NVIDIA GTC 2026 Keynote Delivers Vera Rubin at 50 PFLOPS and Open-Source NemoClaw as Jensen Huang Declares the Agentic Infrastructure Era

NVIDIA GTC 2026 Keynote Delivers Vera Rubin at 50 PFLOPS and Open-Source NemoClaw as Jensen Huang Declares the Agentic Infrastructure Era

The Keynote That Rewrites AI Infrastructure

Jensen Huang took the stage at San Jose's SAP Center on Monday, March 16, before 30,000 attendees from 190 countries and delivered the most hardware-dense keynote in NVIDIA's history. Over two hours, Huang revealed production specifications for the Vera Rubin computing platform, launched the open-source NemoClaw agentic AI framework, and offered a surprise first look at the Feynman architecture built on TSMC's 1nm-class A16 process node.

The keynote's pre-show featured Perplexity CEO Aravind Srinivas, LangChain CEO Harrison Chase, Skild AI CEO Deepak Pathak, and Mistral AI CEO Arthur Mensch — a lineup that framed GTC 2026 as a platform ecosystem event rather than a product launch.

Vera Rubin: 5x Blackwell Inference Performance

The Vera Rubin platform represents NVIDIA's largest generational performance leap. The Rubin GPU packs 336 billion transistors — a 62 percent increase over Blackwell's 208 billion — and ships with 288GB of HBM4 memory manufactured by Samsung and SK Hynix.

Key performance claims:

• 50 PFLOPS inference — 5x improvement over Blackwell
• 17.5 PFLOPS training — 3.5x improvement over Blackwell
• 288GB HBM4 — first GPU to use fourth-generation high-bandwidth memory
• NVLink 6 — 260 TB/s aggregate bandwidth in the NVL72 rack configuration

The full Vera Rubin platform encompasses six components: the Rubin GPU, Vera CPU (successor to Grace), a next-generation DPU, advanced NICs, NVLink 6 scale-up interconnect, and Ethernet switching for scale-out connectivity. This is not a chip launch — it is a complete computing platform designed from the ground up for agentic AI inference workloads.

The Thinking Machines Lab partnership, previously announced as a gigawatt-scale Vera Rubin deployment worth up to $50 billion, was expanded during the keynote with additional deployment timeline details. Mira Murati's company will be among the first production customers for the NVL72 rack configuration.

NemoClaw: Open-Source Agentic AI Framework

The strategic surprise of the keynote was NemoClaw — an open-source platform for building and deploying enterprise AI agents that runs on any hardware, including AMD, Intel, and CPU-only systems.

NemoClaw features:

• Multi-agent architecture with supervisor and worker agents
• Privacy-first design with full enterprise data control
• Built-in security including audit logs and compliance tooling
• Tool integration for browsers, code execution, databases, and APIs
• Hardware-agnostic deployment on any GPU vendor or CPU-only infrastructure

Confirmed enterprise integration partners include Salesforce, Cisco, Google Cloud, Adobe, and CrowdStrike. Each partner receives early access in exchange for contributing to the open-source project.

The hardware-agnostic decision represents a fundamental shift in NVIDIA's strategy. Every previous major NVIDIA software release — CUDA, TensorRT, NeMo, Triton Inference Server — was optimized exclusively for NVIDIA GPUs. NemoClaw's openness follows the Android model: control the platform to drive hardware adoption, rather than using hardware lock-in to force platform adoption.

GTC attendees can build and deploy agents hands-on through the OpenClaw toolkit at GTC Park running March 16-19.

Feynman Architecture: First Look at 2028 Silicon

In the keynote's biggest surprise, Huang provided a first look at the Feynman architecture — NVIDIA's post-Rubin GPU design scheduled for 2028 delivery. Feynman will be built on TSMC's A16 node, marking the first 1nm-class process for a GPU, and will include backside power delivery for improved transistor density and power efficiency.

Reports indicate NVIDIA is considering outsourcing Feynman's I/O die to Intel using 14A or 18A process and EMIB advanced packaging — a supply chain diversification move that would reduce NVIDIA's dependency on TSMC for the first time. The GPU compute die would remain at TSMC.

Showing Feynman two years before launch provides roadmap visibility for enterprises making multi-billion-dollar AI infrastructure commitments. It signals that NVIDIA's architecture cadence of roughly 18-24 months will continue through the end of the decade.

What This Means

GTC 2026 Day One establishes NVIDIA's transition from semiconductor company to AI platform company. The Vera Rubin hardware provides the compute foundation. NemoClaw provides the software platform. Feynman provides the roadmap confidence.

For developers, the immediate action is evaluating NemoClaw for enterprise agent workflows. For enterprises, the calculus shifts toward building on NemoClaw now and migrating to Vera Rubin hardware when production units ship. For competitors, NemoClaw's hardware-agnostic design is both an opportunity (validation of non-NVIDIA silicon) and a threat (NVIDIA controls the platform layer).

The conference continues through March 19 with technical sessions, workshops, and the OpenClaw hands-on experience. Jensen Huang declared that the agentic AI infrastructure era has begun. Based on what shipped today — open-source code and production silicon specifications — the declaration has substance behind it.

Sources:

• NVIDIA GTC 2026 Official Blog
• NVIDIA CEO Jensen Huang Keynote
• CNBC: NVIDIA Plans Open-Source AI Agent Platform NemoClaw
• TechCrunch: How to Watch Jensen Huang's GTC 2026 Keynote