XNews

NVIDIA CES 2026: Rubin chips, Agentic AI & demo highlights

NVIDIA CES 2026: Rubin chips, Agentic AI & demo highlights

NVIDIA unveiled Rubin chips, Agentic AI demos and Alpamayo VLA. - CES stage anecdote: screen outage occurred - Jensen Huang: industry resets every decade - Software shift: program -> train on GPUs - 2025 breakthroughs: inference models, Agentic systems, open-source models - Inference: "test-time scaling" increases answer quality - Agentic: plans, calls tools, executes tasks - Open-source: DeepSeek R1 spurred downloads - Agent superpowers: reasoning, tool use, planning, collaboration - Blueprint demo: DGX Spark assistant managing email/calendar/robot - Privacy routing: local model for emails, frontier models for others - Cursor used internally for agentic coding - Rubin: 100% liquid cooling at 45°C - Rubin: end-to-end encryption - Rubin GPU: 5× FP vs Blackwell - Token cost: Rubin claims 0.1× Blackwell - Vera CPU: 2× perf/watt - NVFP4 Tensor Core: dynamic precision - NVLink-6: 3.6× global internet bandwidth - ConnectX-9: 1.6 Tb/s - BlueField-4 DPU: KV Cache adding 16TB fast-access per GPU - Spectrum-X: 512-channel silicon photonics - Alpamayo 1: 10B VLA chain-of-thought model - Alpamayo dataset: 1,700+ hours driving data - AlpaSim: open driving simulator

Key insights

Rubin architecture and hardware claims

NVIDIA announced the Rubin architecture featuring 100% liquid cooling using 45°C warm water and end-to-end encryption for data in transit, at rest, and in compute. Rubin claims dramatic performance gains: Rubin GPU floating-point throughput is 5× Blackwell, token generation cost is one-tenth of Blackwell for 100T-parameter training workloads, and Vera CPU doubles performance-per-watt with space multithreading. System components include NVLink 6 Switch with bandwidth equaling 3.6× global internet total, ConnectX-9 NIC at 1.6 Tb/s, BlueField-4 DPU providing KV Cache with 16 TB fast-access per GPU, and Spectrum-X switch with 512 silicon-photonic channels.

Alpamayo VLA and open driving stack

NVIDIA introduced the open Alpamayo family and AlpaSim: Alpamayo 1 is a 10B-parameter vision-language-action (VLA) model with chain-of-thought style reasoning for planning safe trajectories and explaining actions. The release includes an open dataset with over 1,700 hours of driving data across varied geographies and rare scenarios, and AlpaSim, an open simulation framework for validating autonomous driving systems.

Agentic AI shift and three 2025 breakthroughs

NVIDIA framed 2025 advances as three interlocking breakthroughs: inference models that perform 'test-time scaling' by spending more reasoning time for higher-quality answers; the explosion of Agentic systems that plan, call tools, and act rather than just answer; and the rise of open-source models like DeepSeek R1 reaching near-state-of-the-art levels. NVIDIA argued these breakthroughs catalyze each other and accelerate real-world adoption.

Agent capabilities and four superpowers

NVIDIA outlined a four-part Agent capability model: (1) reasoning — decomposing unfamiliar problems into solvable steps; (2) tool use — invoking search, calculators, interpreters, databases, and other AI models; (3) planning — formulating multi-step plans and preempting outcomes; (4) collaboration — orchestrating multiple specialized models simultaneously. NVIDIA highlighted multi-model stacks where each model is used for its specialty across the reasoning chain.

Blueprint demo: privacy routing and DGX Spark assistant

NVIDIA demonstrated a 'Blueprint' Agent framework using DGX Spark: a personal assistant that manages email, calendar, to-dos, and a robot. The demo routes private email tasks to a local open-source model and routes other tasks to frontier models via an intent router. Tools were implemented for email, calendar, and camera; the assistant could draft emails, convert sketches to rendered scenes and videos, and identify a pet on camera and command a robot to act — all orchestrated by intent-based model routing.

Implications for software and infrastructure

NVIDIA argued software is transitioning from programmed artifacts to trained, context-aware systems running on GPUs rather than CPUs. Jensen Huang stated industry shifts every decade and called this a dual quake: applications built on AI and a changed software-development approach. He quantified massive infrastructure upgrades, cited investor flows into AI, and suggested enterprise software interfaces will evolve into conversational Agents that translate user natural language into complex workflows.

Source posts