NVIDIA GTC 2026 Keynote Day One — Vera Rubin, NemoClaw, and Feynman Signal the Agentic AI Infrastructure Era

Thirty thousand people from 190 countries packed into San Jose's SAP Center on Monday morning. Jensen Huang walked onstage in his trademark leather jacket at exactly 11:00 AM Pacific, and over the next two hours delivered what may be the most consequential technology keynote since Steve Jobs introduced the iPhone. The difference is that Huang was not introducing a single product. He was introducing an entire computing paradigm — one where autonomous AI agents run on purpose-built silicon, connected by networking fabric that did not exist twelve months ago, deployed through an open-source framework that makes NVIDIA's own hardware optional.

That last part is the real story. NVIDIA, the company that built a $3 trillion market cap on selling GPUs, just released an agentic AI platform that runs on AMD, Intel, and CPU-only infrastructure. If you want to understand why that decision is either the most brilliant strategic move in semiconductor history or the beginning of NVIDIA's commoditization nightmare, you need to understand everything that happened on stage today.

As I covered in my pre-GTC analysis of NVIDIA's full-stack AI takeover, Jensen Huang has been methodically positioning NVIDIA across every layer of the AI stack — from energy to applications. Today's keynote confirmed the strategy and accelerated the timeline. What nobody expected was how aggressively NVIDIA would embrace open source as the delivery mechanism for that strategy.

The Pre-Show: A Who's Who of AI Leadership

Before Huang took the main stage, GTC's pre-show featured a lineup that read like a Fortune 500 of AI companies. Perplexity CEO Aravind Srinivas discussed the convergence of search and agents. LangChain CEO Harrison Chase outlined the middleware layer for agentic systems. Skild AI CEO Deepak Pathak previewed embodied AI progress. Mistral AI CEO Arthur Mensch talked about open-weight models meeting enterprise requirements.

The pre-show itself was a statement. These are not NVIDIA employees. They are the leaders of companies building the agentic AI ecosystem that NVIDIA wants to power. By putting them on stage first, Huang was framing GTC 2026 not as a product launch but as a platform ecosystem event — the way Apple frames WWDC or Google frames I/O.

Vera Rubin: The Numbers That Change Everything

The centerpiece of the keynote was the full reveal of the Vera Rubin computing platform. Named after the astronomer who proved the existence of dark matter, the platform is NVIDIA's most ambitious hardware release since the original CUDA-capable GPUs launched in 2006.

Raw Specifications

The Rubin GPU packs 336 billion transistors onto a single die. For context, the Blackwell B200 had 208 billion. That is a 62 percent increase in transistor density in a single generation, enabled by a combination of architectural improvements and TSMC's advanced packaging technology.

Memory is where the generational leap becomes stark. Vera Rubin ships with 288GB of HBM4 memory — the first GPU to use the fourth generation of high-bandwidth memory. Samsung and SK Hynix are both manufacturing the HBM4 stacks, and NVIDIA has secured production capacity through 2027.

Performance Claims

Huang claimed Vera Rubin delivers 5x inference performance and 3.5x training performance compared to Blackwell. Those are not incremental improvements. They represent the kind of generational leap that rewrites the economics of every AI deployment currently running on Blackwell clusters.

The inference number is the one that matters most for the agentic AI thesis. Autonomous agents that execute multi-step workflows require continuous inference — not the burst inference patterns of chatbots. A system that can handle 5x more inference throughput per watt fundamentally changes what kind of agent architectures become economically viable.

The Full Platform

Vera Rubin is not just a GPU. The platform encompasses six components working in concert: the Rubin GPU for compute, the Vera CPU as the host processor (successor to Grace), a next-generation DPU for data processing, advanced NICs for networking, NVLink 6 for scale-up interconnect, and Ethernet switching infrastructure for scale-out connectivity.

