Five years ago NVIDIA was the company that made your models train faster. Today they’re building the runtime your agents live inside. That trajectory tells you everything about where enterprise AI infrastructure is heading.

The NeMo Era: Models and Training

When I first started using NVIDIA NeMo, it was a model training and fine-tuning framework. You’d take a foundation model, customise it with your data, optimise it with TensorRT, and deploy it on NVIDIA hardware. The value proposition was performance. Faster training, better inference, tighter hardware integration.

That was a GPU-adjacent software play. Important, but bounded. NVIDIA made the chips and the libraries that helped you use the chips.

The Agent Toolkit: From Models to Systems

At GTC 2026, NVIDIA announced the Agent Toolkit — and the scope expanded dramatically. It’s not a model framework anymore. It’s a full-stack agent development platform.

Agent Toolkit includes open models like Nemotron, open agents like AI-Q (which topped the DeepResearch Bench accuracy leaderboards), open skills like cuOpt for decision optimisation, and open runtimes like OpenShell. The stack now covers everything from the model weights to the sandbox the agent executes in.

The partner list reads like an enterprise software who’s who. Adobe, Atlassian, Box, Cadence, Cisco, CrowdStrike, Salesforce, SAP, ServiceNow, Siemens, Synopsys. These aren’t experimental integrations. These are production-grade platform vendors building on NVIDIA’s agent infrastructure.

NemoClaw: The Consumer-to-Enterprise Bridge

NemoClaw is where the strategy gets interesting. OpenClaw is the fastest-growing open source project in history — it’s become the default runtime for autonomous personal AI agents. But OpenClaw was built for developers, not enterprises. It didn’t have sandboxing, policy enforcement or privacy controls.

NemoClaw wraps OpenClaw with NVIDIA Agent Toolkit software. One command installs Nemotron models locally and the OpenShell runtime. The result is an agent platform that runs on anything from an RTX laptop to a DGX Station, with enterprise-grade security baked in from the start.

This is a classic NVIDIA play: take something the community built, add the infrastructure layer the enterprise needs, and ship it as an open source stack that runs best on NVIDIA hardware.

The Strategic Pattern

Looking at this trajectory from the outside, the pattern is clear.

NeMo gave NVIDIA a position in model development. TensorRT and NIM gave them inference. Agent Toolkit gave them the orchestration and evaluation layer. OpenShell gives them the runtime and governance layer. NemoClaw packages it all for the largest developer community in AI.

Each step moves NVIDIA further up the stack. And each step makes the hardware stickier. If your agent runtime, model serving, policy engine and privacy routing all run on NVIDIA infrastructure, switching costs compound at every layer.

Why This Matters for Architects

I see two implications for enterprise architects.

First, the “GPU vendor” framing is obsolete. NVIDIA is a full-stack AI platform company. Any architecture evaluation that only considers them at the compute layer is missing the majority of their offering.

Second, the competition isn’t just about models anymore. It’s about who controls the agent runtime — the layer that determines what agents can see, do and access. Microsoft is approaching this from the cloud identity side. OpenAI from the model side. NVIDIA from the infrastructure side. These are converging on the same control plane, and your architecture decisions need to account for that convergence.

Where This Is Going

The next 12 months will determine whether the enterprise agent stack consolidates around a few dominant platforms or stays fragmented. NVIDIA is betting on open source at every layer — Apache 2.0 across NeMo, Agent Toolkit and OpenShell — while making sure everything runs best on their silicon.

That’s a strategy that worked for CUDA. It’s a strategy that worked for TensorRT. Whether it works for agent runtimes depends on whether enterprises decide that governance and security need to live at the infrastructure level rather than the application level.

From what I’ve seen in production deployments, they do. And if that’s the case, NVIDIA’s move from NeMo to NemoClaw is the most strategically significant expansion they’ve made since CUDA.

• What NVIDIA NemoClaw Signals About the Future of Enterprise Agent Architecture
• Microsoft Agent Framework Foundry MCP and Aspire in Practice
• What Architects Can Learn From NVIDIA NemoClaw
• NemoClaw and the Rise of Secure Agent Platforms in Enterprise AI
• OpenAI Just Bought Promptfoo. That’s a Bigger Deal Than Most People Realise