explainx / blog
The definitive guide to NVIDIA Computex 2026. Every announcement from Jensen Huang's keynote: Nemotron 3 Ultra (550B parameters), Cosmos 3 Physical AI omnimodel, RTX Spark superchip, DGX Station, and 25+ major updates.
Jun 17, 2026
The NVIDIA DGX Spark turns a $4,679 personal machine into a 200B-parameter inference engine — with 128GB unified memory, full CUDA support, and an electricity bill that replaces a $1,000/month cloud subscription. Here is the complete 2026 buyer's guide.
Jun 16, 2026
GPU prices have surged 40–60% because AI data centers are buying every chip NVIDIA can make. An RTX 5090 now costs $2,000+. GeForce Now Ultimate gives you a 4090-class GPU for $20/month with DLSS 4 AI upscaling — no hardware to buy, no drivers to update, and it runs on a Mac, Chromebook, or phone. This guide covers everything you need to know about cloud gaming in 2026.
May 30, 2026
With cryptic coordinates pointing to Taipei, NVIDIA, Microsoft, and ARM have teased what could be the biggest disruption to PC computing in decades: NVIDIA's N1X ARM processor with 20 CPU cores and RTX 5070-class graphics. As Jensen Huang prepares for his Computex 2026 keynote, the PC industry faces its first real challenge to x86 dominance since the 1980s.
TL;DR: Jensen Huang transformed NVIDIA's image at Computex 2026, positioning the company not just as a chipmaker but as a full-stack AI platform. Nemotron 3 Ultra (550B parameters, 55B active) tops US open-weights rankings with an Intelligence Index of 48, delivering 300+ tokens/second. Cosmos 3 becomes the world's first open Physical AI omnimodel, ranking #1 across 7+ robotics benchmarks. RTX Spark reinvents Windows PCs with Grace+Blackwell architecture and 128GB unified memory. DGX Station brings trillion-parameter models to desktops. Plus: Vera CPUs, DLSS 4.5, Agent Toolkit, and Nemotron 4 preview. Here's everything.
NVIDIA CEO Jensen Huang opened the Computex 2026 keynote at Taipei Music Center on June 1, 2026, with a transformational message: NVIDIA is no longer just a chip company—it's a full-stack AI platform company.
The numbers validate this shift:
This wasn't another GPU launch. NVIDIA is fundamentally repositioning from semiconductor vendor to AI intelligence infrastructure provider, with Computex 2026 marking the official transition.
The headline announcement. Nemotron 3 Ultra is NVIDIA's largest and most capable open-weights model to date, designed specifically for agentic AI workloads.
| Specification | Details |
|---|---|
| Total Parameters | 550 billion |
| Active Parameters | 55 billion per token |
| Architecture | Hybrid Mamba-Transformer MoE |
| Context Length | Up to 1 million tokens |
| Training Precision | NVFP4 (4-bit) on Blackwell |
| Special Features | LatentMoE, Multi-Token Prediction |
Hybrid Mamba-Transformer MoE:
LatentMoE Architecture:
Multi-Token Prediction (MTP):
NVFP4 Training:
Nemotron 3 Ultra tops US open-weights rankings across key metrics:
| Benchmark | Nemotron 3 Ultra | GPT-OSS-120B | Qwen3.5-122B | Notes |
|---|---|---|---|---|
| Intelligence Index | 48.0 | 46.2 | 45.8 | Artificial Analysis composite score |
| HumanEval (Coding) | 92.1% | 87.3% | 88.6% | Code generation accuracy |
| MMLU | 89.4% | 87.1% | 86.9% | Multitask language understanding |
| RULER (256K) | 94.2% | 88.7% | 89.1% | Long-context retrieval |
| Output Speed | 300+ tps | 135 tps | 40 tps | Tokens per second (8K input/16K output) |
Key advantages:
Nemotron 3 Ultra occupies the "most attractive quadrant" with both high intelligence (48.0 Index) and exceptional output speed (300+ tokens/second). Chart courtesy of Artificial Analysis.
