March 18, 2026
By Bob O'Donnell
Want to know where the tech industry is headed? While that’s always been a difficult question to answer in a comprehensive way, these days, listening to Nvidia CEO Jensen Huang’s keynote at the company’s annual GTC event is a good way to get started.
The message from this year’s event was clear. The age of agentic AI is already upon us, and the world needs to evolve to meet that new reality...quickly. The same applies to Nvidia itself, and the company made numerous announcements at the show that reflect this huge and important shift.
The most important message from Nvidia’s GTC keynote wasn’t simply that more powerful chips are coming. It was that the shift from chatbots to autonomous agents is reshaping the entire AI stack. That is why Nvidia used the event to pair agent software, such as NemoClaw and OpenShell, with production-ready inference software in Dynamo 1.0 and a broader Vera Rubin architecture that now spans GPUs, CPUs, networking, storage, and Groq-based inference accelerators inside a single AI-factory design.
One of the buzziest announcements was the debut of a new agent tool called NemoClaw, which builds on the popular new open-source software application called OpenClaw. Nvidia describes NemoClaw as the stack for OpenClaw; it uses NVIDIA Agent Toolkit and installs OpenShell, which is the runtime/sandbox layer that adds privacy and security guardrails to the standard OpenClaw. As powerful and popular as OpenClaw has become, it’s widely recognized that file access limitations and other security protections are needed to make it a more reasonable option for businesses (and security-sensitive individuals) to use, so it’s good to see Nvidia tackling this issue so quickly. Even better is that Nvidia is making OpenShell open source so that security experts, software developers, and more can verify and enhance its capabilities.
While there are many types of AI-powered agents available, OpenClaw has become the spark that’s triggered a meteoric rise in attention focused on agentic AI. By getting people to understand that it turns AI from something that can look up and synthesize information to something that can perform real-world actions on your behalf, it’s created a whole range of new ideas on how computing can be done. Want a tool that can create personalized automations of regular workflows, build your own custom applications or put together comprehensive research reports all via a simple chatbot-style prompt? That’s what these new “claw” tools can do.
Taking that idea even further, Nvidia’s Huang discussed how he believes AI agents will lead to an evolution of existing software tools and companies to the point where they will start offering “agents as a service”. After several decades of traditional software evolution, the idea that software is shifting from tools we use to labor we manage is a very big change.
But the impact of this agentic shift goes even further and reaches into Nvidia’s latest generation AI infrastructure hardware as well. Notably, while the company shared the full details around its upcoming Vera Rubin platform with its latest generation Arm-based CPU and new GPU designs as expected, the bigger news was on new Vera CPU and Groq-based LPU (Language Processing Unit) racks.
Because of how AI workloads are evolving—particularly when agents are involved—there’s a growing recognition that it’s going to take a suite of different chip technologies to most efficiently tackle new agent-driven inference queries that tap into the growing range of sophisticated MoE (Mixture of Experts) models. CPUs, for example, are proving to be the best option for coordinating and orchestrating the requests that agents generate, while Groq’s LPU architecture is proving to be incredibly speedy at generating output tokens (e.g., answers) with larger context windows. Underlying these changes is the company’s Dynamo software, which provides a new level of orchestration across all the elements in the system as it processes inference-based workloads.
To address this evolution, Nvidia unveiled two entirely new rack architectures—one that only uses Vera CPUs and another that uses the Nvidia Groq LPX3, leveraging the Groq technology that Nvidia only licensed a few months back. The LPX-focused rack, which is intended for advanced inferencing requests, includes 256 interconnected LPX3 chips, each of which includes 500 MB of speedy SRAM to quickly generate inference tokens. Both of the new racks attach to existing Vera Rubin NVL72 racks via the company’s SpectrumX networking technology. Collectively, they are part of a new bigger picture AI Supercomputer architecture that is the complete fulfillment of the new Vera Rubin platform. This AI Supercomputer pod design now includes 7 different chips: the Vera CPU, Nvidia Rubin GPU, Nvidia NVLink 6 Switch, Nvidia ConnectX-9 SuperNIC, Nvidia BlueField-4 DPU and Nvidia Spectrum-6 Ethernet switch, along with the newly integrated Nvidia Groq LPX3 across five rack systems. Beyond these technical details, the more important point is that Nvidia is now positioning the full data center as the core element of computing—not a single chip or even a single system. Even more, Nvidia is working to make it clear that it sees itself as the full-stack infrastructure company for the agentic era.
As Nvidia’s Huang explained, initial usage for the LPX chips is for premium tiers of token performance (over 1,000 tokens a second), but it’s not hard to imagine those capabilities being integrated into the most basic levels of Nvidia AI infrastructure designs in the future.
Knowing that these systems have grown immensely complex to install and operate, Nvidia also leveraged its Omniverse line of AI-based digital twin modeling and simulation tools, to create an AI Factory blueprint it’s calling Vera Rubin DSX AI Factory Reference Design. Not only does the model provide simulations and management tools for the AI factory, but it also offers a variety of API connections that companies involved with the physical and electrical construction of these facilities can tap into.
The move to agentic AI is happening at a staggeringly fast pace, and it highlights how rapidly this market continues to evolve. Just as many people were starting to get used to interacting with chatbots, the focus has definitely shifted towards AI agents and the productivity enhancements they can enable. Kudos to Nvidia for being able to switch so quickly to this new agentic reality by creating new software offerings and AI infrastructure enhancements optimized to meet these new needs.
Here’s a link to the original column: https://www.linkedin.com/pulse/nvidia-uses-gtc-recast-itself-agentic-ai-era-bob-o-donnell-p02ec
Bob O’Donnell is the president and chief analyst of TECHnalysis Research, LLC a market research firm that provides strategic consulting and market research services to the technology industry and professional financial community. You can follow him on LinkedIn at Bob O’Donnell or on Twitter @bobodtech.
| 
