Jensen Huang took the stage at GTC 2026 and did what he usually does: made the infrastructure story sound like the most important thing happening in technology right now. Whether you think that’s true depends on how much you care about what’s running underneath the AI applications everyone is actually building. This year, the big themes weren’t just new chips — they were enterprise AI agents, open-weight models that genuinely compete with the closed giants, and a physical AI push that’s starting to feel less theoretical. Here’s everything that matters from the keynote, with the context you need to understand why.
NemoClaw: NVIDIA’s Enterprise Agent Play
The announcement that got the most attention in enterprise circles was NemoClaw — a platform built on top of OpenClaw that layers in exactly what large organizations say they can’t live without before deploying AI agents: security, privacy guardrails, and policy enforcement at scale.
The structure here matters. OpenClaw is the foundation — an open-source runtime for building and running agents. NemoClaw wraps that with enterprise-grade controls and plugs into NVIDIA’s NeMo software suite, which handles the agent software layer. Think of it like the difference between raw Linux and Red Hat Enterprise Linux. Same kernel, very different support and compliance story.
What’s notably smart about NVIDIA’s positioning: NemoClaw is hardware agnostic. It doesn’t require NVIDIA GPUs. That’s not altruism — that’s a land-and-expand strategy. Get into enterprise environments through the software layer, and the hardware follows. Jensen Huang made the parallel explicit in the keynote, comparing OpenClaw’s potential to Linux and Kubernetes: “OpenClaw gave us exactly what it needed at exactly the right time… like Linux, like Kubernetes.”
That’s a high bar to set in public. But the comparison isn’t crazy. Kubernetes didn’t win because Google forced it — it won because it became the lowest-friction standard for container orchestration at the moment the industry needed one. If agents are the next layer of infrastructure that enterprises need to orchestrate at scale, OpenClaw is positioning to be that standard before anyone else locks it in.
NVIDIA also announced the NVIDIA Agent Toolkit — open-source models and software for building enterprise agents — alongside NVIDIA OpenShell, an open-source runtime specifically for “self-evolving agents and claws” with built-in safety and security controls. The “self-evolving” framing is worth watching carefully. It suggests agents that can modify their own behavior or tooling over time, which is where real capability and real risk both live.
Nemotron 3 Super: The Model That Should Change How You Think About Open Weights
Announced March 11, 2026, Nemotron 3 Super is the most technically interesting model release from this GTC. 120 billion total parameters, 12 billion active — a hybrid Mixture-of-Experts architecture that means you’re getting 120B-class reasoning while only activating a fraction of the parameters on any given inference call. That’s the core efficiency trick of MoE, and NVIDIA is implementing it with a novel approach they’re calling LatentMoE routing, combined with native NVFP4 pretraining.
The benchmarks are striking enough to be worth quoting directly rather than paraphrasing:
| Benchmark | Nemotron 3 Super | GPT-OSS-120B |
|---|---|---|
| SWE-Bench Verified | 60.47% | 41.90% |
| RULER (1M token context) | 91.75% | 22.30% |
| Inference Throughput | Baseline | 2.2x lower than Nemotron 3 Super |
That RULER number at one million tokens is the one that should stop you. 91.75% versus 22.30% is not a marginal improvement — it’s a different category of capability for long-context retrieval and reasoning. If you’re building applications that need to process entire codebases, legal documents, or large research corpora in a single context window, that gap is the difference between a useful tool and an unreliable one.
The SWE-Bench score matters too. SWE-Bench Verified measures whether a model can actually resolve real GitHub issues — it’s one of the harder evals for coding capability because it requires understanding codebases, writing correct patches, and making them pass tests. 60.47% puts Nemotron 3 Super ahead of where GPT-4 class models were performing on the same benchmark not long ago. For a developer using this in an agentic coding workflow, that improvement compounds fast.
Critically: open weights, open training recipe. The research community can inspect this, fine-tune it, and build on it without going through NVIDIA’s API. That’s a genuine contribution to the open-source AI ecosystem, and it’s how NVIDIA competes with Anthropic and OpenAI on a dimension neither of them can easily match.
AI-Q Blueprint: Agentic Search That’s Actually Winning Benchmarks
NVIDIA’s AI-Q Blueprint is their bet on agentic search — the class of systems that don’t just retrieve documents but reason across them to produce synthesized answers. The claim is that it tops the DeepResearch Bench accuracy leaderboards, which is a specific and verifiable claim worth holding them to as independent evaluations come in.
DeepResearch Bench is designed to test exactly this: can an AI agent conduct multi-step research, synthesize findings across sources, and produce accurate, well-reasoned outputs? It’s the kind of benchmark that Aravind Srinivas at Perplexity would be watching closely, because it’s directly relevant to what search-native AI products need to do well.
The practical question for anyone building with AI-Q Blueprint is: what does “agentic search” mean inside an enterprise context where data lives in SharePoint, Salesforce, ServiceNow, and SAP — and where you can’t just send everything to an external API? The combination of AI-Q with NemoClaw’s privacy guardrails is where this becomes interesting. An agent that can search across enterprise data sources, maintain access controls, and produce accurate synthesized answers without leaking data across permission boundaries is genuinely useful in a way that public-web RAG systems aren’t.
Physical AI: Cosmos, GR00T, and the Humanoid Robot Push
NVIDIA’s physical AI work got significant stage time at GTC 2026. Two main pillars: Cosmos models for world model development, and GR00T open models specifically for humanoid robots.
The GR00T release continuing as open models is notable because humanoid robotics is a space where data scarcity is the core bottleneck. Every serious player — Figure, 1X, Boston Dynamics, Apptronik — needs massive amounts of training data for embodied AI systems. Open foundation