Sources: Artificial Analysis, Crypto Briefing, DataCamp
Nemotron 3 Ultra is purpose-built for autonomous agent workloads:
Agent Productivity:
Use Cases:
Multi-Environment RL Post-Training:
Inference-Time Budget Control:
Given its scale, Nemotron 3 Ultra requires substantial infrastructure:
Sources: NVIDIA Nemotron Research, AI CERTs News
Ultra joins Nano and Super in the complete Nemotron 3 lineup:
| Model | Parameters | Active | Context | Best For | Availability |
|---|---|---|---|---|---|
| Nano | 31.6B | 3.2B | 1M tokens | Edge devices, real-time inference | ✅ Available now |
| Super | 120B | 12B | 1M tokens | Enterprise agents, high-volume workloads | ✅ Available now |
| Ultra | 550B | 55B | 1M tokens | Strategic reasoning, deep analysis | 🔜 Q2-Q3 2026 |
Nemotron 3 Nano:
Nemotron 3 Super:
Nemotron 3 Ultra:
Sources: NVIDIA Blog - Nemotron 3 Super, NVIDIA Developer
NVIDIA unveiled Cosmos 3, the world's first fully open omnimodel combining native vision reasoning with world and action generation - purpose-built for Physical AI and robotics.
Unlike language models that work with text or vision models that process images, Cosmos 3 unifies multiple AI capabilities in a single model:
Release: Super (32B parameters) and Nano (8B parameters) variants available now.
Cosmos 3 introduces a novel Mixture-of-Towers architecture that separates reasoning from generation:
| Component | Function | Technology |
|---|---|---|
| Reasoner Tower | Physical understanding, spatial relationships, motion prediction | Autoregressive Transformer |
| Generator Tower | High-quality video synthesis, controlled scene generation | Diffusion model |
Why this matters: Previous models separated world generation, physical understanding, and scene control into different systems. Cosmos 3 unifies all three, enabling the model to think before it acts - crucial for robotics and embodied AI.
Cosmos 3 ranks first among open models across major Physical AI benchmarks:
| Benchmark | Rank | What It Measures |
|---|---|---|
| Physics-IQ | 🥇 #1 | Physical reasoning and common sense |
| PAI-Bench | 🥇 #1 | Physical AI understanding |
| R-Bench | 🥇 #1 | World generation accuracy |
| RoboLab | 🥇 #1 | Robot action policies |
| RoboArena | 🥇 #1 | Multi-step robotic tasks |
| VANTAGE-Bench | 🥇 #1 | Vision understanding for robotics |
| TAR Leaderboard | 🥇 #1 | Vision reasoning |
Artificial Analysis ranking: Tops open models for Physical AI capabilities.
Sources: NVIDIA Newsroom, NVIDIA Blog - Cosmos 3
1. Vision Reasoning
Cosmos 3 understands:
2. World Simulation
3. Robot Policy Training
4. Image-to-Video Generation
Cosmos 3 excels at controlled video generation from single images:
Example use case (from NVIDIA demo):
Input image prompt:
"Generate a 16:9 image from a dashcam view of a formula 1 racing event"
Video prompt:
"A high-speed racing event where a car navigates multiple winding turns"
Output:
→ Realistic 16-second Formula 1 race video
→ Proper motion blur, camera shake, environmental audio
→ Physics-accurate vehicle dynamics
→ Consistent lighting and weather
Sound generation: Cosmos 3 also generates ambient audio matching the physical scene (engine sounds, wind, tire screeches in the F1 example).
Cosmos 3 was trained on one of the largest multimodal Physical AI datasets:
This massive scale enables the model to understand physical interactions it has never directly experienced.
Autonomous Vehicles:
Robotics:
Gaming & Simulation:
Scientific Simulation:
Film & Content Creation:
| Model | Parameters | Best For | Availability |
|---|---|---|---|
| Cosmos 3 Nano | 8B | Edge devices, real-time robotics, mobile deployment | ✅ Available now |
| Cosmos 3 Super | 32B | High-quality simulation, content creation, research | ✅ Available now |
Open release: Both variants released as open models under permissive license, continuing NVIDIA's commitment to open Physical AI.
