At ServiceNow Knowledge 2026, the headline was not another feature panel. NVIDIA and ServiceNow used the stage to describe a fuller infrastructure layer for enterprise agents: a desktop agent on employee machines, Nemotron 3 behind it, and a governance wrapper around the work it performs. The stack is called Project Arc.
What Project Arc is meant to be
In practical terms, Project Arc is a desktop agent designed to live on a Mac or Windows PC and carry out multi-step work over time.
It is not a Copilot-style assistant waiting for a prompt. The pitch is closer to an independent work process that can read local files, use the terminal, operate installed software and adapt as it works. The first audience is developers, IT engineers and system administrators who already move between many tools every day.
NVIDIA supplies the model layer through Nemotron 3 Super, which it says ranked first among open models on the EnterpriseOps-Gym benchmark. ServiceNow supplies the workflow layer through Action Fabric, tying desktop activity back to tickets, processes and approvals inside ServiceNow.
Project Arc represents the next step...delivering the governance and security that enterprise AI requires.
That line from ServiceNow AI platform general manager Jon Sigler points to the actual sales argument. The tools that let agents act already exist. The reason many enterprises hesitate is that they still lack audit trails, controls and replay mechanisms. Project Arc is not being sold mainly as a smarter agent. It is being sold as a controllable one.
OpenShell may matter even more
The second important release is OpenShell, an open source secure runtime from NVIDIA for autonomous agents.
The problem is concrete. Enterprises need to define which tools an agent may call, keep file operations inside a sandbox, and audit each decision the agent makes.
- Which tools can the agent invoke?
- How is file access isolated?
- How can every step be reviewed after the fact?
OpenShell's answer is to give each agent an isolated environment, govern capabilities through policy files, log actions and let administrators inspect, stop or roll back activity.
That may outlast Project Arc itself. OpenShell is open source, model-neutral and not limited to ServiceNow. Any enterprise building its own agent platform could use it as sandbox infrastructure.
The strategy feels familiar from NVIDIA: spread the tools, then make the hardware the natural place to run them. CUDA followed that pattern. OpenShell could become part of the operating-system layer for enterprise agents.
The numbers NVIDIA wants buyers to remember
Several performance claims framed the announcement.
- Blackwell versus Hopper: 50 times more tokens per watt.
- Cost per million tokens: down by 35 times.
- Nemotron 3 Super: first among open models on EnterpriseOps-Gym.
The 35x cost figure matters because the biggest cost in enterprise LLM deployment has often been inference, not the model license. If a 500-person team lets agents run all day, compute can quickly become the limiting budget item. NVIDIA is arguing that Blackwell pushes the marginal cost low enough for all-employee agent deployment to become plausible.
That is not just about defending GPU margins. It expands the market to a scale that needs more GPUs.
Governance is the enterprise opening
In the weeks before the launch, The Hacker News warned that agents were already inside corporate networks while many companies still could not produce a basic inventory. Gartner has forecast that 40% of enterprise applications will embed task-specific agents by the end of this year, up from less than 5% last year.
That is the environment Project Arc is entering. For enterprise IT leaders, the question has shifted from whether to use agents to how to keep them from creating unmanaged risk.
- Which employee deployed which agent?
- What data did the agent access?
- Is there an audit log?
- Can the company stop it when something goes wrong?
ServiceNow has a natural advantage here. It already sits at the center of IT operations in many large companies, with asset inventories, change approvals and incident response workflows. Folding agents into that system is a logical extension.
NVIDIA brings compute and models. ServiceNow brings the enterprise process layer. Together, Project Arc becomes one of the first complete enterprise-agent packages that looks deployable rather than experimental.
Why model vendors should pay attention
The agent market is starting to separate into layers.
| Layer | Companies |
|---|---|
| Models | Anthropic, OpenAI, Google, Alibaba |
| Agent frameworks | LangChain, Crew, Anthropic Cowork |
| Desktop control | Claude Code, Cursor, Codex |
| Enterprise governance and workflow integration | ServiceNow, Microsoft Agent 365, SAP |
Anthropic and OpenAI are trying to move deeper into the enterprise layer. Anthropic has introduced Claude Cowork, while OpenAI has pushed The Deployment Company. But IT process ownership is not their native territory.
For a ServiceNow customer, Action Fabric plus NVIDIA's Nemotron can look like an agent added on top of existing IT workflows. That is a lower-friction path than adopting an outside agent platform from scratch.
The next fight in enterprise AI may not be won by the best model alone. It may be won by whoever controls governance, approvals, logs and workflow. Project Arc is an early move in that direction.
Sources: NVIDIA and ServiceNow Partner on New Autonomous AI Agents for Enterprises (NVIDIA Blog); Your AI Agents Are Already Inside the Perimeter (The Hacker News); CocoLoop; Microsoft Agent 365 now generally available (Microsoft Security Blog)