Nvidia CEO Jensen Huang CES 2026 Keynote – Next Gen Rueben GPU in Full Production. 5X Blackwell FP

by

Jensen Huang revealed that latest Nvidia technology and applications including the new Ruben GPU chips.

Vera Rubin addresses skyrocketing AI computation demand (models 10x larger yearly, reasoning adds compute).

Post-training uses reinforcement learning and inference is multi-token. Tokens use increase 5x yearly. Costs decline 10x yearly due to race.

NVIDIA advances computation yearly. GB200/GB300 in production and Vera Rubin now in full production.

Vera Rubin Architecture is six chips.
Vera CPU (2x performance),
Reuben GPU (co-designed for low-latency data sharing),
compute board (100 petaflops AI, 5x predecessor).

Connect X9 (1.6 TB/s bandwidth), Bluefield 4 DPU (offloads storage/security), NVLink 6 switch (scales 72 GPUs as one), Spectrum X Ethernet Photonix (512 lanes, 200 Gbit optics for AI factories). 15,000 engineer-years.

First rack: 18 trays, 9 switches, 220T transistors, ~2 tons.

Scale is 1152 GPUs in 16 racks and each rack 72 Reubens (two dies each).

Extreme Co-Design was used. They redesigned all chips simultaneously despite rule against it, due to Moore’s law slowdown. Enables performance beyond transistor limits.

Vera CPU Details are88 cores (176 threads via spatial multi-threading), insane IO/data rates, 2x performance/watt vs. advanced CPUs.

Reuben GPU Details

5x Blackwell in FP performance.
1.6x transistors. NVFP4 tensor core: adaptive precision processor for transformers, revolutionary (potential industry standard).

MGX Chassis Revolution. Zero cables/hoses/fans. Assembly 2 hours to 5 minutes. 100% liquid-cooled.

Networking

Connect X9 (world’s best NIC, programmable RDMA).
Spectrum X (AI Ethernet, made NVIDIA largest networking company, 25% higher throughput worth billions).

Bluefield 4: is for virtualization/security/networking. New for KV cache (AI working memory) to handle growing contexts off-HBM.

Rack Design. Compute nodes (NVLink 72), context memory (150TB/rack, 16TB/GPU additional), management plane (Spectrum X switches).

Key Features. Twice energy-efficient (45°C water, no chillers, saves 6% global data center power); confidential computing (all encrypted). power smoothing (avoids 25% overprovisioning); performance: 3.5x training, 5x inference vs. prior.

Performance Charts: For 10T parameter model on 100T tokens: 1/4 systems for 1-month training; 10x factory throughput (revenue impact). Token cost 1/10th.

NVIDIA’s Role: Builds full-stack AI (chips to applications) for others to create apps.

Advancements in Digital Biology and Other Domains

NVIDIA’s work in proteins and digital biology enabled tools like Proteina for synthesizing and generating proteins, OpenFold 3 for understanding protein structures, and EVO 2 for understanding and generating multiple proteins, representing the beginnings of cellular representation. They also developed Earth 2 AI for understanding laws of physics.

Open Sourcing Models and Libraries

All models are developed transparently and open-sourced, allowing derivatives.
NVIDIA provides a suite of libraries.
Nemo (general AI lifecycle management)
Physics Nemo
Claronemo, and Bioneo.

These handle data processing, generation, training, creation, evaluation, guardrailing, and deployment of AI models—all open-sourced and complex.

NVIDIA as Frontier AI Builder

NVIDIA positions itself as a frontier AI model builder, creating models openly to enable companies, industries, and countries to participate in the AI revolution. Huang highlights NVIDIA’s unmatched contributions to the industry, with plans to accelerate further. Models are world-class, topping leaderboards in intelligence, multimodality (PDFs), speech recognition, retrieval, and semantic search.

AI Agents and Reasoning Capabilities

These models enable building AI agents. Huang notes that early models like ChatGPT hallucinated due to lack of real-time grounding. Now, agents can reason: decide on research, use tools, break problems into steps, and compose sequences for novel tasks. This mimics human reasoning from past experiences.

Breakthroughs in Reinforcement Learning and Multi-Model Use

Agents use reinforcement learning, chain of thoughts, search, and planning—all open-sourced. A key breakthrough, inspired by Perplexity, is using multiple models simultaneously for tasks, making AI multi-modal (speech, images, text, videos, 3D, proteins) and multi-model (best-fit models). It’s also multi-cloud and hybrid cloud for various applications.

