Slide 1: NVIDIA
GI NVIDIA
Slide 2: Partners and customers
. = Ld accenture Adobe & airbnb Alibaba Cloud amazon AMGEN Ansys € oramco Mall arm Sarat Baitae BARCLAYS BlackRock ee BNP PARIBAS Bristol Myers Squibb cadence capitate Ocerenceai Geel, =~ COHESITY «= € CoreWeave &SORTEVA Sdatabricks 2 DataRobot Oddn S) Deloitte. ADELTA “Dienep Research (=) ebay t m-
Slide 3: RES
RES revere | aye eer) BoE TY eel wind
Slide 4: Slide 4
Slide 5: Physical AI
PHYSICAL AI AUTONOMOUS VEHICLES GENERAL ROBOTICS AGENTIC Al CODING ASSISTANT CUSTOMER SERVICE PATIENT CARE y, GENERATIVE Al 2012 ALEXNET Conreit CREAT ON PERCEPTION AI a WSESE gear MEDICAL IMAGING
Slide 6: From One to Three
FROM ONE TO THREE SCALING LAWS TEST-TIME“SCALING “LONG THINKING" “INTELLIGENCE” POST-TRAINING SCALING PRE-TRAINING SCALING 4 PERCEPTION G st a » ENERATVE . AGENTIC > PRYSICAL
Slide 7: From One to Three
FROM ONE TO THREE SCALING LAWS TEST-TIME“SCALING “LONG THINKING" “INTELLIGENCE” REASONING AT INFERENCE COMPUTE >100X ONE-SHOT POST-TRAINING SCALING PRE-TRAINING SCALING 4 \ PERCEPTION > GENERATIVE > AGENTIC > PHYSICAL 7 Al Al Al Al
Slide 8: "Intelligence"
“INTELLI GENCE” FROM ONE TO THREE SCALING LAWS TEST-TIME*SCALING “LONG THINKING" PROBLEM | PROMPTS VERIFIER MODEL POST-TRAINING WITH RLVR >100T TOKENS ANSWER POST-TRAINING SCALING PRE-TRAINING SCALING 4 PERCEP i n a Ene Gu » GENERATIVE , acus > Hires
Slide 9: Top US CSPs
Top 4 US CSPs 3.6M Blackwell GPUs So Far (2-GPUs per Chip) 1.3M Hopper GPUs 2024 2025
Slide 10: Computing at Inflection Point
Computing At Inflection Point $1,000 $1T+ $750 NVDA DC Revenue Data Center Capex $500 (Dell’Oro Projections) $250 2022 2023 2024 2025 2026 2027 2028
Slide 11: CUDA-X for Every Industry
CUDA-X FOR EVERY INDUSTRY : q PARABRICKS OPT AERIAL cuPYNUMERIC cuQUANTUM EARTH-2 MONal SIONNA , . TRTLLM cuDA-a ee MEDICAL __ GENE Decision NUMERICAL Ga MEGATRON TENSOR — ANALYTICS IMAGING SEQUENCING OPTIMIZATION ‘COMPUTING cuSPAR ouanTum 3 oe Lee cru (COMPUTING cuFFT CHEMISTRY PrYsics AMGX COMPUTE
Slide 12: Slide 12
Slide 13: Leading Technology, Full-Stack AI, CUDA-X Diverse Applications
Omniverse-Cosmos NVIDIA AI Enterprise Isaac GR00T DRIVE AV Leading Technology Omniverse-Cosmos Full-Stack AI CUDA-X Diverse Applications Rich Developer Ecosystem Aerial-Sionna AI Factory 6G AI-RAN Blueprint and Full Stack
Slide 14: NVIDIA
CINVIDIA
Slide 15: Announcing NVIDIA Halos
Announcing NVIDIA Halos Chip-to-Deployment AV Safety System
Slide 16: Slide 16
Slide 17: Grace Blackwell in Full Production
Grace Blackwell in Full Production
Slide 18: NVIDIA Blackwell System
NVIDIA Blackwell System 72 Blackwell GPUs 130 Trillion Transistors 1.1 EF FP4 576 Memory Chips 2,592 Grace CPU Cores 20 TB 576 TB/s 72 ConnectX-8 NICs 18 NVLink Switches 18 BlueField DPUs 130 TB/s All-to-All
Slide 19: Inference At-Scale is Extreme Computing
Inference At-Scale is Extreme Computing Revenue Throughput Tokens per Second FLOPS For Factory HBM Memory HBM Bandwidth Software and Architecture Smart AI Fast Response Tokens per Second For One User
