Jensen Huang GTC 2026 keynote unveiled Vera Rubin’s 5x inference leap, a three-generation GPU roadmap through 2028, NemoClaw AI agents, physical AI factories, and the most significant reset of AI computing economics since the Transformer model era.
Jensen Huang GTC 2026 Keynote: Vera Rubin, Feynman, Physical AI, and Everything NVIDIA Just Announced
Jensen Huang’s GTC 2026 keynote on March 16 in San Jose was the kind of presentation that tends to get quoted and dissected for months afterward, because it was not just a product launch. It was a comprehensive statement about where NVIDIA sees the AI industry heading — and where it plans to sit at the center of all of it.
More than 30,000 people attended the event at the SAP Center, with hundreds of thousands more watching the live stream. Jensen Huang walked on stage in his signature leather jacket, and for the next several hours he laid out what NVIDIA has built, what is coming, and the larger argument for why the company believes it is no longer just a chip maker. By the end of the keynote, the message was clear: NVIDIA is now positioning itself as the full-stack infrastructure layer for the age of AI.
Let us walk through everything that was announced and what it actually means.
=> Jensen Huang just redrew the map for AI’s next three years. Here is everything from GTC 2026 you need to know right now.
Vera Rubin Is Here — And the Numbers Are Staggering
The hardware headline of the entire event was the formal deep-dive into Vera Rubin, NVIDIA’s custom AI accelerator platform and the direct successor to Blackwell. Vera Rubin is now in full production, and the performance figures Huang presented were enough to reframe what AI computing infrastructure looks like from this point forward.
The numbers that matter most: Vera Rubin delivers a 5x inference performance improvement over Blackwell Ultra in FP4 workloads. More significantly, it cuts the cost per inference token by 10x compared to Blackwell — a reduction that NVIDIA and industry analysts are calling the most significant shift in AI compute economics since the Transformer architecture changed everything.
For anyone not deep in the AI infrastructure world, here is why those two numbers matter so much. Training large AI models gets most of the attention, but inference — running those trained models to actually generate answers, code, images, or decisions — is where the day-to-day cost of AI sits. Every time you use ChatGPT, Gemini, Claude, or any AI-powered application, inference is happening. Lowering the cost of inference by 10x does not just make AI applications cheaper to run. It fundamentally changes which AI applications are economically viable to build and deploy at scale
The flagship hardware configuration is the VR200 NVL144 rack system — 144 Vera Rubin GPUs combined with the new Vera CPU, pulling HBM4 memory with over 3 terabytes per second of bandwidth. To put that bandwidth figure in context, it runs about 30% higher than AMD’s comparable MI350 offering.
The partnership announcement that accompanied Vera Rubin was equally significant. NVIDIA and Thinking Machines Lab confirmed a multiyear deal for at least one gigawatt of Vera Rubin system deployment specifically for frontier model training. That is the first confirmed gigawatt-scale deployment of the Vera Rubin platform and a signal that the largest AI research and inference workloads are committing to NVIDIA’s new hardware generation at a scale that would be very difficult to unwind.
The Three-Generation Roadmap: Vera Rubin, Vera Ultra, and Feynman
One of the most strategically notable moments of the keynote was not a single product — it was the roadmap visibility Jensen Huang offered to hyperscalers, enterprise buyers, and investors simultaneously.
NVIDIA laid out three generations of future hardware in a single presentation. Vera Rubin is the current generation, in production now. Vera Ultra arrives in the second half of 2027. Feynman is confirmed for 2028, with silicon photonics as a core architectural feature.
This level of multi-year roadmap transparency is unusual for NVIDIA, and it is not accidental. Large-scale AI infrastructure purchases — the kind made by Microsoft, Amazon, Google, Meta, and the sovereign AI initiatives that governments are now funding — require planning horizons of years, not months. By confirming what comes after Vera Rubin right now, NVIDIA is effectively making the switching cost calculation harder for anyone thinking about diversifying their AI chip stack.
When hyperscalers know that Vera Ultra in 2027 and Feynman in 2028 are on the way — and that both will share the same ecosystem, tooling, and software stack as Vera Rubin — the incentive to stay inside the NVIDIA ecosystem across all three generations becomes very strong. Every investment made in CUDA, NeMo, and NVIDIA’s software layer pays off more with each successive generation.