Unified Multimodal Learning:
Physics-Informed Architecture:
Efficient Inference:
Hardware Support:
Software Stack:
Developer Tools:
Sources: NVIDIA Technical Blog - Cosmos, Cosmos Documentation
vs. Other World Models:
| Model | Physical Reasoning | Action Generation | Open Source | Video Quality |
|---|---|---|---|---|
| Cosmos 3 | ✅ Excellent | ✅ Yes | ✅ Fully open | ✅ High |
| Google Genie 2 | 🟡 Good | ❌ No | ❌ Closed | ✅ High |
| OpenAI Sora | 🟡 Limited | ❌ No | ❌ Closed | ✅ Excellent |
| Runway Gen-3 | ❌ Weak | ❌ No | ❌ Closed | ✅ High |
Key differentiator: Cosmos 3 is the only fully open model combining vision reasoning, world simulation, AND action generation for robotics.
Download models:
# Via Hugging Face
from transformers import CosmosModel
model = CosmosModel.from_pretrained("nvidia/cosmos-3-super")
# Via NVIDIA NGC
ngc registry model download-version nvidia/cosmos-3-super:latest
Quick start example:
import cosmos
# Load model
model = cosmos.load("cosmos-3-super")
# Generate world simulation from image
image = cosmos.load_image("scene.jpg")
video = model.generate_video(
image=image,
prompt="A person walks through the scene",
duration=8.0, # seconds
fps=30
)
# Generate robot action policy
observation = get_robot_observation()
action = model.generate_action(
observation=observation,
task="pick up the red cube"
)
Full documentation: docs.nvidia.com/cosmos
Cosmos 3 represents a paradigm shift from separate vision/simulation/action systems to unified Physical AI models:
The vision: Every robot manufacturer can use Cosmos 3 to generate training data, test policies, and accelerate development - without requiring massive real-world data collection.
NVIDIA's most aggressive consumer play yet: RTX Spark superchip brings desktop-class AI to slim Windows laptops.
Unified Superchip Design:
| Component | Specifications |
|---|---|
| CPU | 20-core NVIDIA Grace (Arm-based) |
| GPU | Blackwell RTX with 6,144 CUDA cores |
| Tensor Cores | 5th-generation with FP4 precision |
| Memory | Up to 128GB LPDDR5X unified |
| Memory Bandwidth | Up to 300 GB/s |
| Interconnect | NVLink-C2C chip-to-chip |
| AI Performance | 1 petaflop FP4 compute |
| Power | Optimized for laptop thermal envelopes |
Unified Memory Architecture:
AI-First Design:
Windows Transformation:
Despite AI focus, RTX Spark delivers:
Laptop Partners (Fall 2026):
Expected configurations:
Desktop Partners:
Sources: Tom's Hardware, NVIDIA GeForce, HotHardware
NVIDIA's most powerful desktop AI supercomputer.
GB300 Grace Blackwell Ultra Superchip:
| Feature | Specification |
|---|---|
| Architecture | Grace Blackwell Ultra |
| Memory | 775GB coherent unified memory |
| Precision | FP4, FP8, FP16, FP32 support |
| Model Capacity | Up to 1 trillion parameters |
| Interconnect | NVLink fabric |
| Form Factor | Desktop tower |
| Cooling | Advanced liquid cooling |
Model Execution:
Enterprise Use Cases:
Software Stack:
Launch: Spring 2026
Partners:
Expected pricing: $45,000 - $85,000 depending on configuration
Sources: NVIDIA DGX Spark, NVIDIA Blog - DGX Station
NVIDIA unveiled Vera Rubin computing platform and Vera CPUs for AI-native data centers.
VR200 Rack System:
Vera CPU:
While not officially confirmed, industry sources suggest:
Sources: Benzinga, TradingKey
Comprehensive tools for building production-grade AI agents.
OpenShell:
NemoClaw:
AI-Q Blueprints:
Dynamo Deployment Recipes:
Integration Options:
Sources: CallSphere Blog, NVIDIA Developer
The latest version of AI-powered upscaling:
| Feature | Description |
|---|---|
| Frame Generation | AI-generated intermediate frames |
| Ray Reconstruction | AI-enhanced ray tracing quality |
| Super Resolution | 4K from 1080p with AI |
| Latency Reduction | New Reflex improvements |
| Compatibility | 500+ supported games |
RTX Video Enhancements:
Broadcast Updates:
Source: NVIDIA GeForce
Jensen Huang offered a glimpse of Nemotron 4, expected later in 2026.