Slide 20: Reasoning model vs traditional LLM tokens
| need to seat 7 people around a table at my wedding reception, but my parents and in-laws should not sit next to each other. Also, my wife insists we look better in pictures when she's on my left, but | need to sit next to my best man. How do | seat us on a round table? But then, what happens if we
Slide 21: Announcing NVIDIA Dynamo
Announcing NVIDIA Dynamo Distributed Inference Serving Library Query KV $ Prefill Decode Disaggregated Inference GPU Resource Allocation KV Cache Routing Communication Library (NIXL)
Slide 22: Blackwell Giant Leap in Inference Performance
Blackwell Giant Leap in Inference Performance 10,000,000 9,000,000 8,000,000 Throughput 7,000,000 6,000,000 5,000,000 TPS / MW 4,000,000 3,000,000 2,000,000 Hopper 1,000,000 0 0 100 200 300 400 500 600 Smart AI TPS for 1 User Fast Response
Slide 23: Blackwell Giant Leap in Inference Performance
Blackwell Giant Leap in Inference Performance 10,000,000 9,000,000 8,000,000 Throughput 7,000,000 6,000,000 5,000,000 TPS / MW 4,000,000 3,000,000 Blackwell NVL8 2,000,000 FP8 Hopper 1,000,000 0 0 100 200 300 400 500 600 Smart AI TPS for 1 User Fast Response
Slide 24: Blackwell Giant Leap in Inference Performance
Blackwell Giant Leap in Inference Performance 10,000,000 9,000,000 8,000,000 Throughput 7,000,000 6,000,000 5,000,000 TPS / MW 4,000,000 3,000,000 Blackwell Blackwell NVL8 NVL8 2,000,000 FP8 FP4 Hopper 1,000,000 0 0 100 200 300 400 500 600 Smart AI TPS for 1 User Fast Response
Slide 25: Blackwell Giant Leap in Inference Performance
Blackwell Giant Leap in Inference Performance 10,000,000 9,000,000 8,000,000 Throughput 7,000,000 6,000,000 5,000,000 TPS / MW 4,000,000 3,000,000 Blackwell Blackwell Blackwell NVL8 NVL8 NVL72 2,000,000 FP8 FP4 FP4 Hopper 1,000,000 0 0 100 200 300 400 500 600 Smart AI TPS for 1 User Fast Response
Slide 26: Blackwell Giant Leap in Inference Performance
Blackwell Giant Leap in Inference Performance 10,000,000 9,000,000 8,000,000 Throughput 7,000,000 6,000,000 5,000,000 TPS / MW 4,000,000 3,000,000 Blackwell Blackwell Blackwell NVL8 NVL8 NVL72 Blackwell 2,000,000 FP8 FP4 FP4 Dynamo NVL72 Hopper FP4 1,000,000 Dynamo 0 0 100 200 300 400 500 600 Smart
Slide 27: Blackwell 25X Hopper
Blackwell 25X Hopper FP4, NVL72, Dynamo, and TRT-LLM Continuous Optimization 1K ISL / 2K OSL 10,000,000 EP8, Batch 3072, Disagg Off EP32, Batch 1792, Disagg Off 9,000,000 EP64, Batch 896, Disagg Off EP64, Batch 512, Disagg Off 8,000,000 EP64, Batch 384, Disagg Off Throughput EP64, Batch 256, Disagg
Slide 28: Blackwell 40X Hopper
Blackwell 40X Hopper FP4, NVL72, Dynamo, and TRT-LLM Continuous Optimization 32K ISL / 8K OSL 1,000,000 EP64PP2+PP4, Batch 480, 52% Context EP64PP2+PP4, Batch 224, 50% Context 900,000 EP64PP2+PP4, Batch 160, 48% Context EP64PP4+PP4, Batch 224, 46% Context 800,000 EP64PP4+PP4, Batch 160, 44% Context
