NVIDIA GTC 2026 Keynote - Jensen Huang's Vision for the AI Factory Era
Quick Navigation
Tags
Share the slides
NVIDIA GTC 2026 Keynote - Jensen Huang's Vision for the AI Factory Era
Detailed view of each slide page, including layout, key content and visual elements.
Opening slide with NVIDIA GTC logo and industry vertical icons: Automotive, Financial Services, Healthcare, Industrial, Media, Quantum, Retail, Robotics, Telco
Dark background with centered GTC logo and 9 icon cards below
Strategic partnership slide showing NVIDIA, Palantir, and Dell Technologies collaboration with enterprise customers (Accenture, CenterPoint Energy, Hertz, Lowe's, Teton Ridge) and Palantir platform tools
Three logos at top with hearts, partner logos in mid-section, Palantir tools grid, NVIDIA AI Platform stack at bottom
Comprehensive ecosystem map of AI-native companies across 8 categories: AI for Auto, Customer Support, Engineering, Healthcare, Robotics, Search, Software Development, plus DL/Inference Frameworks, Agent Frameworks, Frontier Model Builders, Model to Production
Grid layout with company logos organized by category, NVIDIA AI Platform stack at bottom
Cinematic aerial rendering showing a plane flying over mountainous terrain, demonstrating NVIDIA Omniverse simulation capabilities
Full-bleed cinematic image with rounded frame
Timeline showing inference compute growth: 2023 (ChatGPT), 2024 (o1 - 10X models/context, 10X tokens), 2025 (Claude Code - 100X models/context, 100X tokens)
Horizontal timeline with three circular icons representing each era
NVIDIA full-stack expanding to all regions/industries. Blackwell+Rubin pipeline growing from $0.5T to $1T+. New customers since GTC DC 2025: Anthropic, MetaSL, Multiple OSS, xAI, Gemini, OpenAI. CUDA distribution: 60% Cloud/AI Natives, 40% Hyperscalers + NCP/SCC/Sovereign AI/Industrial/Enterprise
Two-chart layout: bar chart for pipeline growth, pie chart for CUDA distribution
GB NVL72 'Inference King': 50X higher perf/watt, 35X lower token cost vs competition. InferenceX by SemiAnalysis benchmarks using DeepSeek R1 0528 FP4 1K/1K
Two side-by-side charts: Tokens per Watt vs Interactivity, Performance vs Token Cost
Output speed benchmark (March 3, 2026): Baseten 185, Fireworks 170, Lightning AI 99 tokens/sec. All leading inference endpoints run on NVIDIA. Kimi K2.5 Reasoning model
Horizontal bar chart ranking inference providers by output speed
Core thesis: Inference is the Workload, Tokens are the New Commodity, Compute is Revenue. Visual flow: Electricity → AI Factory → Tokens → AI Agents
Left-to-right flow diagram with hand-drawn illustrations
Vera Rubin POD specs: AI FLOPS 60,000 PF, Memory Bandwidth 92 TB/s, All-to-All Bandwidth 260 TB/s, Scale-Out Radix 131,072. Shows chip-to-rack progression
Hardware showcase with chips at top, server trays in middle, full rack at bottom
Live keynote stage photo of Jensen Huang presenting inference economics chart to packed audience
Wide-angle stage photography
Framework showing throughput (TPS/MW) vs interactivity (TPS/User) with pricing tiers: Free (Qwen 3, $0), Medium (Kimi K2.5, $3), High (GPT MoE 2T, $6), Premium (GPT MoE 2T 400K, $45), Ultra ($150)
Scatter plot with tier labels at bottom
Hopper performance curve showing baseline throughput across Free and Medium tiers
Line chart with single Hopper curve
Blackwell NVL72 curve added showing 35X improvement over Hopper, extending to High tier ($6)
Line chart with two curves, Blackwell in green
