NVIDIA founder and CEO Jensen Huang delivered a sweeping keynote in Taipei on Sunday, rolling out a new generation of AI hardware, open-source models and robotics infrastructure — all oriented around a single strategic bet: that the next era of AI will be defined not by chatbots, but by autonomous systems that observe, plan and act in the world.
The address, held at the Taipei Popular Music Center during GTC Taiwan 2026, amounted to one of NVIDIA's most comprehensive product announcements in recent years.
Vera Rubin: Built for the Age of AI Agents
The headline announcement was Vera Rubin, NVIDIA's next-generation AI computing platform, which Huang confirmed has entered full mass production. The supply chain supporting it, he said, is twice the scale of what the company built for its predecessor, Grace Blackwell.
Huang was emphatic that Vera Rubin is not a chip or a GPU — it is a complete multi-rack, pod-scale AI supercomputer, integrating GPUs, CPUs, networking, storage, security processors and liquid-cooled cabinet infrastructure. Where Grace Blackwell was engineered primarily for AI inference, Vera Rubin is purpose-built for agentic AI: systems that must manage large context windows, working and long-term memory, tool calls and coordination across multiple sub-agents simultaneously.
Vera CPU: Rethinking the Processor for Nanosecond-Scale AI
Alongside Vera Rubin, Huang introduced the Vera CPU, framing it as a fundamental rethink of what a processor needs to do when its primary user is no longer a human. Conventional CPUs were designed around human latency — interactions measured in seconds. AI agents operate at nanosecond timescales, and even minor delays translate directly into lost GPU utilization and reduced token throughput.
The Vera CPU prioritizes low latency, high single-thread performance, wide memory bandwidth and energy efficiency. As AI agents increasingly call external tools, query databases and orchestrate complex workflows, Huang argued that the CPU will become one of the most consequential variables inside any AI factory.
Open Models: Nemotron 3 Ultra, Cosmos 3 and Alpamayo 2
On the software side, NVIDIA launched Nemotron 3 Ultra, an open-source model released alongside its training data and methodology — allowing enterprises to build custom AI agents on top of a shared foundation. Huang said the model runs five times faster than its predecessor at 30% lower operating cost.
For physical AI, NVIDIA introduced Cosmos 3, a foundation model aimed at robots, autonomous vehicles and industrial equipment. Huang noted a structural challenge: most large language models are trained on data captured from a human vantage point, while machines operating in the physical world need to understand their environment from an entirely different perspective — one where training data is inherently scarce. Cosmos 3 is designed to help machines simulate and reason about real-world conditions.
Rounding out the model announcements, NVIDIA unveiled Alpamayo 2, an open autonomous driving model built around reasoning. Huang said it allows vehicles to move beyond recognizing objects on a road, instead enabling them to interpret traffic situations, evaluate options and articulate the logic behind their decisions.
Isaac Groot Opens the Humanoid Robotics Stack
Perhaps the most forward-looking announcement was Isaac Groot, an open development platform for humanoid robots. Huang acknowledged that building humanoid robots remains extraordinarily difficult — research teams typically spend enormous time assembling simulators, teleoperation systems, data pipelines and training infrastructure from scratch before any real development can begin.
Isaac Groot consolidates open models, simulation and training libraries, data generation tools and the Jetson Thor computing platform into a unified development environment. NVIDIA demonstrated a reference-design robot built on the platform during the keynote.
A Coherent Strategic Direction
Taken together, the six product lines announced Sunday — Vera Rubin, Vera CPU, Nemotron 3 Ultra, Cosmos 3, Alpamayo 2 and Isaac Groot — trace a clear strategic line. NVIDIA is positioning itself not merely as a GPU supplier for AI training, but as the full-stack infrastructure provider for a phase of AI development in which systems no longer just answer questions, but take actions, use tools and operate in the physical world.
Huang also introduced Isaac Groot, an open platform for humanoid robot development. He acknowledged that building humanoid robots is extraordinarily difficult, with research teams typically forced to assemble simulators, teleoperation systems, data pipelines and training infrastructure from scratch.
Isaac Groot integrates open models, simulation and training libraries, data generation tools and the Jetson Thor computing platform into a single development environment, Huang said, with the goal of allowing research institutions and developers to accelerate humanoid robotics work. NVIDIA simultaneously demonstrated a reference-design robot built on the platform.
Taken together, the products announced Sunday — Vera Rubin, Vera CPU, Nemotron 3 Ultra, Cosmos 3, Alpamayo 2 and Isaac Groot — trace a coherent strategic line: NVIDIA is repositioning itself for a phase of AI development in which systems no longer just converse, but act, use tools and operate in the physical world. （Related： NVIDIA Declares the Age of Physical AI Has Begun ｜ Latest ）
