NVIDIA GTC 2026 Recap: Keynote, Announcements, and Verdict

TL;DR: NVIDIA GTC 2026 ran March 16-19 at the 17,000-seat SAP Center in San Jose. Jensen Huang unveiled the Vera Rubin computing platform (NVIDIA’s next-gen AI infrastructure stack), agentic AI framework OpenClaw, a six-family Nemotron open-model coalition, and a $1 trillion revenue outlook through 2027. Pre-event projections put attendance north of 30,000 from 190-plus countries, and in-person passes sold out before the show started.

I track NVIDIA GTC every year for one reason: what gets announced on that stage in March tends to show up as shipping product inside the AI tools I review six to twelve months later. Vera Rubin, the AI infrastructure platform NVIDIA unveiled at GTC 2026, is the kind of announcement that reshapes the pricing and speed of the AI SaaS tools most readers here actually touch, even though almost none of them will ever touch a GPU rack directly. GTC sits at the top of my AI events hub for that reason.

This is a recap, not a should-you-attend guide. What NVIDIA said, what shipped, how many people showed up, and an honest verdict on whether the trip (or the free livestream) was worth it. Every claim below is sourced. Where NVIDIA’s own numbers do not fully agree with each other, I say so.

GTC 2026 turned downtown San Jose into an AI campus

NVIDIA GTC 2026 ran March 16-19, 2026 in San Jose, California, spanning 10 venues across the city and anchored by the San Jose McEnery Convention Center (checked 2026-08-24, blogs.nvidia.com/blog/gtc-2026-news/). Huang’s keynote came from a different location entirely: the SAP Center, San Jose’s 17,000-seat hockey arena, the same venue GTC used in 2025 to accommodate keynote demand that had outgrown the convention center’s own auditorium.

The keynote opened Monday, March 16 at 11 a.m. PT, preceded by a GTC Live pregame show at 8 a.m. PT featuring venture investors Sarah Guo (Conviction), Gavin Baker (Atreides Management), and Alfred Lin (Sequoia Capital) interviewing AI-native CEOs including Perplexity’s Aravind Srinivas, LangChain’s Harrison Chase, and Mistral AI’s Arthur Mensch. NVIDIA also ran a separate financial analyst Q&A for investors on Tuesday, March 17 at 9 a.m. PT.

That financial-analyst detail matters more than it looks. A developer conference does not usually need a dedicated investor session the day after the keynote. GTC has become a hybrid event that serves NVIDIA’s stock story and its developer ecosystem simultaneously, and that dual purpose shapes what gets announced on stage: big platform names, trillion-dollar demand figures, and splashy demos alongside the technical substance.

Detail	Info
Event	NVIDIA GTC 2026
Dates	March 16-19, 2026
Keynote venue	SAP Center, San Jose, CA (17,000 seats)
Conference footprint	10 venues across downtown San Jose
Format	Hybrid, in-person plus free virtual keynote livestream
Scale	1,000+ sessions, 2,000+ speakers, 450+ sponsors, 240+ Inception startups
Keynote	Jensen Huang, NVIDIA founder and CEO
On-demand replay	Free at nvidia.com/gtc/keynote, no registration required

Vera Rubin was the biggest hardware bet of the year

Vera Rubin was the headline hardware announcement. The platform packages seven chip types (a new Vera CPU, Rubin GPUs, NVLink 6 switches, ConnectX-9 NICs, BlueField-4 DPUs, Spectrum-X optical NICs, and a Groq 3 LPU) into five rack-scale systems that function as a single AI supercomputer. NVIDIA’s own specs: 3.6 exaflops of compute, 260 terabytes per second of all-to-all NVLink bandwidth, fully liquid-cooled at 45 degrees Celsius, with rack installation time cut from two days to two hours. The platform is in production now, not a future promise (checked 2026-08-24, nvidianews.nvidia.com/news/nvidia-vera-rubin-platform).

Huang opened his two-hour keynote by framing the “token” as the fundamental unit of modern AI, then moved into what NVIDIA has branded the “five-layer cake”: energy, chips, infrastructure, models, and applications. Vera Rubin sits at the chips-plus-infrastructure layer.

The Groq acquisition NVIDIA quietly closed in late 2025 paid off here. The Groq 3 LPU handles the decode phase of AI inference where token-generation speed is the bottleneck. Vera Rubin handles the prefill phase. A software layer called NVIDIA Dynamo splits the work. Independent analysis from SemiAnalysis, cited in Atlan’s technical recap of the keynote, put the combined throughput at roughly 35x per megawatt versus Blackwell alone and roughly 50x more tokens per watt versus the prior-generation Hopper H200. That kind of independent verification is worth flagging specifically: vendor keynote claims should never be taken at face value, and in this case a third party checked the math and the numbers held up.