Blackwell curve extended to Premium tier ($45), showing sustained performance at 400 TPS/User
Same chart with extended Blackwell curve
Rubin NVL72 curve added: 2X over Blackwell at 50 TPS, 2X at 100 TPS, 3X at 200 TPS, 10X at 400 TPS
Three-curve comparison chart
Revenue per gigawatt comparison: Blackwell vs Rubin across tiers. Total annual revenue jumps from $30B to $150B (5X) with Rubin
Bar chart with stacked Blackwell (teal) and Rubin (green) bars
Same three-curve chart showing Hopper → Blackwell → Rubin performance progression
Performance chart with three curves
Rubin + LPX curve extends to Ultra tier ($150, 1000+ TPS/User) maintaining 35X over Hopper across all tiers
Four-curve chart with LPX extension in yellow-green
Agent architecture diagram: NemoClaw at center with connections to Multi-Modal Prompt, Files, Computer Use, Tools (CLI/MCP), OpenShell, Sub-Agents, Skills, LLM, Memory. Powered by Nemotron/NeMo/Dynamo/NIM and cuDF/cuVS/vGPU/cuOPT/AI-Q
Circular architecture diagram with glowing green borders on side panels
NVIDIA is world's largest contributor to open-source AI (400+ repos, surpassing Alibaba Cloud). Showcases BioNeMo, Earth-2, Nemotron, Cosmos, GR00T, Alpamayo models
Center chart with model cards flanking both sides
Earth-2 climate/weather AI model showcase with photorealistic Earth rendering
Split view with model cards on left and large Earth globe on right
12 benchmark results across Nemotron 3 Super, NemotronVL, Nemotron VoiceChat, Cosmos Predict/Reason, GR00T N, Alpamayo. All showing top-tier or SOTA performance
4x3 grid of benchmark result tables
Leaderboard: claude-sonnet-4.6 (86.9%), claude-opus-4.6 (86.3%), gpt-5.4 (86.0%), nvidia/nemotron-3-super-120b-a12b (85.6%) - best open model
Horizontal bar chart leaderboard
5X efficiency and highest reasoning accuracy on GB200 NVL72. Comparison vs GLM and Kimi K2 across Peak Throughput, MMLU Pro, HumanEval/MBPP, GSM8K, Global MMLU
Two-chart comparison: throughput bars and accuracy bars
Global map showing Nemotron adoption: US, EuroLLM, Linagora (France), Bielik.ai (Poland), SOOFI, Dicta (Israel), TII (UAE), HUMAIN, BharatGen/Gnani.ai/Sarvam AI (India), Viettel, NAVER/Trillion Labs, Institute of Science Tokyo, AI Singapore, Indosat, WideLabs (Brazil)
Night-illuminated world map with labeled pins
Global AI leaders joining NVIDIA Nemotron Coalition: Black Forest Labs, Cursor, LangChain, Mistral AI, Perplexity, Reflection, Sarvam, Thinking Machines
NVIDIA logo at top with hearts, partner logos in two rows below
12 enterprise agent architecture diagrams from Adobe, Atlassian, Cadence, Cisco, CrowdStrike, Dassault Systemes, Palantir, Salesforce, SAP, ServiceNow, Siemens, Synopsys. Plus 40+ additional partner logos
4x3 grid of detailed architecture diagrams with partner logo strip at bottom
12 more enterprise agents: Abridge, CodeRabbit, Cohesity, Edison, IQVIA, Mastercard, Microsoft Security, PayPal, Perplexity, Xiaohongshu, Rockwell Automation, Schneider Electric
4x3 grid of architecture diagrams
Same agent architecture diagram as page 20
Circular architecture diagram
Omniverse-rendered robot showroom featuring 30+ robots from Boston Dynamics, Figure, Foxconn, Agility, 1X, LG Electronics, XPENG, Disney Research, and many more. Industrial and humanoid robots alongside vehicles
3D rendered exhibition hall with robots arranged on floor
Split-screen showing simulation (left, pastel colored) vs real-world (right, metallic) robot arm performing assembly tasks
