NVIDIA releases Nemotron 3 Ultra with 17 enterprise suppliers embedding the OpenShell sandbox

NVIDIA made the Nemotron 3 Ultra available on 4 June through Hugging Face, ModelScope, OpenRouter, and build.nvidia.com as NIM microservices, concluding the announcement made by Jensen Huang at GTC Taipei during Computex. The model is a mixture of experts comprising 550 billion parameters, which the company claims runs up to 5 times faster and about 30% cheaper than the leading open models of the previous generation. Alongside, the NVIDIA Agent Toolkit was launched with 17 enterprise suppliers embedding the OpenShell runtime for long-duration agents.

The list of adopters shifts the discussion regarding agent stack selection. Included are Adobe, Salesforce, SAP, ServiceNow, Siemens, Atlassian, Cisco, CrowdStrike, Box, Cohesity, IQVIA, Palantir, Red Hat, Cadence, Synopsys, Dassault Systèmes, and Amdocs. SAP, ServiceNow, and Salesforce together account for roughly 70% of corporate SaaS spending in back-office markets such as Germany, the United Kingdom, and Brazil, according to Gartner data for 2025, and each brings its own production agent into NVIDIA’s sandbox.

SAP integrates Joule into OpenShell

SAP has integrated OpenShell with the Joule Studio runtime on the Business Technology Platform, positioning each agent in an isolated environment with configurable policies. This directly addresses an objection that German and Japanese CIOs voiced to Christian Klein, CEO of SAP, during Sapphire in May: production agents need guardrails that meet audit requirements similar to SoD in SAP S/4HANA. The Joule Studio comes in the wake of the 4 June release with NeMo embedded for customisation.

The ServiceNow Project Arc, announced at Knowledge 2026 and running in production since 4 June, is an autonomous desktop agent that executes workflows from the ServiceNow AI Platform connected to CMDB and Action Fabric, also within OpenShell. CrowdStrike and Palantir have adopted Nemotron 3 Ultra as the engine for long-duration agents; CrowdStrike uses it for chained SOC investigations, while Palantir employs it for continuous analysis in Foundry. Perplexity, on the consumer side, uses the model for search and for the Perplexity Computer.

What distinguishes it from Bedrock and AI Foundry

The architecture proposed stacks three elements that other stacks treat separately: an open model optimised for agents, a sandboxed runtime with security policies, and plugable CUDA-X skills in any agent, including cuDF, cuOpt, NeMo, PhysicsNeMo, and CUDA-Q. These skills are accessible through the Claude Code plugin marketplace and the Hermes Skills Hub. For the CIO evaluating AWS's Bedrock, Microsoft's AI Foundry, and Google's Vertex, the question becomes tangible: will they operate agents on the hyperscaler runtime or on a portable runtime across clouds.

The benchmark published by NVIDIA places the Nemotron 3 Ultra ahead of American open models like Llama 4 and GPT-OSS, but behind Moonshot's Kimi K2 and Alibaba's Qwen 3 in reasoning in Chinese and advanced mathematics. This gap is the real pressure point for the White House: Joe Tsai and Lin Junjie have a competitive open model running on the same NVIDIA hardware, under a permissive licence. For enterprise customers in Japan and Singapore, this means vendor choice; for customers in Brazil, it still means stability of the roadmap linked to NVIDIA more than to the laboratory.

The perspective beyond the United States

The pipeline that matters for Germany is straightforward. SAP running agents in OpenShell within the BTP is the first enterprise offering with a sandboxed runtime by default that theoretically satisfies DORA requirements in the European Union and BaFin for German banks. Deutsche Bank, Allianz, and Bayer run SAP at scale and are likely to be among the first pilot tests of the agentic Joule Studio. For India, the importance is different: TCS, Infosys, and Wipro deliver SAP and ServiceNow to half of the Fortune 500 and now need to train tens of thousands of consultants in agents running within the sandbox.

In Brazil, the most immediate effect is felt through Itaú and Bradesco, which maintain significant deployments of ServiceNow for ITSM and are on the public pilot list for Salesforce Agentforce. The Nemotron 3 Ultra being open and runnable in local data centres reduces regulatory dependency under LGPD for tasks involving customer data. The point worth noting is the time between upstream release and availability in NVIDIA's regional NIMs: the company has not published a timeline for São Paulo in the announcement.