Edition GPUs. Source

4. Confidential Bare Metal Lets Model Builders Run Their Own Stack on Third-Party GPUs

NVIDIA. NVIDIA Confidential Computing gains a third deployment option, Confidential Bare Metal, alongside Confidential Containers for Kubernetes and Confidential VMs. New platform controls protect data in transit, measure platform integrity, and enforce hardware-level access restrictions, so frontier model builders can operate their own host and platform software on bare-metal instances at clouds, neoclouds, and enterprises while keeping models and data protected. The capability is expanding across Grace Blackwell and Rubin systems, and Cohere said it plans to integrate it into Cohere Model Vault. Source

5. cuDSS Sparse Direct Solver Reaches General Availability

NVIDIA. cuDSS, the CUDA-X library for solving large sparse linear systems on GPUs, is now generally available and scales from single workstations to clusters. Early integrators report results including up to 10x faster sparse solves in Applied Materials’ Ginestra versus CPU-only, 8-10x speedups in Dassault Systèmes’ Abaqus and CST electromagnetic solvers, and a 4x speedup for ZEISS SMT in a COMSOL EUV lithography plasma simulation. Cadence, Siemens Simcenter STAR-CCM+, and Synopsys Sentaurus also use it. Source

6. BioNeMo RFD4-Proteina Open Protein Design Model Released With the Baker Lab

NVIDIA. Among a slate of open model updates (Cosmos 3.5, Alpamayo 2 Edge, Kumo, new Nemotron Retriever models), NVIDIA released BioNeMo RFD4-Proteina, an open protein design model built with the University of Washington’s Baker Lab. It generates protein sequences and 3D structures for binders, assemblies, and small-molecule binding, and NVIDIA reports leading performance across more than 10 design tasks and roughly 13x faster generation than RFdiffusion3 on 200-residue proteins. NVIDIA also introduced Cosmos Data Factory blueprints and agent skills for curating and generating physical AI training data. Source

7. DSX POD Cuts On-Site AI Factory Deployment to 14 Days

NVIDIA. In his GTC Berlin keynote, Jensen Huang announced NVIDIA DSX POD, a reference design for complete Vera Rubin AI factories built by partners inside modular containers and validated before delivery. NVIDIA says on-site deployment drops from six months to 14 days once site utilities are ready. Huang also said NVIDIA has committed nearly $10 billion to the European technology ecosystem. Source

8. DGX Spark 64GB Ships October 23, DGX Station for Windows Due in December

NVIDIA. NVIDIA’s local AI updates include a 64GB DGX Spark available October 23, NVIDIA Sync support for up to eight DGX Sparks plus a Model Launcher this month, and a Windows version of DGX Station arriving in December. NVIDIA Cloud Functions is entering early access on DGX Station, a new local coding model called Recursive Flint is available for NVIDIA hardware, and new llama.cpp and vLLM accelerations ship alongside checkpoints optimized for local devices. Source