Side-by-side comparison in rounded frame
Photorealistic rendering of Disney's Frozen-themed village created with NVIDIA Omniverse
Full-bleed cinematic render
Whimsical Pixar-style CG scene of Jensen Huang cartoon roasting marshmallows at a campfire with various robots and a lobster mascot
Full-bleed cinematic render
Clean NVIDIA logo on dark background - section divider
Centered logo
Revenue per gigawatt with VR+LPX: Total annual revenue from $30B to $300B (10X). Three-color stacked bars: Blackwell, Rubin, VR+LPX across all pricing tiers
Stacked bar chart with three hardware generations
Chip comparison: Rubin GPU (288GB HBM4, 22 TB/s, 50 PFLOPs NVFP4, 336B transistors) vs Groq 3 LPU (500MB SRAM, 150 TB/s bandwidth, 1.2 PFLOPs FP8, 98B transistors)
Side-by-side chip photos with specs below
Scaled comparison: 8 Groq 3 LPUs = 4GB SRAM, 1,200 TB/s (55X vs Rubin), 9.6 PFLOPs, 784B transistors. 'Uniting Processors of Extreme Performances'
Side-by-side with multiplied Groq chips
Dynamo inference architecture: Vera Rubin NVL72 handles Prefill+Decode Attention, Groq 3 LPX handles Decode FFN. KV Cache shared, Activations exchanged bidirectionally
Technical architecture diagram with three components
Groq 3 LPX rack: 315 PFLOPS AI Inference, 128GB SRAM, 40 PB/s bandwidth, 256 chips, 640 TB/s scale-up. Available 2H26. Shows rack and node detail with 8 LPUs, FPGA, Host CPU, BF4
Product photo with specs list on right, detailed node diagram below
AI Labs: Anthropic, Cursor, Meta, OpenAI, Perplexity, Black Forest Labs, Cohere, Harvey, Mistral AI, OpenEvidence, Runway, Thinking Machines. Cloud: AWS, Google Cloud, Azure, Oracle, CoreWeave, Crusoe, Lambda, Nebius, NScale. OEMs: Dell, HPE, Lenovo, Supermicro, etc.
Product photo left, three-tier partner grid right
Cloud partners: Meta, Oracle, CoreWeave, Crusoe, Lambda, Nebius, NScale, together.ai. OEMs: Cisco, Dell, HPE, Lenovo, Supermicro, ASRock, ASUS, COMPAL, Foxconn, etc. 256 Vera CPUs, 300 TB/s LPDDR5X
Product photo left, two-tier partner grid right
Cloud: CoreWeave, Crusoe, CREN, Lambda, Mistral AI, Nebius, Oracle, Vultr. OEMs: AIC, Cloudian, DDN, Dell, HPE, Hitachi, IBM, MinIO, NetApp, Nutanix, etc. 5x Tokens/sec, 50 Tb/s, 16TB Shared Context/GPU
Product photo left, two-tier partner grid right
Complete system: 1GW AI Factory comparison X86+Hopper (600K GPUs, 1.2 ZFLOPS, 2M TPS) vs Vera Rubin (300K GPUs, 16 ZFLOPS, 700M TPS). Shows all 7 chips and 6 system form factors
Comparison table at top, chip/tray images in two rows, rack at bottom
Three-generation roadmap: Blackwell (2024), Rubin (2026), Feynman (2028). Each with chip lineup, rack systems. Feynman introduces Die Stacking Custom HBM, LP40 NVLink, Rosa CPU, BlueField-5, NVLink 8 CPO
Left-to-right progression with stacked hardware renders
Extreme Co-Design at Infrastructure Scale. Three layers: Reference Designs (Rubin DSX, DSX Ecosystem, Omniverse DSX Blueprint), Libraries/APIs/Software (DSX Max-Q, DSX Flex, DSX Exchange, DSX Sim), Chips/Systems/Facilities (Power → Liquid Cooling)
Three-tier horizontal architecture with AI factory rendering at bottom
Photorealistic aerial view of a large-scale AI factory data center with rows of liquid-cooled server racks
Full-bleed cinematic render
DSX Flex, DSX Max-Q, DSX Sim, DSX Exchange partner lists. 40+ partners per category including CoreWeave, Crusoe, Digital Realty, NSCALE, Red Hat, Dell, Cadence, Siemens, VAST, WEKA, etc.
