NVIDIA GTC 2026: Jensen Huang Keynote March 16 — Vera Rubin, Feynman Chips, and Why Developers Should Watch

The Most Important AI Hardware Event of 2026 Is in 15 Days

NVIDIA GTC 2026 runs March 16–19 in San Jose, California. The keynote — Jensen Huang live on stage — is confirmed for Monday, March 16 at 8:00 a.m. PDT. Virtual attendance and livestream are free.

This is NVIDIA's primary developer conference. Think of it as Apple WWDC or Google I/O, but for the hardware layer that every AI application in the world actually runs on. The chips announced here will define the infrastructure for model training and inference for the next two to three years.

Jensen Huang has been unusually specific about what is coming. In the buildup to GTC 2026, he stated explicitly: "At GTC 2026, we'll unveil a chip that will surprise the world." That is not standard marketing language for Jensen Huang. When he says something will surprise the world, it is worth paying attention.

Here is what is confirmed, what is credibly expected, and why it matters to you as a developer.

What Is Confirmed: Vera Rubin Deep Dive

NVIDIA announced the Vera Rubin platform at CES 2026 in January. GTC is where the engineering details arrive.

Vera Rubin is NVIDIA's next major compute platform, succeeding Blackwell. It consists of six newly designed chips — Vera CPUs and Rubin GPUs — built specifically for the AI inference and training demands of 2026 and beyond.

The headline numbers are significant:

• 10x reduction in inference token cost compared to Blackwell
• 4x reduction in GPUs needed for training mixture-of-experts models
• First Vera Rubin products arriving in the second half of 2026

The 10x inference cost reduction deserves emphasis. Inference cost is the primary constraint on deploying AI at scale for most companies and developers. A 10x reduction changes the economics of what you can build. Applications that are currently too expensive to run continuously become viable. AI features that require cost justification at every product review become baseline infrastructure.

Blackwell already represented a significant step down in inference costs from Hopper. Vera Rubin continuing that trajectory means the cost curve for AI is still dropping fast — and the applications that become economically viable will continue expanding as it does.

GTC will provide the detailed engineering deep-dive: memory bandwidth, interconnect architecture, power envelopes, software stack compatibility, and the migration path from existing Blackwell deployments. For developers and infrastructure teams, this session is the one to attend or watch.

What Is Expected: The Chip That Will Surprise the World

This is the part of GTC 2026 that nobody knows exactly but everyone is speculating about.

Huang's "surprise the world" framing points toward something beyond Vera Rubin, which was already announced at CES. The leading theory among hardware analysts is a first public reveal of Feynman — NVIDIA's architecture generation after Rubin.

What is known or credibly reported about Feynman:

• Likely built on TSMC's 1.6nm process — the most advanced semiconductor manufacturing node that will exist in this timeframe
• May include LPU (language processing unit) integration via 3D chip stacking — a fundamentally different approach to AI inference acceleration than conventional GPU architecture
• If the LPU integration is real, it would represent a structural shift in how NVIDIA's chips handle transformer model inference, not just an incremental improvement

A first reveal of Feynman at GTC 2026 would mean NVIDIA is showing its roadmap two to three generations out — giving the industry, cloud providers, and hyperscalers visibility to plan infrastructure commitments accordingly. This is strategically important: cloud providers commit to multi-year chip purchases. Seeing the roadmap lets them make those commitments with confidence.

The alternative theory is that the "surprise" is something in the software or systems layer — a new inference framework, a physical AI platform, or a systems architecture — rather than a chip reveal. But Huang specifically said "chip," so Feynman remains the primary expectation.

Physical AI and Robotics: The Theme Running Through GTC

Beyond the chips, GTC 2026 will have a significant focus on what NVIDIA calls physical AI — AI systems that operate in and interact with the physical world.

This includes:

Cosmos: NVIDIA's open-world foundation models, trained on video data and robotics simulation. Cosmos is designed to give robots and autonomous systems a model of how the physical world works — a prerequisite for real-world AI that goes beyond the controlled environments where most robotics AI currently operates.

Newton physics engine: The sim-to-real simulation engine developed jointly with Google DeepMind and Disney Research (which we covered in detail separately). GTC will likely include Newton-related sessions and potentially updates on which robotics companies have integrated it.

Alpamayo: NVIDIA's autonomous vehicle models, representing the automotive AI stack that car manufacturers and mobility companies build on.

The physical AI theme reflects where NVIDIA believes the next wave of AI value creation is. The current wave — foundation models, large language models, image generation — runs primarily in data centres. The next wave runs in factories, hospitals, roads, and public spaces. NVIDIA wants to be the hardware and platform provider for that wave in the same way it became the default for the current one.

Agentic AI: Going Beyond Chatbots

GTC 2026 will include substantial content on agentic AI — systems where AI models take sequences of actions to achieve goals rather than responding to single prompts.

The infrastructure requirements for agentic AI are different from conversational AI. Agents run longer, make more API calls, require tool access, need persistent memory across sessions, and often run in parallel as multi-agent systems. The inference economics and hardware profiles are different as a result.

NVIDIA's interest here is in the infrastructure layer: how do you run agent workloads at scale efficiently? How do you orchestrate multi-agent systems that might involve dozens of models running simultaneously? These are engineering problems that require hardware and software solutions, not just model improvements.

For developers building agentic applications in 2026, the GTC sessions on inference infrastructure for agents are likely to be practically relevant.

How to Watch

The keynote is free to stream. NVIDIA typically puts it on:

• YouTube: search "NVIDIA GTC 2026 keynote"
• NVIDIA's website: nvidia.com/gtc

Time zones for the March 16 keynote:

• PDT (San Jose): 8:00 a.m.
• EDT (New York / US East): 11:00 a.m.
• BST (London): 4:00 p.m.
• IST (India): 8:30 p.m.
• SGT (Singapore): 11:00 p.m.
• AEST (Sydney): 2:00 a.m. March 17

Individual sessions beyond the keynote require free registration at nvidia.com/gtc.

Why This Matters for Developers

Three concrete reasons GTC 2026 should be on your radar:

1. Inference economics determine what you can ship. The Vera Rubin 10x cost reduction will cascade through cloud provider pricing over 18–24 months. Applications you cannot cost-justify today may be straightforward to build and run by mid-2027. Understanding the trajectory helps you plan what to build.

2. The chip roadmap shapes the model landscape. The models researchers train over the next 18 months will be trained on hardware announced at GTC. More capable hardware typically means more capable models. The GTC announcements give you a view into what the model landscape will look like in 2027–2028.

3. Physical AI is the next platform shift. If you believe AI will move from cloud-based software into physical systems — robotics, autonomous vehicles, smart infrastructure — GTC is where NVIDIA lays out the platform it intends to provide for that shift. The same pattern played out with mobile (Apple WWDC 2007), cloud (AWS re:Invent), and AI software (NVIDIA GTC 2023). Being early on a platform shift compounds over time.

The keynote runs about three hours. Jensen Huang's GTC keynotes have a strong track record of being substantive rather than ceremonial. Set a reminder for March 16.