Skip to content
LDS Logo
LDS

Jensen Huang Promised to Surprise the World. GTC Starts Monday.

DS
LDS Team
Let's Data Science
10 minAudio · 3 listens
Listen Along
0:00/ 0:00
AI voice
Four days out, NVIDIA's annual conference is drawing 30,000 attendees from 190 countries to an NBA arena in San Jose — and the CEO says he's carrying chips "the world has never seen before."

It started over Korean fried chicken.

Sitting at a table in Santa Clara with 30 NVIDIA and SK hynix engineers after dinner last month, Jensen Huang pulled out his phone and did what he often does before a big keynote: he set the mood. "We've prepared several new chips the world has never seen before," he told the Korean Economic Daily. "A chip that will surprise the world will be unveiled at GTC."

That was weeks ago. On Monday, March 16, Huang takes the stage at SAP Center — the 17,500-seat arena where the San Jose Sharks play — to deliver the centerpiece keynote of NVIDIA's GPU Technology Conference. It runs through Thursday, March 19. More than 30,000 people from over 190 countries will be in attendance. Hundreds of thousands more will watch the livestream.

GTC has become the AI industry's closest thing to a Super Bowl. And this year, there are more chips on the table than ever.

The SAP Center Is Not an Accident

NVIDIA has held GTC in San Jose before. But holding the keynote in an NBA arena, at the scale of a stadium rock concert, is a deliberate statement.

Last year, when Huang unveiled Blackwell Ultra and put the Vera Rubin roadmap on the wall for the first time, he called GTC "the Super Bowl of AI." It stuck. This year, Fortune called it "the real March Madness." CNBC analysts told investors to buy NVDA ahead of what they called the "AI Super Bowl" conference.

The company is now the most valuable semiconductor firm in history. Every word Huang says on Monday gets parsed by 38 analysts, most of whom hold Buy ratings with an average price target of $273, roughly 50% above where NVDA closed on March 13 at approximately $180.

For context: After GTC 2025, NVDA stock actually fell 3.4% on the day, as investors judged the Blackwell and Vera Rubin announcements as incremental. The pressure on Monday is real: Huang needs to show that the next architecture cycle justifies the expectations baked into the stock.

Vera Rubin Is Already in the Wild

The centerpiece of GTC 2026 is almost certainly Vera Rubin, NVIDIA's next major compute platform, which we covered in depth here when the architecture was first announced.

Vera Rubin launched publicly at CES in January, when Huang confirmed the chips were in full-scale production. NVIDIA has since shipped its first samples to customers. The NVL72 rack — 72 Rubin GPUs, 36 Vera CPUs, built around 288 GB of HBM4 memory per GPU — is now at hyperscalers. AWS, Google Cloud, Microsoft, and Oracle Cloud are listed as early deployment partners. So are CoreWeave, Lambda, and others.

The performance numbers are striking. Each Rubin GPU delivers 50 PFLOPS of inference performance using NVFP4, a 5x improvement over Blackwell GB200. Each 88-core Vera CPU (codenamed Olympus, built on Arm v9.2-A) replaces the Grace CPU from the previous generation. The full NVL72 rack promises 10x lower cost per token at inference compared to Blackwell. For the hyperscalers racing to bring inference costs down, that number matters enormously.

NVIDIA hasn't confirmed exactly what Vera Rubin content appears in Monday's keynote. But analysts expect shipping timelines, pricing signals, and possibly first-look benchmarks from early hyperscaler deployments.

The first Vera Rubin chips shipped in early 2026. Production volume ramp is scheduled for the second half of the year.

The Mystery Chip

Huang's pre-GTC tease wasn't specifically about Vera Rubin. He already announced that at CES. "Chips the world has never seen before" implies something beyond the roadmap.

The leading theory among analysts is Rubin Ultra — a higher-end configuration of the Vera Rubin architecture with 576 GPUs, delivering 14.4x the performance of Grace Blackwell. NVIDIA put Rubin Ultra on the roadmap at GTC 2025 as a 2027 product. If Huang is teasing it a year early, that would be a significant pull-forward signal.

The longer-shot theory is Feynman. According to TrendForce, NVIDIA may offer a first look at Feynman at GTC 2026, even though production isn't expected until 2028. Feynman is planned for TSMC's A16 1.6nm process, the most advanced semiconductor node that TSMC will have ever put into mass production. The architecture is being designed as an "inference-first" chip, built specifically for the long-context, multi-step reasoning requirements of AI agents — not just training big models. If Huang shows even a roadmap slide with Feynman specs on it, that alone moves markets.

On the software side, the picture is clearer.

NemoClaw: NVIDIA Enters the Enterprise Agent Market

Wired broke the story first: NVIDIA has been quietly building NemoClaw, an open-source enterprise AI agent platform, and plans to formally launch it at GTC.

NemoClaw is designed to let companies deploy AI agents that execute multi-step tasks on behalf of employees — processing data, managing workflows, running automations — without depending on any specific cloud vendor or model provider. Unlike OpenAI's or Anthropic's enterprise products, it runs on any hardware. Open source, no lock-in.