Slide 29: 100 MW AI Factory H100 NVL8
100 MW AI Factory H100 NVL8 GPU Dies 45K Racks 1,400 Data Center Productivity 300M DeepSeek R1 Context and Generation, , ISL=32K, OSL=8K @ Pareto optimal TPS/User
Slide 30: Blackwell 40X Hopper Inference Performance
Blackwell 40X Hopper Inference Performance NVLink Flops Per Watt ~ AI Factory Output 100 MW AI Factory H100 NVL8 GB200 NVL72 GPU Dies 45K 85K Racks 1,400 600 Data Center Productivity 300M 12,000M 40X More Token Revenue DeepSeek R1 Context and Generation, , ISL=32K, OSL=8K @ Pareto optimal TPS/User
Slide 31: Omniverse
be ee cise sana | (a gp samme! www |+-7 — —<— i usp ame i — Omniverse
Slide 32: Slide 32
Slide 33: Blackwell Ultra NVL72
Grace Blackwell Ultra NVL72 Second Half 2025 1.1 EF Dense FP4 Inference 0.36 EF FP8 Training 1.5X GB200 NVL72 New Attention Instructions Blackwell Ultra 2X 20 TB HBM | 40 TB Fast Memory 1.5X 14.4 TB/s CX8 2X 2 Reticle-Sized GPUs 15PF Dense FP4 | 288GB HBM3e Oberon Rack Liquid Cooled
Slide 34: Vera Rubin NVL144
Vera Vera Rubin NVL144 Second Half 2026 88 Custom Arm Cores 176 Threads 3.6 EF FP4 Inference 1.8 TB/s NVLink-C2C 1.2 EF FP8 Training 3.3X GB300 NVL72 13 TB/s HBM4 Rubin 75 TB Fast Memory 1.6X 260 TB/s NVLink6 2X 28.8 TB/s CX9 2X 2 Reticle-Sized GPUs 50PF FP4 | 288GB HBM4 Oberon Rack Liquid Cooled
Slide 35: Rubin Ultra NVL576
Vera Rubin Ultra NVL576 Second Half 2027 88 Custom Arm Cores 176 Threads 15 EF FP4 Inference 1.8 TB/s NVLink-C2C 5 EF FP8 Training 14X GB300 NVL72 4.6 PB/s HBM4e Rubin Ultra 365 TB Fast Memory 8X 1.5 PBs NVLink7 12X 115.2 TB/s CX9 8X 4 Reticle-Sized GPUs 100PF FP4 | 1TB HBM4e Kyber Rack Liquid Coole
Slide 36: NVIDIA Rubin System
NVIDIA Rubin System 576 Rubin GPUs 1,300 Trillion Transistors 15 EF FP4 2,304 Memory Chips 12,672 Vera CPU Cores 150 TB 4600 PB/s 576 ConnectX-9 NICs 144 NVLink Switches 72 BlueField DPUs 1500 PB/s Vera Rubin NVLink576
Slide 37: NVIDIA Blackwell System
NVIDIA Blackwell System 72 Blackwell GPUs 130 Trillion Transistors 1.1 EF FP4 576 Memory Chips 2,592 Grace CPU Cores 20 TB 576 TB/s 72 ConnectX-8 NICs 18 NVLink Switches 18 BlueField DPUs 130 TB/s
Slide 38: Fast Roadmap Drives AI Factory Economics
Fast Roadmap Drives AI Factory Economics Perf = Leadership TCO / Perf = Cost 900 1 0.13 68 0.03 1 Hopper Blackwell Rubin Hopper Blackwell Rubin *Without sparsity
Slide 39: Spectrum-X Supercharged Ethernet
Spectrum-X “Supercharged” Ethernet Powering the World’s Most Advanced AI Clouds and Factories Spectrum-X SuperNIC Spectrum-X Ethernet Switch
Slide 40: NVIDIA Photonics
NVIDIA Photonics CPO Co-Invention With Ecosystem Partners 1st 1.6T Silicon Photonics CPO Chip - New Micro Ring Modulators (MRM) 1st 3D-Stacked Silicon Photonics Engine with TSMC Process High-Power, High-Efficiency Lasers Detachable Fiber Connectors 100’s of Patents, Licensed to Partners Spectrum-X I
Slide 41: Co-Packaged Optics components