NVL72: The Rack-Scale Supercomputer

The Vera Rubin NVL72 rack configuration connects 72 GPUs through NVLink 6, delivering 260 TB/s of aggregate interconnect bandwidth. To put that in perspective, the entire bandwidth of the global internet is estimated at roughly 1,000 TB/s. A single NVL72 rack delivers a quarter of that bandwidth between 72 chips sitting in the same cabinet.

This bandwidth is not academic. Training frontier models at the scale of GPT-5 or Claude Opus requires moving enormous volumes of data between GPUs during every training step. The NVLink 6 fabric eliminates the interconnect bottleneck that has constrained multi-GPU training for the past three years. When Thinking Machines Lab announced their gigawatt-scale partnership with NVIDIA for Vera Rubin deployment, the NVL72 architecture was the centerpiece — not the individual GPU specs.

NemoClaw: NVIDIA's Open-Source Bet on Agentic AI

If Vera Rubin was the expected hardware announcement, NemoClaw was the strategic surprise. NVIDIA formally launched NemoClaw as an open-source platform for building and deploying enterprise AI agents — and the most striking feature is what it does not require: NVIDIA hardware.

What NemoClaw Actually Is

NemoClaw is a complete framework for creating, orchestrating, and managing autonomous AI agents in enterprise environments. It includes a supervisor-worker agent architecture, built-in tool integration for browsers, code execution, databases, and APIs, enterprise-grade security with audit logs and compliance tooling, and a privacy-first design that keeps all data under the enterprise's control.

The platform ships with native connectors for the tools that enterprise agents actually need: web browsers for research tasks, code interpreters for data analysis, database connectors for information retrieval, and API clients for integration with existing enterprise systems. NVIDIA is not building another chatbot framework. They are building the operating system for autonomous enterprise workflows.

The Hardware-Agnostic Decision

Here is where the strategy gets interesting. NemoClaw runs on NVIDIA GPUs, AMD GPUs, Intel GPUs, and even CPU-only infrastructure. NVIDIA has never released a major software platform that was hardware-agnostic. CUDA was explicitly designed to lock developers into NVIDIA silicon. TensorRT requires NVIDIA GPUs. Even their networking software assumes NVIDIA NICs and switches.

NemoClaw being hardware-agnostic is not altruism. It is the same strategy that made Android dominant in mobile. Google gave away Android for free because controlling the platform was more valuable than controlling the hardware. NVIDIA is betting that controlling the agentic AI platform layer — the framework that enterprises standardize on for building and deploying agents — will drive more GPU purchases than any hardware lock-in could achieve.

The logic works like this: if NemoClaw becomes the enterprise standard for agentic AI, developers will build and test on whatever hardware they have. But when they need to scale to production — when they need the throughput, latency, and power efficiency that real-time agent workloads demand — they will buy Vera Rubin. The platform creates the demand. The hardware fulfills it.

Enterprise Partnerships

NVIDIA has already been pitching NemoClaw to major enterprise software companies. Confirmed integration partners include Salesforce, Cisco, Google Cloud, Adobe, and CrowdStrike. Each partner gets early access to the platform in exchange for contributing to the open-source project and building native integrations.

The partnership strategy mirrors the early CUDA ecosystem play. NVIDIA seeded CUDA adoption by working directly with researchers and framework developers. Now they are seeding NemoClaw adoption by working directly with enterprise software vendors. The pattern is identical — the scale is exponentially larger.

This aligns directly with my prediction that 50 percent of enterprise agentic AI pilots will reach production by Q4 2026. NemoClaw dramatically lowers the barrier to production deployment by providing the security, compliance, and orchestration tooling that has been missing from existing agent frameworks. The prediction was made before NemoClaw was announced. Its arrival makes the timeline even more plausible.

OpenClaw: The Developer On-Ramp

GTC attendees can participate in NVIDIA's build-a-claw event at GTC Park from March 16-19, where they can customize and deploy proactive AI assistants using the OpenClaw toolkit. OpenClaw is the developer-friendly wrapper around NemoClaw — think of it as Create React App for enterprise AI agents.

