Jensen Huang walked on stage at the SAP Center in San Jose wearing the same leather jacket he always wears, and proceeded to spend two hours reframing what the next decade of computing is going to look like. I've watched a lot of keynotes in thirty years of technology. Most of them are product announcements dressed up as vision. GTC 2026 was vision dressed up as product announcements — and the distinction matters.
The headline numbers are staggering enough on their own. NVIDIA revealed it sees at least $1 trillion in visible order flow for its Blackwell and Vera Rubin platforms through 2027 — roughly double the projections from the prior year. The AI capital expenditure boom that everyone has been waiting to see slow down is not slowing down. If anything, the demand signal is getting louder.
But the numbers aren't the story. The story is what those numbers are being spent to build.
Vera Rubin: a system, not a chip
The centerpiece of the keynote was the unveiling of Vera Rubin — NVIDIA's next-generation platform, described not as a GPU but as a full-stack computing system purpose-built for agentic AI. Seven chips, five rack-scale systems, one supercomputer. The Vera CPU, the Rubin GPU, NVLink 6, ConnectX-9, BlueField-4, Spectrum-6 Ethernet, and the newly integrated Groq 3 LPU — all vertically integrated, optimized as a single system.
The Groq integration is worth pausing on. NVIDIA acquired Groq for $20 billion late last year — its largest deal ever — and the Groq 3 LPU made its debut at GTC integrated directly into the Vera Rubin rack. The LPU is purpose-built for inference, handling token generation at speeds that GPU architectures weren't designed for. In agentic scenarios — where a model isn't answering a single question but executing a multi-step workflow — the throughput gains are up to 35 times over prior generation systems. That's not an incremental improvement. That's a different category of performance.
Huang was explicit about the framing: NVIDIA isn't building chips anymore. It's building the operating system for the intelligent age.
The agentic layer is becoming infrastructure
The software announcements at GTC were as significant as the hardware. NVIDIA launched Dynamo 1.0, described as an AI inference operating system — software that manages the compute, routing, and orchestration of AI workloads across an entire data center the way an operating system manages resources on a single machine. It also unveiled NemoClaw, an enterprise agent platform — effectively a governed, secure stack for deploying AI agents at scale within organizational environments.
The framing that landed for me: companies will deploy AI agents powered by Vera Rubin the same way they deploy software today. The agent isn't a feature inside an application anymore. The agent is the application.
This matters for every company thinking about AI adoption right now. The infrastructure that makes enterprise agentic deployment safe, auditable, and scalable is arriving. The question shifts from "is agentic AI ready for enterprise?" to "how fast can your organization design the workflows to take advantage of it?"
The physical AI moment
The part of the keynote that I found most viscerally striking wasn't the chip announcements — it was the robots.
NVIDIA's GR00T N1.7 humanoid robot, trained entirely in simulation, walked successfully in the real world on its first attempt. Jensen called it on stage and it delivered. Disney brought an Olaf robot — trained in an NVIDIA simulation jointly developed with Disney — that walked and talked with Huang in front of the crowd. New automotive partners were announced: BYD, Hyundai, Nissan, Geely. Huang declared, with characteristic confidence, that "the ChatGPT moment for autonomous driving has arrived."
I've been skeptical of robotics timelines for most of my career. The gap between laboratory performance and real-world deployment has historically been measured in decades, not years. What I saw at GTC suggests that gap is closing faster than most operating models assume. Simulation-trained physical AI that works on first deployment in the real world is a qualitatively different capability than anything we've seen before.
I'm not predicting a specific timeline. I am saying that the leaders of the companies I work with need to start thinking about physical AI as a medium-term operational reality, not a science fiction scenario.
What I keep coming back to
After two hours of announcements — the trillion-dollar order book, the Vera Rubin platform, Groq integration, NemoClaw, Dynamo, the robots, the orbital data center preview — the thing I keep coming back to is the cumulative weight of it.
CUDA turned 20 this year. For two decades, NVIDIA built the foundational layer of modern AI without most of the world paying attention. Then the world paid attention all at once, and NVIDIA found itself in the position of being the company whose infrastructure every major AI capability runs on. GTC 2026 was Jensen making the case that this is not the peak — it's the inflection before the next phase.
The $1 trillion order book isn't a prediction. It's a measurement of committed demand that already exists. The physical AI demonstrations aren't concept cars. They're systems shipping to customers this year.
I've spent thirty years watching technology waves arrive. Some of them were real. Some were overhyped. Some arrived slower than expected and some faster. What I observed at GTC 2026 is a convergence of infrastructure, software, and physical capability that doesn't feel like hype. It feels like the kind of moment you want to understand clearly before it's already behind you.
Sources: NVIDIA Newsroom, CNBC, Data Center Knowledge, Tom's Hardware, Medium — March 16–20, 2026.