Four-quadrant partner grid with factory render at bottom
Announcement of NVIDIA Space-1: Vera Rubin compute module designed for satellite deployment
Chip render on left, solar panel satellite array on right, Earth in background
GitHub star history chart showing openclaw/openclaw surpassing facebook/react and torvalds/linux in 2026. Lobster mascot featured
Star history chart with mascot illustration
Terminal screenshots showing nemoclaw installation (curl -fsSL nvidia.com/nemoclaw.sh | bash) and onboarding (nemoclaw onboard)
Terminal window screenshots stacked
General agent architecture: Agent at center with connections to Multi-Modal Prompt, Files, Computer Use, Tools (CLI/MCP), Sub-Agents, Skills, LLM (multiple providers), Memory
Circular architecture diagram with hand-drawn style
Traditional IT model: Enterprise IT Expense → Software/SaaS + GSI → Files → Data Centers
Vertical flow diagram, left side only
New model: Enterprise Information Workers → Software/SaaS + GSI + AI Providers + AI Natives → Tokens → AI Factories. Side-by-side before/after comparison
Two-column before/after comparison
NVIDIA logo on dark background
Centered logo
Photorealistic water/wave simulation scene showing industrial floodgate mechanism with dynamic water rendering
Full-bleed cinematic render
3D rendered aerial view of GTC 2026 conference venue in San Jose with NVIDIA-branded buildings and event spaces
Cinematic city render with GTC logo overlay
Speaker lineup: Sarah Guo (3rd appearance), Alfred Lin (2nd, 2026 host debut), Gavin Baker (2nd, 2026 host debut). Plus 20+ additional speakers including Harrison Chase, Michael Dell, Anirudh Devgan, etc.
Three featured speakers in large cards at top, grid of smaller speaker cards below
250+ company logos including ABB, Apple, Adobe, Anthropic, AWS, BYD, Cisco, CoreWeave, Disney Research, Meta, Microsoft, OpenAI, Samsung, Tesla, Uber, Walmart, Xiaomi, and many more
Dense logo grid with 3D venue render at bottom
Circular flywheel diagram: Breakthroughs → Developers → Installed Base → Ecosystems → back to Breakthroughs
Centered circular diagram on black background
Product hero shot of GeForce GTX 1080 graphics card - historical reference to CUDA's gaming roots
Centered product photo on black background
Hogwarts Legacy game screenshot with DLSS 5 On, showcasing 3D-Guided Neural Rendering quality
Full-bleed game screenshot
EA Sports FC screenshot with DLSS 5 3D-Guided Neural Rendering announcement
Full-bleed game screenshot
$120B structured data ecosystem map showing cuDF accelerating CSP engines (AWS, GCP, Azure), commercial OSS platforms (Clickhouse, Cloudera, Databricks), proprietary engines, and OSS engines (DuckDB, Pandas, Polars, Spark)
Network diagram with cuDF at bottom connecting to all layers
Unstructured data ecosystem: 100s of zettabytes growing exponentially. cuVS accelerates vector search across enterprise databases, CSP engines, OSS engines (Elasticsearch, FAISS, Milvus, etc.)
Network diagram with cuVS at bottom
IBM watsonx.data with NVIDIA cuDF: Order-to-Cash Data Mart updates 3 min vs 15 min, 83% cost savings (CPU vs GPU). Quote from Chris Wright, CIO of Nestle
Quote at top, architecture diagram left, bar chart right
Aerial photo of Nestle manufacturing facility
Full-bleed photo
Dell AI Data Platform with NVIDIA cuVS/cuDF: Enterprise data processing from hours to minutes (10TB, 3X faster). Quote from Abhijit Dubey, CEO of NTT DATA
Quote at top, Dell architecture diagram left, benchmark chart right
Google Cloud AI Hypercomputer with cuDF: A/B Experimentation Platform 45k CPUs to 1k GPUs, 76% cost savings. Quote from Saral Jain, CIO of Snap
Quote at top, Google Cloud architecture left, savings chart right
Full partnership overview: customers (baseten, CrowdStrike, GM, Puma, Salesforce, Snapchat, etc.), Google services (Vertex AI, GKE, BigQuery, etc.), NVIDIA AI Platform stack
Three-tier partnership layout
Full AWS partnership: customers (CrowdStrike, databricks, OpenAI, Perplexity, Palantir, etc.), AWS services (Braket, EKS, Nitro Enclaves, SageMaker AI, EC2, etc.)
Three-tier partnership layout
Full Azure partnership: customers (Anthropic, BlackRock, BMW, Krones, Rockwell, Synopsys), Azure services (Foundry, Windows 365, Fabric, AKS, Copilot, Bing)
Three-tier partnership layout
Full Oracle partnership: customers (Cohere, Fireworks AI, Neospace, Siemens, Uber, Zoom), OCI services (Superclusters, AI Database, Generative AI, Data Science, OKE)
Three-tier partnership layout
Full CoreWeave partnership: customers (Canva, Cohere, Cursor, Mercado Libre, Mistral AI, Morgan Stanley, OpenAI, Runway), services (W&B Models/Training/Inference/Weave, CoreWeave Inference/Slurm/K8s/Compute/Storage/Mission Control)
Three-tier partnership layout
Common questions about this slide and the underlying presentation content.