The naming is deliberate. "Claw" references the robotic claw machines that grab prizes — but in NemoClaw's context, the claw grabs data, executes workflows, and delivers results. The branding positions AI agents not as threatening autonomous systems but as helpful tools that fetch things for you. It is the kind of subtle positioning that Jensen Huang has mastered over three decades of product launches.

Feynman: The 2028 Architecture That Nobody Expected to See

The biggest surprise of the keynote was a first look at the Feynman architecture — NVIDIA's post-Rubin GPU design that is not scheduled to ship until 2028. Showing Feynman at GTC 2026 follows the pattern Huang established at GTC 2025 when he previewed Vera Rubin two years before its launch. But the Feynman details revealed today go far beyond a typical early preview.

Technical Details

Feynman will be built on TSMC's A16 node — a 1nm-class process technology that represents the most advanced semiconductor manufacturing process ever used for a GPU. Current Blackwell GPUs use TSMC's 4NP node. Vera Rubin advances to 3nm-class. Feynman leaps to 1nm-class.

The A16 process includes backside power delivery — a technique where power is routed through the back of the chip rather than competing with signal routing on the front. This allows NVIDIA to pack more transistors into the same area while maintaining power delivery to all of them. The result is a GPU that will likely exceed 500 billion transistors while potentially consuming less power per operation than Vera Rubin.

The Intel Supply Chain Angle

One of the more intriguing details to emerge is that NVIDIA is reportedly considering outsourcing Feynman's I/O die to Intel — specifically using Intel's 14A or 18A process and EMIB advanced packaging. This would be the first time NVIDIA has used Intel as a manufacturing partner for a major GPU component.

The strategic logic is risk management. By splitting Feynman's manufacturing between TSMC (for the GPU compute die) and Intel (for the I/O die), NVIDIA diversifies its supply chain dependency. The Taiwan Semiconductor Manufacturing Company produces the vast majority of NVIDIA's chips today. Any disruption to TSMC — geopolitical, natural disaster, or capacity constraint — would cripple NVIDIA's entire product line. Intel as a secondary manufacturer provides a hedge that NVIDIA has never had before.

What Feynman Means for the Roadmap

Showing Feynman in 2026 sends a clear message to customers and competitors: NVIDIA's roadmap is locked in through 2028 and beyond. Companies making multi-year AI infrastructure investments — the kind measured in billions of dollars — need visibility into at least three hardware generations before committing. Feynman gives them that visibility.

The cadence is accelerating. Hopper to Blackwell was roughly 18 months. Blackwell to Vera Rubin is about 18 months. Vera Rubin to Feynman may be closer to 24 months, but the transistor density improvements are larger at each step. NVIDIA is maintaining Moore's Law-like scaling through a combination of process node advances, architectural innovations, and advanced packaging techniques — even as the semiconductor industry at large has declared Moore's Law dead.

The Agentic AI Infrastructure Stack

The most important takeaway from GTC 2026 is not any single product announcement. It is the coherent vision of an agentic AI infrastructure stack that connects hardware, networking, frameworks, and applications into a unified platform.

Five Layers of the Stack

Huang explicitly laid out NVIDIA's five-layer model for agentic AI infrastructure during the keynote:

Layer 1: Silicon — Vera Rubin GPUs and Vera CPUs provide the compute foundation. The 5x inference improvement over Blackwell makes always-on agent workloads economically feasible for the first time.

Layer 2: Networking — NVLink 6 and next-generation Ethernet switching connect GPUs at 260 TB/s aggregate bandwidth per rack. This eliminates the interconnect bottleneck that has constrained agent system performance.

Layer 3: Platform — NemoClaw provides the framework for building, deploying, and managing AI agents. Its hardware-agnostic design ensures adoption regardless of a customer's current GPU vendor.