NemoClaw: NVIDIA Moves Deeper Into Software
Beyond the hardware announcements, Huang introduced NemoClaw, an open-source enterprise AI agent deployment platform. This is a significant move because it extends NVIDIA’s presence from being primarily a chip and hardware company into the software layer that runs on top of that hardware.
NemoClaw is designed to make it straightforward for enterprises to build and deploy AI agents at scale — the kind of always-on, task-completing software agents that handle scheduling, coding, customer service, data analysis, and any other defined workflow without needing human prompting at every step. It integrates with NVIDIA’s existing NeMo framework and is positioned as the software glue that connects NVIDIA’s accelerated compute infrastructure to real-world enterprise AI applications.
The strategic logic here is straightforward but important. If NVIDIA only sells chips, it is always at risk of being displaced by a competing chip that is faster or cheaper. If NVIDIA also controls the software framework that enterprises use to build their AI applications, the relationship becomes much stickier. Moving customers up the stack from hardware buyers to platform users is exactly what Microsoft did with Azure, what Salesforce did with enterprise CRM, and what NVIDIA is now doing with AI infrastructure.
Physical AI: Robots, Cars, and the World Beyond Screens
The section of the keynote that generated some of the most striking visuals was Huang’s presentation on Physical AI — the extension of AI from software running on servers to intelligent systems operating in the real world.
NVIDIA’s GR00T robotics platform was front and center. GR00T is the foundation that NVIDIA provides for humanoid robot development, offering simulation environments, training infrastructure, and pre-built models for physical world navigation and manipulation. The message was that the humanoid robot industry is not years away from real-world deployment — several of NVIDIA’s robotics partners are already using GR00T-based systems in warehouse, logistics, and manufacturing environments.
On the autonomous vehicle side, Huang confirmed that BYD, Hyundai, Nissan, and Geely have all joined NVIDIA’s robotaxi-ready platform. He described the trajectory of robotaxi-ready vehicles with a kind of quiet confidence: “The number of robotaxi-ready cars is going to be incredible.”
NVIDIA’s vision for Physical AI sits inside a broader concept Huang presented through what he calls the AI Layer Cake — a five-layer framework describing the full AI infrastructure stack from energy generation through data centers, chips, software platforms, and ultimately to applications. NVIDIA’s stated goal is to provide the technology across all five layers, not just the GPU tier. That is an enormous scope, and the fact that NVIDIA can plausibly claim progress at each layer right now is a meaningful departure from the company’s position even three years ago.
The N1X: Bringing AI Inference to Personal Computers
Huang also took time during the keynote to introduce the N1X, a new chip designed specifically for AI PC applications. This positions NVIDIA to compete directly in the on-device AI inference market that Qualcomm, Intel, and AMD are all chasing aggressively with their own neural processing unit offerings.
The N1X is built for local AI workloads — the kind of AI tasks that run on a laptop or desktop without needing a cloud connection. With AI features increasingly becoming part of everyday operating systems and applications, on-device inference capability is becoming a standard expectation for new PC hardware. The N1X is NVIDIA’s bid to own a position in that market the same way it has owned the data center GPU market. Whether it gains traction against Qualcomm’s well-established Snapdragon X Elite and Intel’s Meteor Lake NPU will depend on software ecosystem development over the next year.
DLSS 5: The GPT Moment for Graphics
In a moment that landed differently from the data center announcements, Jensen Huang declared that DLSS 5 — a real-time neural rendering technology for games that rewrites lighting and material detail using AI — represents “the GPT moment for graphics.”
DLSS 5 does not just boost frame rates or sharpen resolution like previous DLSS versions. It takes a rendered game frame and uses a neural model to rebuild its lighting and surface detail to look more like photoreal visual effects. The demonstrations in Starfield, Resident Evil Requiem, and Hogwarts Legacy were striking enough that even developers used language you do not usually hear at a hardware announcement — words like “amazing” and “changes what we can promise to players.” DLSS 5 arrives this fall exclusively for RTX 50 series GPUs.
