NVIDIA AI Updates: May 6, 2026

1. NVIDIA and ServiceNow Launch Project Arc and OpenShell Runtime

NVIDIA. Announced at ServiceNow Knowledge 2026 by Jensen Huang and ServiceNow CEO Bill McDermott, the expanded partnership pairs ServiceNow’s Action Fabric and AI Control Tower with NVIDIA’s Blackwell platform, Nemotron open models, and a new secure runtime called NVIDIA OpenShell, which runs agents in sandboxed, policy-governed environments. ServiceNow introduced Project Arc, a self-evolving autonomous desktop agent for knowledge workers, developers, and IT teams that connects natively to the ServiceNow AI Platform for governance and auditability while accessing local files, terminals, and applications to complete multistep tasks. The companies also released NOWAI-Bench, an open benchmarking suite for evaluating enterprise AI agents. Source

2. NVIDIA Lays Out an In-Vehicle AI Agent Stack from Cloud to Car

NVIDIA. A developer-blog deep dive describes how to deploy 7B+ parameter agentic models inside vehicles with sub-500ms latency and 30+ tokens/second decode, using a pipeline of Nemotron speech models for ASR, the NeMo Agent Toolkit for orchestration, TensorRT Edge-LLM for on-device inference, and Magpie for TTS. The post outlines three hardware paths — an add-on DRIVE AGX AI Box, the Blackwell-based DRIVE AGX Thor consolidating cabin and AV workloads, or a central computer pairing DRIVE AGX with MediaTek’s Dimensity AX C-X1 SoC — and cites ABI Research projecting agentic-AI vehicle shipments to grow from roughly 5 million in 2025 to 70 million by 2035. Source

3. NVIDIA Pitches Vera Rubin as the Substrate for Agentic Workloads

NVIDIA. A second developer post argues that agentic systems consume up to 15x more tokens than chat workloads — citing Anthropic data and a Claude Code session whose context grew from 15,000 to 156,000 tokens across 283 requests in 33 minutes — and frames the Vera Rubin platform as an “extreme co-design” answer. The stack combines Vera Rubin NVL72 compute, the Vera CPU for tool execution and KV-cache management, NVLink 6, ConnectX-9, BlueField-4, and Spectrum-X, with NVIDIA targeting 400+ tokens per second per user on trillion-parameter MoE models with 400k context. Source