Layer 4: Models — NVIDIA is investing heavily in open-weight models through its $26 billion open-source commitment. These models are optimized for NVIDIA hardware but run anywhere.

Layer 5: Applications — Partner integrations with Salesforce, CrowdStrike, Adobe, and others bring NemoClaw-powered agents into existing enterprise workflows.

This is the full-stack strategy I analyzed before GTC, but the keynote made the execution timeline concrete. Every layer has a product shipping in 2026. Every layer connects to every other layer. And critically, every layer except silicon is available to non-NVIDIA hardware users — creating platform adoption that eventually drives hardware sales.

The Economic Case for Agentic Infrastructure

The economic argument Huang presented was straightforward: AI inference costs are dropping exponentially, but agent workloads are growing even faster. The net result is that total AI compute spending will increase even as per-query costs fall.

By 2028, according to NVIDIA's projections, agent-specific compute spending will exceed both inference and training spending combined. This is the market that Vera Rubin's 5x inference improvement and NemoClaw's deployment framework are designed to capture. The Morgan Stanley analysis predicting a massive AI breakthrough in H1 2026 cited exactly this compute scaling dynamic as the foundation for their thesis.

The Thinking Machines Partnership: Gigawatt-Scale Computing

Huang used the keynote to expand on the partnership with Mira Murati's Thinking Machines Lab — a multiyear deal to deploy at least one gigawatt of Vera Rubin systems. The scale is staggering. One gigawatt is enough power to run approximately 800,000 homes. NVIDIA is proposing to use that same power to run a single AI compute cluster.

The partnership validates two things simultaneously. First, there is real customer demand for Vera Rubin at scale — this is not vaporware. Second, the AI compute buildout is accelerating, not decelerating. Despite a recent tech sector pullback and questions about AI ROI, Thinking Machines is committing tens of billions of dollars to NVIDIA's newest hardware.

Murati's involvement adds credibility. As the former CTO of OpenAI, she has direct experience with the compute requirements of frontier model training. If she is betting her company on Vera Rubin at gigawatt scale, the hardware performance claims are likely conservative, not aspirational.

Market Implications: What GTC Day One Means

For Developers

The immediate developer action item is NemoClaw. If you are building enterprise AI agents — or planning to — NemoClaw just became the default framework to evaluate. The hardware-agnostic design means you can start building today on whatever infrastructure you have. The enterprise security and compliance features mean you can actually ship to production without building those layers yourself.

The shift from AI assistant to AI operator that GPT-5.4 and Cursor Automations kicked off earlier this month just got its infrastructure layer. NemoClaw is the missing piece between "AI can theoretically do this task" and "AI is actually doing this task in a production enterprise environment."

For Enterprises

If you are an enterprise CTO evaluating AI infrastructure investments, GTC Day One just compressed your decision timeline. Vera Rubin's 5x inference improvement means that waiting for the next generation actually has a clear payoff — but it also means that current Blackwell deployments will look increasingly expensive per inference operation as Vera Rubin enters production.

The smart play is to start NemoClaw adoption now on existing infrastructure, then scale to Vera Rubin when production units ship. NemoClaw's hardware-agnostic design makes this migration path seamless. Build the agent workflows today. Upgrade the silicon tomorrow.

For Competitors

AMD, Intel, and custom silicon startups face a strategic dilemma. NemoClaw running on their hardware validates their AI compute capabilities — but it also means NVIDIA is defining the platform layer. When the platform vendor also makes the best hardware, the "works everywhere but works best here" dynamic is powerful. Android runs on every phone, but Google's Pixel consistently gets the best Android experience first.

NVIDIA's 80 percent market share in AI training hardware is well-documented. The question GTC Day One raises is whether NemoClaw will extend that dominance into the agentic AI inference market — or whether the hardware-agnostic design will actually enable competitors to capture meaningful share. The answer probably depends on whether NemoClaw's "works best on NVIDIA" performance differential is 10 percent (competitors can compete) or 50 percent (competitors are irrelevant).

