Jensen Huang walked onto the GTC 2026 stage and did something unusual for a hardware CEO: he spent more time talking about software infrastructure and autonomous agents than he did about GPUs. That’s not an accident. NVIDIA’s bet — the one they’re staking billions on — is that the next decade of AI value doesn’t come from selling more chips. It comes from owning the platform that runs the agents that run everything else. GTC 2026 was where that strategy came into full view.
The OpenClaw Moment: NVIDIA’s Linux Analogy Isn’t Casual
When Jensen Huang compared OpenClaw to Linux and Kubernetes, he wasn’t reaching for a flattering metaphor. He was making a specific claim: that OpenClaw is becoming the foundational runtime that the agent ecosystem builds on top of, the same way Linux became the substrate that cloud computing couldn’t function without.
His exact words — “OpenClaw gave us exactly what it needed at exactly the right time” — are worth sitting with. The framing is deliberate. Linux didn’t win because it was technically perfect; it won because it was open, composable, and showed up at the moment the internet needed a free, stable operating layer. Kubernetes won the container orchestration wars for similar reasons: it solved a real operational problem with an open standard at the moment enterprises were drowning in container sprawl.
NVIDIA is positioning OpenClaw as the equivalent for agentic AI: the open runtime that handles how agents are defined, how they communicate, how they’re orchestrated, and how they’re deployed — regardless of what hardware they run on. Which brings us to the part enterprises actually care about.
NemoClaw: What It Is and Why “Hardware Agnostic” Is the Headline
NemoClaw is NVIDIA’s enterprise layer built on top of OpenClaw. Think of it as OpenClaw with a suit on: same underlying runtime, but wrapped with enterprise security, privacy guardrails, and policy enforcement. It integrates with NVIDIA’s NeMo AI agent software suite, giving enterprises a coherent stack from model to deployment.
But the detail that should stop enterprise IT leaders cold is this: NemoClaw doesn’t require NVIDIA GPUs.
That’s a significant strategic move. Historically, NVIDIA’s software has been most valuable when running on NVIDIA hardware — CUDA is the canonical example of a software moat that only runs on their chips. NemoClaw breaks that pattern. By making the enterprise agent platform hardware agnostic, NVIDIA is prioritizing platform adoption over hardware lock-in. The logic isn’t hard to follow: if your agent runtime becomes the standard, you win even when companies run inference on AMD, Intel, or custom silicon. You get the telemetry, the ecosystem, the enterprise relationships — and you still sell a lot of GPUs to the customers who want maximum performance.
For enterprises evaluating agent infrastructure right now, NemoClaw represents a credible answer to the “what runtime do we standardize on” question that most companies are currently answering with a combination of LangChain, custom glue code, and hope. The partners announced at GTC — Adobe, Atlassian, Cisco, CrowdStrike, SAP, Salesforce, ServiceNow, and Siemens — aren’t logos on a slide. They’re organizations that have committed to building on this stack, which meaningfully de-risks the bet for anyone else considering it.
Nemotron 3 Super: The Numbers That Actually Matter
Announced on March 11, Nemotron 3 Super is NVIDIA’s most technically interesting model release in years — and the benchmarks deserve more attention than they’ve gotten outside of ML circles.
The architecture is a hybrid Mixture-of-Experts model: 120 billion total parameters, but only 12 billion are active at inference time. That’s what makes the efficiency numbers possible. Two novel contributions power it: LatentMoE routing (a new approach to which experts get activated and when) and native NVFP4 pretraining (training directly in 4-bit floating point, rather than post-training quantization). The result is 2.2x higher throughput than GPT-OSS-120B at comparable parameter counts.
The benchmark numbers are specific enough to be meaningful:
| Benchmark | Nemotron 3 Super | GPT-OSS-120B |
|---|---|---|
| SWE-Bench Verified (coding) | 60.47% | 41.90% |
| RULER at 1M tokens (long context) | 91.75% | 22.30% |
| Inference throughput vs GPT-OSS-120B | 2.2x higher | baseline |
The SWE-Bench gap is notable — nearly 20 percentage points better at real software engineering tasks. But the RULER result at one million tokens is the one that should get more attention. A 91.75% vs 22.30% spread at one million token context isn’t a marginal improvement; it’s a different capability class. This matters enormously for enterprise agents that need to reason over large codebases, long document histories, or complex multi-turn workflows where earlier context can’t be dropped without losing critical information.
Both the weights and the training recipe are open. Andrej Karpathy has been vocal in the past about how much the field benefits from open training recipes rather than just open weights — you can fine-tune weights, but you need the recipe to actually understand and replicate what was done. NVIDIA releasing both is a meaningful commitment to the open model ecosystem, and it makes Nemotron 3 Super a serious option for any enterprise that wants to run capable models on their own infrastructure.
Physical AI: Robots Are No Longer a Separate Category
One of the cleaner narratives from GTC 2026 is that NVIDIA has stopped treating physical AI as a side project. Cosmos models and the GR00T open models for humanoid robots were featured alongside the enterprise agent stack — not in a separate robotics track, but as part of the same coherent vision.
The framing matters. NVIDIA is arguing that the agent infrastructure they’re building for software — the runtimes, the orchestration, the safety layers — is the same infrastructure that will run physical agents. A humanoid robot in a warehouse and a software agent managing your Salesforce CRM are, at the infrastructure level, not that different: both need to perceive state, reason about goals, take actions, and operate within policy constraints.
OpenShell, the open source runtime for “self-evolving agents and claws” with built-in safety and security, is the piece that ties this together. The language around self-evolving agents is worth flagging — it’s ambitious, and the practical meaning in production environments will take time to understand. But the direction is clear: NVIDIA wants to own the safety and orchestration layer for agents that learn and adapt over time, whether they’re running in a data center or walking around a factory floor.
The Vera Rubin platform, with H300 GPUs targeting trillion-parameter models, is the hardware bet that supports all of this. Trillion-parameter models aren’t a near-term enterprise deployment reality for most organizations, but they