Optical Electronic IC Sub-Assembly 3D Stacked Electronic COUPE uLens with Fiber and Photonic ICs surface coupling Connector Photonic IC External Laser Source Module Laser Source Package Co-Packaged Optics Quantum-X Photonic Switch Interposer
Slide 42: LLL)
LLL) a|| Sy) ee)
Slide 43: Announcing NVIDIA Photonics Switch Systems
Announcing NVIDIA Photonics Switch Systems World’s Most Advanced Co-Packaged Optical Switches Spectrum-X Photonics Quantum-X Photonics 2nd Half 2026 2nd Half 2025 144 Ports of 800G | 576 x 200G 128 Ports of 800G | 512 x 200G 512 Ports of 800G | 2K x 200G
Slide 44: NVIDIA Paves Road to Gigawatt AI Factories
NVIDIA Paves Road to Gigawatt AI Factories One-Year Rhythm | Full-Stack | One Architecture | CUDA Everywhere Feynman Blackwell Rubin COMPUTE Blackwell Blackwell Ultra Rubin Rubin Ultra Feynman 8S HBM3e 8S HBM3e 8S HBM4 16S HBM4e Next–Gen HBM Grace CPU Vera CPU Vera CPU NVLINK (SCALE-UP) 5th Gen NVL
Slide 45: Reinventing $500B Enterprise IT For the Age of AI
Reinventing $500B Enterprise IT For the Age of AI AI OS INFRASTRUCTURE Compute Networking Storage
Slide 46: SISA ARS
= SISA ARS
Slide 47: Slide 47
Slide 48: DGX Station
DGX Station The Ultimate Workstation for AI and Data Science GB300 Superchip 784 GB Unified System Memory 20,000 AI TFLOPS ConnectX-8 SuperNIC
Slide 49: NVIDIA AI Infrastructure for Enterprise Computing
NVIDIA AI Infrastructure for Enterprise Computing DGX Spark 1 PFLOPS DGX GB300 DGX B200 15 EFLOPS DGX RTX PRO RTX PRO Station Enterprise Server Workstation 20 PFLOPS
Slide 50: Storage Leaders Build AI Data Platforms for Enterprise AI
Announcing Storage Leaders Build AI Data Platforms for Enterprise AI Semantic Query Agent NVIDIA AI-Q Blueprint Llama Nemotron Compute Node NVIDIA Blackwell NVIDIA BlueField NeMo Retriever NVIDIA Spectrum-X Enterprise Files Vector Database Storage Node NVIDIA cuVS NVIDIA BlueField / ConnectX
Slide 51: Industry's Broadest End-to-End
e . ° * ° . D@LL + >| INDUSTRY’S BROADEST END-TO-END Technologies NVIDIA. NVIDIA Al ENTERPRISE INFRASTRUCTURE — Se eee Ne | Reda @nw4 ———— Fiascan entre SC tT! Beer AIPCs Spark Station Compute Networking Storage 2 @) @) @) + Customers 4 QO + New releases | 4 9 ++ Exaflops J lle wee __) eee) . e ~ 2
Slide 52: Announcing NVIDIA Llama Nemotron Reasoning
Announcing NVIDIA Llama Nemotron Reasoning Distilled, Quantized, Aligned, and Optimized by NVIDIA 3,000 NVIDIA Llama Nemotron Super 49B Tokens/s 2,000 1,000 Llama 3.3 DeepSeek R1 Llama 70B 70B 0 50 55 60 65 70 75 80 Average Accuracy Across Agentic Tasks (%) Nano Super Ultra
Slide 53: Accenture
> accenture
Slide 54: Amdocs amAIz
©: amdocs amdocs amAlz
Slide 55: AT&T Ask AT&T Gen AI Platform
(@ AT&T Ask AT&T Gen Al Platform [ATT Knowledge Base S wean Tar KDA NVIDIA NeMo Curator ATT Does ATT Data Kubeflow Training AtVoleano Scheduler Operator i Training Job KBs GPU ‘Cluster NVIDIA & ‘Customizer a& DataStore — | NVIDIA & Evaluator SS ATET Domain ZS services @ a NVIDIA LLM NIM Inference Se
Slide 56: BlackRock's Aladdin Platform
BlackRock’s Aladdin Platform BlackRock.