Framework for Future Applications

Future applications are built on agentic AIs
multi-model, customizable with open tools like Neimotron and Nemo.

Users can teach unique skills. A smart router (manager) directs prompts to the best model based on intent.

Benefits of the Architecture

Results in customizable AIs with domain expertise, always at the frontier. NVIDIA provides “blueprints” integrated into enterprise SaaS platforms worldwide.

Demo of Personal Assistant Blueprint

Huang demonstrates building a personal assistant using Brev to turn DGX Spark into a personal cloud. It handles calendar, emails, to-do lists, home monitoring. Integrates local open models for privacy, intent-based router, Hugging Face’s Reachi mini robot for interaction, and 11 Labs for voice. The demo shows tasks like checking to-do lists, sending emails, rendering sketches, creating videos, and monitoring pets—all trivial now but impossible years ago.

Architecture of Modern AI Applications

Combines pre-trained proprietary frontier models with customized ones in an agentic reasoning framework, accessing tools, files, and other agents. Enables fast creation of applications that reason through novel data.

Integration with Enterprise Platforms

NVIDIA works with leading platforms: Palantir (AI and data processing), ServiceNow (customer/employee service), Snowflake (cloud data), Code Rabbit (used internally at NVIDIA), CrowdStrike (AI threat detection), NetApp (semantic AI and agentic systems for customer service). Agentic systems become the user interface, replacing traditional tools like Excel or command lines for multimodal interaction.

NVIDIA Physical AI – Reasoning self driving cars and robots

Physical AI moves intelligence from screens/speakers to interacting with the world, understanding common sense (object permanence, causality, friction, gravity, inertia).

System for Physical AI

Requires teaching AI physics laws from scarce data, evaluating via simulation.

Involves three computers
– training AI models,
inferencing (robotics edge computing), and
simulation (NVIDIA’s strength).

Stacks for Physical AI

Omniverse (digital twin simulation),
Cosmos (foundation model of the world, aligned with language),
robotics models like Groot and Alpha Mayo.

Synthetic Data Generation Breakthroughs

Key revelation is the need transform compute into data using physics-grounded synthetic generation. Cosmos generates data from simulators for training.

Cosmos Capabilities Demo

Cosmos is an open frontier world foundation model, pre-trained on videos, driving/robotics data, and 3D simulation. Aligns language, images, 3D, actions. Performs generation, reasoning, trajectory prediction. Generates realistic videos from 3D scenes, motion from telemetry, surround video from simulators/prompts. Enables interactive simulations, reasons about edge cases. Turns compute into data for AVs and robots.

Cosmos Usage and Impact

Downloaded millions of times, used worldwide for physical AI. NVIDIA uses it for self-driving cars (scenario generation, evaluation), simulating trillions of miles virtually.

Announcement of Alpha Mayo World’s first thinking/reasoning autonomous vehicle AI

Trained end-to-end (camera to actuation) using human demonstrations and Cosmos-generated data. Hundreds of thousands of labeled examples. Reasons about actions, explains them, and plans trajectories—all coupled.

Alpha Mayo Performance

Drives naturally like humans. Reasons in every scenario. Handles long-tail events by decomposing into normal sub-scenarios.

Self-Driving Demo

One-shot, no-hands demo shows routing, buckling up, navigating traffic, arriving safely.

NVIDIA’s AV Journey

Started 8 years ago to reinvent computing stack. AV is a five-layer cake: land/power/shell (car), chips (GPUs/networking/CPUs), infrastructure (Omniverse/Cosmos), models (Alpha Mayo, open-sourced), application (e.g., Mercedes-Benz).

Inflection Point for AVs

Next 10 years, large percentage of cars autonomous. Technique applies to all robotics (manipulators, mobile, humanoid).

Robotics Ecosystem

Partners like Neurorobot, Agibot, LG, Caterpillar, Surf Robot (Uber Eats delivery), Agility, Boston Dynamics, surgical/manipulator robots (Franka, Universal Robotics).

Revolutionizing Design Industries

Integrating NVIDIA tech (CUDA-X, AI physics) into EDA/SDA tools from Cadence, Synopsys, Siemens. Enables agentic designers, simulates chips/systems/robots/factories virtually.

Siemens Partnership Announcement

Integrating CUDA-X, AI models, Omniverse into EDA/CAE/digital twins. Covers industrial lifecycle: design, simulation, production, operations. Addresses labor shortages with physical AI/automation for factories (chips, computers, drugs, AI).

Model Landscape
OpenAI leads in tokens. Open-source models will dominate due to diversity.