NVIDIA GTC 2026 Keynote: Agenda, Timestamps and Where to Start

Updated 2026-10-01

NVIDIA’s GTC 2026 keynote ran on March 16, 2026, at 11 a.m. PT in the SAP Center, San Jose. Jensen Huang spoke for a little over two hours; the official recording is 2:18:56 long. This page is the map: what each segment covers, where to click, and which of our notes go deeper.

Agenda

Start here

1. Tokens per watt, explained. Every hardware claim in the talk is expressed in this unit, so read it first.
2. Token economics and the inference inflection. This is the business case: why NVIDIA thinks inference demand keeps compounding.
3. Vera Rubin platform, then Groq LPX and disaggregated inference. These are the products that are supposed to deliver that case.
4. Roadmap: Rubin Ultra to Feynman. What ships after Vera Rubin.
5. Pick by interest: agents and NemoClaw, open models, physical AI, DLSS 5.

If you want one page that walks the whole keynote in order, use the chapter-by-chapter notes.

What this event was really about

On the surface GTC 2026 was a product launch: Vera Rubin moved to full production and a new rack, Groq 3 LPX, appeared beside it. The more important change was the unit of account. Jensen spent most of the hardware hour on one idea: a data center is a power-limited factory, its output is tokens, and the only number that matters is how many tokens each watt produces at a given response speed. FLOPS barely came up. That framing lets NVIDIA argue that buying its newest system is cheaper than running an older one for free. It also shows where the company thinks competition will come from.

The second theme was specialization inside inference. A year earlier, Dynamo split prefill and decode across different GPUs. This year NVIDIA went further and split the decode step itself: attention stays on Rubin GPUs that hold the large KV cache, and the expert feed-forward layers move to SRAM-heavy Groq LPUs. That is an admission that one chip design cannot be best at both high throughput and very low latency. NVIDIA’s answer is to sell both chips and let its software decide which one does what.

The third theme was demand. The software half of the talk, covering OpenClaw, NemoClaw, the Nemotron coalition and robotaxis, served the hardware half. Agents that run all day, coding tools that burn millions of tokens per engineer, and robots trained on synthetic data all turn into more tokens. Jensen’s line was that NVIDIA is “vertically integrated, horizontally open”. The keynote was built to show both halves of that sentence, and to make the case that each one sells the other.