What the Pharmaceutical Sector Adds to the Story
One announcement that deserves more attention than it received in most coverage was the confirmation of a major pharmaceutical AI deployment. Eli Lilly, the global pharmaceutical company, confirmed it is now deploying NVIDIA AI infrastructure for drug discovery and development work.
This matters as a signal more than as a single customer win. The combination of Vera Rubin’s 10x reduction in inference token costs and the expansion of AI into regulated, high-stakes industries like drug development tells you something about where the industry is heading. AI is not just a productivity tool for tech companies. It is being integrated into the research pipelines of industries where the stakes — and the economic value of the outcomes — are orders of magnitude higher.
The Bigger Picture From the SAP Center Stage
If you step back from the individual product announcements and look at what GTC 2026 communicated as a whole, a few things stand out.
First, NVIDIA is now describing itself explicitly as a full-stack AI infrastructure company, not a chip maker. The hardware, the software platforms, the agentic AI tools, the physical AI robotics frameworks, the partner ecosystem of automakers and pharmaceutical companies — all of it is presented as an interconnected stack, with NVIDIA at the foundation and running through every layer.
Second, the three-generation roadmap is a competitive moat being built in real time. By the time Feynman ships in 2028 with silicon photonics, hyperscalers that have built their infrastructure on Vera Rubin will have three generations of investment in NVIDIA’s ecosystem. That is a switching cost that becomes harder to justify crossing with every passing year.
Third, the inference era has officially replaced the training era as the driver of AI infrastructure spending. Vera Rubin’s 10x token cost reduction is specifically calibrated for inference workloads, not training. That tells you exactly where NVIDIA thinks the next phase of AI value creation is concentrated.
FAQ
What did Jensen Huang announce at GTC 2026?
The major announcements included: Vera Rubin in full production with 5x inference performance and 10x lower token costs over Blackwell Ultra, a three-generation GPU roadmap through 2028, the NemoClaw AI agent platform, the N1X AI PC chip, Physical AI expansion including GR00T robotics and new autonomous vehicle partners, DLSS 5 for gaming, the Thinking Machines Lab gigawatt deployment deal, and Eli Lilly’s pharmaceutical AI deployment.
What is NVIDIA Vera Rubin?
Vera Rubin is NVIDIA’s next-generation custom AI accelerator platform, succeeding the Blackwell architecture. It delivers up to 5x better inference performance and 10x lower inference token costs compared to Blackwell Ultra, with HBM4 memory at over 3 TB/s bandwidth.
What is the NVIDIA GPU roadmap after Vera Rubin?
Vera Ultra arrives in the second half of 2027. Feynman is confirmed for 2028, featuring silicon photonics as a core architectural component.
What is NemoClaw?
NemoClaw is NVIDIA’s open-source enterprise AI agent deployment platform, designed to help businesses build and run always-on AI agents for tasks like coding, scheduling, data analysis, and workflow automation.
What is DLSS 5?
DLSS 5 is a neural rendering technology for PC games that uses AI to rewrite the lighting and material detail of rendered frames in real time, producing significantly more photoreal graphics. It arrives this fall, exclusively for RTX 50 series GPUs.
What is Physical AI according to NVIDIA?
Physical AI is NVIDIA’s term for AI systems that operate in the real world — robots, autonomous vehicles, and industrial systems — rather than purely digital environments. NVIDIA’s GR00T platform provides the foundation for humanoid robot development.
GTC 2026 was Jensen Huang at his most comprehensive and his most ambitious. The argument he made — that NVIDIA has built the full-stack infrastructure for the age of AI and that each new generation makes leaving that ecosystem harder — was backed by production hardware, confirmed partner deployments, and a roadmap that gives enterprise buyers clarity through 2028. Whether you are an investor, a developer, a gamer, or simply someone paying attention to where technology is headed, the announcements from San Jose on March 16 are worth understanding fully. The next chapter of AI infrastructure was mapped out on that stage.
=> 10x cheaper AI. Photoreal games. Robots that actually work. NVIDIA’s GTC 2026 keynote changed everything — read the full breakdown.