What Developers Should Build Now

GTC Day One creates a clear priority list for developers and engineering teams working in the AI space.

Priority 1: Evaluate NemoClaw

Download the open-source release. Build a proof-of-concept agent that solves a real problem in your organization. The supervisor-worker architecture maps naturally to most enterprise workflows — a supervisor agent that understands the business objective, delegating to worker agents that have specific tool access (database queries, API calls, browser automation, code execution).

Priority 2: Instrument for Agentic Workloads

The biggest gap in most AI deployments is observability for agent systems. When a chatbot fails, you get a bad response. When an autonomous agent fails, it might execute the wrong workflow, make incorrect API calls, or corrupt data before anyone notices. NemoClaw includes audit logging, but you need to build the monitoring and alerting layer on top of it.

Priority 3: Plan Your Compute Migration

Vera Rubin will not ship in volume until later this year, but the pricing and availability signals from GTC will shape procurement decisions now. If your agent workloads are currently inference-bound on Blackwell, start planning the Vera Rubin migration. The 5x inference improvement will likely change the economics of your entire AI budget.

Priority 4: Rethink Your AI Architecture

If you built your AI system around a single-model, request-response pattern — the way most chatbot deployments work — NemoClaw's multi-agent architecture requires a fundamental rethink. The shift from "user asks, model responds" to "supervisor delegates, workers execute, results aggregate" is as significant as the shift from monolithic applications to microservices. Start that architectural planning now, even if you do not adopt NemoClaw immediately.

The Bigger Picture: NVIDIA as an AI Platform Company

Step back from the individual announcements and the pattern is unmistakable. NVIDIA is no longer a semiconductor company that makes software to sell chips. It is a platform company that makes chips to anchor its platform.

The evidence is in the numbers. NVIDIA's open-source investment exceeds $26 billion. NemoClaw is hardware-agnostic. The Thinking Machines deal is structured around a platform commitment, not just a hardware purchase order. Jensen Huang spent more keynote time on NemoClaw's software architecture than on Vera Rubin's transistor count.

This transformation mirrors what happened to Microsoft in the 2010s. Satya Nadella took a company defined by Windows and Office — products that ran on Microsoft platforms — and rebuilt it around Azure, a platform that runs everything. The Windows tax was replaced by the Azure platform fee. NVIDIA is attempting the same transformation: the CUDA lock-in is being replaced by the NemoClaw platform gravity.

The difference is scale and timing. Microsoft's transformation took a decade. NVIDIA is attempting it during the most explosive growth phase in computing history. If agentic AI delivers on even half its promise — and the enterprise partnerships suggest it will — NVIDIA's platform play could make its hardware revenue look modest by comparison.

What Happens Next at GTC

The keynote was Day One of a four-day conference. Over the next three days, GTC will host hundreds of technical sessions, workshops, and partner announcements. Key sessions to watch include deep dives on NemoClaw's architecture, Vera Rubin deployment best practices, Feynman roadmap details, and partner integration showcases.

The build-a-claw event at GTC Park runs all four days, giving attendees hands-on experience with NemoClaw's agent framework. This is where the rubber meets the road — where the keynote's vision gets tested against real developer workflows and enterprise requirements.

Jensen Huang opened GTC 2026 by declaring that we have entered the agentic AI infrastructure era. Based on what shipped today — not what was promised, but what is available as open-source code and production silicon — he may be right. The infrastructure for autonomous AI agents just went from theoretical to tangible. What the industry builds on that infrastructure over the next twelve months will determine whether GTC 2026's keynote joins the short list of technology presentations that actually changed the world.

---

Further Reading:

• NVIDIA's Full-Stack AI Takeover: Pre-GTC Analysis
• NVIDIA Backs Thinking Machines Lab With Gigawatt-Scale Vera Rubin Deal
• Prediction: 50% of Enterprise Agentic AI Pilots Reach Production by Q4 2026