Slide 57: Cadence JedAI Platform
cadence Cadence’ JedAl Platform ‘Cadence Agentic Al Verification PCBIBDAC Mutiphysies Cadence® JedA\ Piatform Data Refinement || Embedding Promt Foundaton Models S & ‘COMPUTE. Jence’ ovis HT “Anvioa, srayese em
Slide 58: Capital One MACAW
Capital One MACAW ~—s Data Generation Customization Evaluation 8 x S 2 = 6 @ "| ay * NVIDIA NeMo NVIDIA Triton im) opyrech ovens A sores o. @ ~ as ood the optimization overs ua Capital @ g- 68» NVIOW Tensor PALE MACAW optinied A of nares) (2 @& & @ NVICIAAIOG —-NVIDIANeMo—«NUIDIATensockT™ == Kuberne
Slide 59: Deloitte Zora AI
Deloitte. Deloitte Zora Al
Slide 60: EY.ai Agent Platform
EY EY.ai Agent Platform
Slide 61: Nasdaq Generative AI Platform
jl Nasdaq Nasdaq Generative Al Platform authorzer Skil ‘ Gen Al Account O-B-® GPU Inference Account
Slide 62: SAP Business AI
SAP Business Al Serving/Inference Engine
Slide 63: ServiceNow AI Agents
servicenow. ServiceNow Al Agents So» Triggers & } y/ iney secondary User Message “a MB Agent Collaboration ServiceNow Agentic Workflow sunnomous,atapte, eaten etigees Tools — NVIDIA Blueprint ent ils Context | Knowledge LLM (Planners) LLM (Execution)
Slide 64: Robots - The Next Multi-Trillion Dollar Industry
ROBOTS - THE NEXT MULTI-TRILLION DOLLAR INDUSTRY SIMULATE 98 | cocoo pote
Slide 65: Slide 65
Slide 66: Post-Training Human Demonstration
Post-Training Human Synthetic Data Generation Demonstration and RL in Physics Sim Internet Video Pre-Training
Slide 67: NVIDIA Omniverse With Cosmos
NVIDIA Omniverse With Cosmos Physical AI Digital Twin Operating System NVIDIA Omniverse Cosmos
Slide 68: NVIDIA Warp and Isaac Lab
‘DYenep Research = <GYNVIDIA © Google DeepMind MuJoCo Disney Research Custom Solver Solver Solver : — MuJoCo Playground os be es Other Simulator NVIDIA Warp Isaac Lab
Slide 69: Slide 69
Slide 70: Announcing NVIDIA Isaac GR00T N1 Humanoid Foundation Model
Announcing NVIDIA Isaac GR00T N1 Humanoid Foundation Model Sensor Tokens Action Tokens Vision Diffusion Language Transformer Model Model Pick up the industrial object and place in yellow bin. Text Tokens
Slide 71: NVIDIA GTC 2025
NVIDIA GTC 2025 BLACKWELL IN BLACKWELL NVL72 AND VERA RUBIN AI INFRASTRUCTURE AI INFRASTRUCTURE FULL PRODUCTION DYNAMO 40X FOR INFERENCE NVIDIA PHOTONICS FOR $500B ENTERPRISE IT FOR ROBOTS AI INFRASTRUCTURE FOR AI CLOUDS $1T Computing Reasoning 100X One-Shot Annual Rhythm for the New Compute, Netwo
Slide 72: Slide 72
Slide 73: NVIDIA
GI NVIDIA