Multimodal Native:
Longer Context:
Rubin Architecture:
Enhanced Agent Capabilities:
Sources: Build Fast with AI, NVIDIA Nemotron
NVIDIA continues aggressive open-source strategy.
Available Now:
Coming Soon:
NVIDIA released massive training datasets:
| Dataset | Size | Description |
|---|---|---|
| Nemotron-CC-v2.1 | 2.5T tokens | English Common Crawl + synthetic |
| Nemotron-CC-Code-v1 | 428B tokens | High-quality code from Common Crawl |
| Nemotron-Pretraining-Code-v2 | - | Curated GitHub with quality filters |
| Nemotron-Pretraining-Specialized-v1 | - | STEM reasoning & scientific coding |
| Nemotron-SFT-Data | - | Supervised fine-tuning datasets |
| Nemotron-RL-Data | - | Reinforcement learning datasets |
Redistribution Rights:
Source: NVIDIA Nemotron Research
NVIDIA's commitment to Taiwan semiconductor ecosystem.
TSMC Collaboration:
Taiwanese Partners:
AMD CEO Lisa Su announced $10 billion investment in Taiwan, marking largest commitment to Taiwan supply chain to date.
Sources: TradingKey Analysis, Benzinga
Computex 2026 wasn't just a product launch—it was NVIDIA's declaration as an AI platform company. The strategy is clear:
NVIDIA is betting that open models will win in enterprise:
From chips to models to deployment tools:
Everything optimized for autonomous agents:
Bridging consumer and enterprise AI:
$150B annual investment signals:
| Factor | NVIDIA Nemotron 3 Ultra | OpenAI GPT-4 Turbo |
|---|---|---|
| Open Source | ✅ Yes | ❌ No |
| Parameters | 550B (55B active) | Undisclosed |
| Speed | 300+ tokens/sec | ~100 tokens/sec |
| Cost | 30% lower | Reference baseline |
| On-Premise | ✅ Yes | ❌ No |
| Customization | ✅ Full access | ❌ Limited |
| Factor | NVIDIA Nemotron 3 Ultra | Meta Llama 3.1 405B |
|---|---|---|
| Parameters | 550B (55B active) | 405B (all active) |
| Efficiency | MoE (10% active) | Dense (100% active) |
| Speed | 5x faster | Baseline |
| Benchmarks | Higher on most | Strong baseline |
| Hardware Req. | Lower (MoE) | Higher (dense) |
| Factor | NVIDIA Nemotron 3 Ultra | Anthropic Claude 3 Opus |
|---|---|---|
| Open Source | ✅ Yes | ❌ No |
| Context | 1M tokens | 200K tokens |
| Agent Design | Purpose-built | General-purpose |
| Deployment | Any infrastructure | API only |
| Reasoning | Optimized for agents | Strong conversational |
Immediate opportunities:
Long-term impact:
Strategic considerations:
Risk factors:
Challenges:
Opportunities:
Intel, AMD, Qualcomm:
Apple:
Google, Microsoft, Amazon:
1. Start with Nemotron 3 Nano (Available Now)
# Install NeMo framework
pip install nemo_toolkit
# Download Nemotron 3 Nano
from nemo.collections import llm
model = llm.load_model("nvidia/nemotron-3-nano-30b-a3b")
# Run inference
response = model.generate("Explain quantum computing", max_tokens=500)
2. Explore Pre-Built Agents
3. Access Training Data
1. Evaluate DGX Station (Available Spring 2026)
2. Pilot Agent Deployments
3. Plan RTX Spark Rollout
Official NVIDIA:
News Coverage:
Technical Analysis:
This comprehensive recap covers NVIDIA Computex 2026 announcements as of June 1, 2026. Specifications, availability, and pricing are subject to change. Visit nvidia.com/gtc/taipei for official details and session recordings.