The major announcements include the Vera Rubin NVL72 system with 2X performance over Blackwell, the NVIDIA Groq 3 LPX (a new LPU-based inference accelerator with 315 PFLOPS and 128GB SRAM), the complete Vera Rubin platform with 7 chips and 5 rack systems, the Feynman architecture roadmap for 2028, NVIDIA Space-1 satellite compute module, and the DSX AI Factory Platform for infrastructure-scale deployment.
Vera Rubin delivers 2X throughput improvement over Blackwell NVL72 at both 50 and 100 TPS/User, scaling to 3X at 200 TPS and 10X at 400 TPS. A 1GW AI Factory with Vera Rubin uses only 300K GPUs (vs 600K with X86+Hopper) while delivering 16 ZFLOPS (vs 1.2) and 700M tokens per second. It unlocks a $150B annual revenue opportunity per gigawatt, a 5X increase from Blackwell.
NVIDIA and Groq are uniting two processors of extreme performance: the Rubin GPU (288GB HBM4, 50 PFLOPs) for compute-heavy prefill operations, and the Groq 3 LPU (150 TB/s SRAM bandwidth) for bandwidth-heavy decode operations. NVIDIA Dynamo orchestrates the split, sending prefill to Vera Rubin NVL72 and decode FFN to Groq 3 LPX, achieving 35X performance over Hopper alone.
NemoClaw is NVIDIA's reference implementation for OpenClaw, an open agent toolkit for building specialized AI agents. It provides multi-modal prompt handling, file processing, computer use, CLI/MCP tools via OpenShell sandbox, sub-agents, skills, LLM integration, and memory. It's powered by Nemotron 3 Super (the best open model for OpenClaw at 85.6% on benchmarks) and leverages cuDF, cuVS, vGPU, and cuOPT for GPU-accelerated agent capabilities.
NVIDIA presents five inference tiers based on model size and interactivity: Free (Qwen 3, 235B, 32K context, $0), Medium (Kimi K2.5, 1T, 128K, $3), High (GPT MoE 2T, 128K, $6), Premium (GPT MoE 2T, 400K, $45), and Ultra (GPT MoE 2T, 400K, $150). Vera Rubin + LPX enables profitable serving across all tiers, unlocking $300B total annual revenue per gigawatt.
Over 24 major enterprises showcased agent architectures: Adobe, Atlassian, Cadence, Cisco, CrowdStrike, Dassault Systemes, Palantir, Salesforce, SAP, ServiceNow, Siemens, Synopsys, Abridge, CodeRabbit, Cohesity, Edison, IQVIA, Mastercard, Microsoft Security, PayPal, Perplexity, Xiaohongshu, Rockwell Automation, and Schneider Electric.
Vera Rubin NVL72 launch partners include AWS, Google Cloud, Microsoft Azure, Oracle Cloud, CoreWeave, Crusoe, Lambda, Nebius, and NScale. The deck also showcases deep integrations with each provider's specific services (e.g., Amazon SageMaker AI, Google Vertex AI, Azure Foundry, OCI Superclusters). OEM partners include Dell, HPE, Lenovo, Supermicro, Foxconn, and many more.
NVIDIA describes an 'Enterprise IT Renaissance from SaaS to Agent-as-a-Service.' The traditional model of Enterprise IT Expense flowing through Software/SaaS and GSIs to data centers transforms into Enterprise Information Workers supported by Software/SaaS, GSIs, AI Providers, and AI Natives — all consuming tokens from AI Factories. This represents a fundamental shift where compute becomes revenue and tokens become the new commodity.
Create Your Own Slides
Turn your ideas into professional presentations in seconds with 2slides AI.
Reference professional designs, choose your style, and generate slides with perfect text rendering. Powered by Nano Banana—start creating your presentation now.
Your AI Agent for slides. Save time, shine faster with intelligent presentation creation.
© 2026 2slides. All rights reserved.