OpenClaw is NVIDIA’s real bet on agentic AI

I care most about the OpenClaw and NemoClaw announcements because that is the layer of the AI stack that shows up in the SaaS tools I test six to twelve months later. OpenClaw is an open-source agentic AI framework built by independent developer Peter Steinberger, and Huang called it “the most popular open source project in the history of humanity” on stage. NVIDIA announced platform-wide support plus NemoClaw, an enterprise-hardened stack that adds policy enforcement, network guardrails, and privacy routing so companies can deploy always-on AI agents without losing control of what those agents can touch.

If you build or evaluate AI agent tooling, this is worth a closer look. My breakdown of the best MCP servers covers the adjacent agent-infrastructure ecosystem this kind of endorsement tends to feed into.

The Nemotron Coalition, physical AI, and DLSS 5

NVIDIA expanded its open-model portfolio into a formal Nemotron Coalition spanning six frontier model families: Nemotron (language and reasoning), Cosmos (world and vision models), Isaac GR00T (general-purpose robotics), Alpamayo (autonomous driving), BioNeMo (biology and chemistry), and Earth-2 (weather and climate modeling). NVIDIA is positioning itself as an open-model backer across nearly every AI application category at once, extending the chip business into model plumbing.

Physical AI got real production news. NVIDIA’s robotaxi-ready DRIVE Hyperion platform picked up BYD, Hyundai, Nissan, and Geely as automaker partners, plus a deployment partnership with Uber to put those vehicles into its ride-hailing network. The IGX Thor edge-AI platform went generally available, with named enterprise adopters including Caterpillar, Hitachi Rail, Johnson & Johnson, and Medtronic building real-time physical AI into construction equipment, rail inspection, and surgical robotics.

Huang also previewed the next architecture generation, Feynman, which includes a new GPU, an LP40 LPU built jointly with the former Groq team, a Rosa CPU (named for Rosalind Franklin), and a next-generation BlueField-5 networking chip. NVIDIA also confirmed it is building toward orbital AI data centers under the name NVIDIA Space-1 Vera Rubin, a genuinely unusual reveal that got less press coverage than it probably deserved. On the graphics side, GTC’s original reason for existing before AI took over, NVIDIA announced DLSS 5, an AI-powered neural rendering upgrade that hits real-time photoreal 4K rendering on local consumer hardware.

The $1 trillion figure needs a caveat

I would treat Huang’s $1 trillion figure the way I treat any company’s own multi-year revenue projection from its own CEO on its own stage: directionally informative, not a guarantee. He told the keynote audience he expects roughly $1 trillion in cumulative NVIDIA revenue from 2025 through 2027, driven by what he framed as an “inference inflection,” a shift from training-heavy AI workloads to inference-heavy ones. The technical announcements around Vera Rubin and OpenClaw are the parts I would bet on. The trillion-dollar number is doing double duty as a stock-story talking point.

Attendance came in two flavors that do not fully agree

NVIDIA’s March 3, 2026 pre-event press release projected “more than 30,000 attendees” from over 190 countries across developers, researchers, business leaders, and AI-native companies (checked 2026-08-24, nvidianews.nvidia.com). That figure covers the full event footprint including virtual attendance. Separately, NVIDIA’s own social announcement confirming that in-person passes had sold out cited “over 20,000 attendees” specifically for the in-person San Jose show.

Both numbers come from NVIDIA directly. A smaller in-person crowd inside a larger global registrant count is a normal pattern for a hybrid event. What is genuinely confirmed, rather than projected, is that passes sold out before the event started. NVIDIA has not published a single official post-event headcount. I would rather flag that plainly than round one of these numbers into something more precise-sounding than it is.

Who should have shown up in person

I would have flown in if I ran GPU procurement or shipped physical AI. For infrastructure engineers, robotics and physical-AI teams, and enterprise buyers making multi-year GPU commitments, GTC 2026 delivered. Vera Rubin’s specs are real and independently verified. The OpenClaw endorsement is the kind of platform-level bet that tends to actually move enterprise adoption, not generate a press cycle.

For everyone else, the honest verdict is that the free on-demand keynote replay covers essentially everything that mattered. NVIDIA does not gate the keynote behind registration. There was no in-person-only announcement this year that changes that calculus. If your job does not involve GPU procurement, chip architecture, or physical AI deployment, the free replay gets you the same information the $2,000-plus in-person pass did, minus the sold-out-arena atmosphere and the Denny’s pancakes NVIDIA has historically handed out before the keynote.

NVIDIA GTC 2027 is confirmed for March 14-18, 2027 at the San Jose McEnery Convention Center. Full details in the NVIDIA GTC 2027 guide.