NVIDIA has reportedly been pitching early partnerships to Salesforce, Cisco, Google, Adobe, and CrowdStrike. The security and privacy tooling is reportedly built in from the start, which is a direct response to the cluster of enterprise AI agent incidents that have shaken confidence in the category over the past year.

The strategic logic is obvious: if every major company is about to deploy AI agents, NVIDIA wants to be the infrastructure layer those agents run on. NemoClaw positions NVIDIA not just as a chip company but as an enterprise software platform. That's a different multiple.

Nemotron 3 Super Drops the Same Week

NVIDIA didn't wait for GTC to release its new open model. On March 11, three days before the conference, it published Nemotron 3 Super: a 120-billion-parameter hybrid Mamba-Transformer Mixture-of-Experts model with 12 billion parameters active at inference time and a 1-million-token context window.

The architecture is genuinely interesting. Nemotron 3 Super combines Mamba sequence modeling, standard transformer attention, and MoE routing in a single model — using what NVIDIA calls "latent MoE," which compresses tokens before routing them to expert layers, calling 4x as many specialists at the same inference cost. It also uses multi-token prediction to generate multiple tokens per forward pass, enabling built-in speculative decoding without a separate draft model.

NVIDIA claims the model delivers over 5x higher throughput than the previous Nemotron Super. It outperforms GPT-OSS and Qwen on throughput benchmarks. The 1M context window means an agent can hold an entire project's worth of state in memory without losing track of what it was doing 50,000 tokens ago.

GTC sessions this week will likely dig into how Nemotron 3 Super integrates with NemoClaw and how it runs on Vera Rubin hardware.

Physical AI Gets Its Own Days

GTC 2026 includes two dedicated "Physical AI Days" covering robotics, autonomous vehicles, industrial AI, and digital twins.

The highlight on the robotics side is Isaac GR00T N1.6, NVIDIA's vision-language-action model for humanoid robots. Like its autonomous vehicle counterpart Alpamayo 1 — announced at CES in January — GR00T N1.6 applies chain-of-thought reasoning to physical control, allowing robots to reason through novel situations step by step rather than pattern-match against training data.

Disney is presenting a session called "Disney's Robotic Characters: From the Screen to Reality via Physical AI," which is an unusual angle for a chip conference but exactly the kind of cross-industry headline that keeps GTC culturally relevant.

Alpamayo, NVIDIA's open-source autonomous driving model family, is also expected to feature prominently. Alpamayo 1 is a 10-billion-parameter model that teaches autonomous vehicles to reason through rare edge cases — a traffic light outage, an unusual road configuration — rather than rely purely on training coverage. When Jensen Huang announced it at CES, he said the goal was to let an AV "think more like a human."

GTC will show where that program stands today.

Beyond the Chips: The Keynote Format

Huang is not just delivering a product announcement on Monday. He has structured part of the week as a public conversation about where AI is going.

On Wednesday, March 18, he will moderate a panel on open frontier models and what comes next, with Harrison Chase of LangChain, leaders from a16z, AI2, Cursor, and Thinking Machines Lab. For NVIDIA, whose chips power the training and inference of almost every frontier model, demonstrating stewardship of the open-model ecosystem is part of the brand.

On Tuesday, Dario Gil — now Undersecretary at the U.S. Department of Energy — joins NVIDIA VP Ian Buck to discuss AI applications in climate and energy research. That session is notable given the scale of power consumption questions hanging over the industry. Huang published an essay before GTC framing AI as a "5 layer cake" — energy, chips, infrastructure, models, and applications — arguing all five layers must scale together. The energy conversation isn't just philosophical.

A "GTC Live Pregame Show" runs from 8 a.m. to 11 a.m. PT on Monday before the main keynote. The main event begins at 11 a.m. PT at SAP Center. The investor and analyst Q&A follows on Tuesday, March 17, at 9 a.m. PT.

The keynote streams free at nvidia.com.

The Bottom Line

NVIDIA enters GTC 2026 with its stock down roughly 11% from its October 2025 high and the AI infrastructure story demanding a second act. Blackwell delivered on its promises. Vera Rubin is in production. What Monday needs to answer is what comes after the infrastructure build-out: who controls the software stack, what does inference look like at trillion-token scale, and how far has physical AI actually progressed.

Huang's chicken-dinner quote was carefully placed. "Chips the world has never seen before" is a promise, and GTC is where NVIDIA collects on promises.

The keynote runs two hours. If the past is any guide, that's enough time to reshape the AI hardware roadmap for the next three years.

Sources

Practice interview problems based on real data

1,500+ SQL & Python problems across 15 industry datasets — the exact type of data you work with.

Try 250 free problems
Free Career Roadmaps8 PATHS

Step-by-step roadmaps from zero to job-ready — curated courses, salary data, and the exact learning order that gets you hired.

Data Analyst
$95K
Data Scientist
$130K
ML Engineer
$155K
AI Engineer
$160K
Data Engineer
$140K
Analytics Eng.
$140K
MLOps Engineer
$160K
Quant Analyst
$175K
Explore all career paths