Sensetime Group, Inc. Class B · HKEX:20

Sensetime Group, Inc. Class B — H1 2023 Earnings Presentation

37 slides from Sensetime Group, Inc. Class B's earnings presentation, August 2023

  1. Slide 1: 2023 Interim Results Presentation

    SenseTime Group Inc. 2023 Interim Results Presentation TO CREATE AN AI-EMPOWERED FUTURE THROUGH INNOVATION 2023.08

  2. Slide 2: Overview

    01 Overview - 2

  3. Slide 3: Foundation Models Lead Evolution of AI Industry in 1H 2023

    Foundation Models Lead Evolution of AI Industry in 1H 2023 Breakthroughs in Generative AI Triple-digit % growth in SenseNova Foundation Product Suite generative AI-related revenue Models 20.3% 10.4% RMB 1.4 billion RMB 2.4 billion ( 7% ） FY 2022 1H 2023 1H 2023 Group Revenue 1H 2023 GAAP net loss2 (

  4. Slide 4: SenseCore Provides Ample Computing Power for Foundation Models

    SenseCore Provides Ample Computing Power for Foundation Models China’s Largest ~6,000+ P 1 Building New AIDCs ~30,000+ GPU1 (vs. March 2023: ↑ ~20%) AIDC at (vs. March 2023: ↑ ~11%) in Guangzhou, Chongqing, Continuously expand Shanghai Lingang Shenzhen, Fujian computing power Note： 1. As of 28 Augus

  5. Slide 5: InternLM-123B Released under SenseNova Framework

    InternLM-123B Released under SenseNova Framework EXAM Rank First Rank Second 80 InternLM InternLM LLaMA-2 70 Test Set ChatGPT GPT-4 -104B -123B -70B 60 GPT-4 MMLU 64.3 72.9 69.1 83.0 69.8 REASONING LANGUAGE 50 C-Eval 54.4 67.5 52.5 69.9 50.1 InternLM-123B 40 Exam AGIEval 37.1 57.8 39.9 55.1 40.0 GPT

  6. Slide 6: Path to Surpass GPT-4: 2023 R&D Roadmap

    Path to Surpass GPT-4: 2023 R&D Roadmap Training know-how 100s Billion Stable and ample computing power Parameter Foundation Model Valuable data input 10s Billion Parameter X 100+ x Models trainings Industry 10 Trillion Billion leading Token Parameter X 1000+ x Total data Expected reserve Models tra

  7. Slide 7: Training the Best Performing Foundation Model with Best Quality Data

    Training the Best Performing Foundation Model with Best Quality Data Algorithm + Human Used in Oversight Foundation Model Refined Cleaning Training Industry 10+ Trillion leading  Language detection Token Additional tokens High Quality Total raw data Expected  Text extraction per month reserve rese

  8. Slide 8: Path to Surpass GPT-4: 2023 R&D Roadmap 104B 123B

    Path to Surpass GPT-4: 2023 R&D Roadmap 104B 123B Parameter Parameter Renewing More Powerful Quarterly Foundation Model 2023 Q2 2023 Q3 Training know-how 100s Billion Stable and ample computing power Parameter Foundation Model Valuable data input 10s Billion Parameter X 100+ x Models trainings Indus

  9. Slide 9: Promote Open-sourced AI Platforms with Research Collaborations

    Promote Open-sourced AI Platforms with Research Collaborations InternLM-7B LLaMA2-7B LLaMA2-13B ChatGLM2-6B MMLU1 52.0 46.8 55.0 46.0 AGIEval1 43.7 21.8 30.9 39.0 Open-sourced in 2018 87,000+ Github stars C-Eval1 54.1 32.5 41.4 52.6 SuperClue-Opt2 53.9 Not evaluated 34.3 53.8 FlagEval3 0.618 0.463 0

  10. Slide 10: Create AI Operating System and Traffic Portal to Enhance Productivity

    Create AI Operating System and Traffic Portal to Enchance Productivity Function calling AI Agents Robotic Control “…… are driven by GPT-4……” Error types on a weather GPT-4 LLaMA StarCoder CodeGen query tool  Function calling Error rate 19% 61% 68% 93% - API selection 0% 22% 22% 30%  Code interpret

  11. Slide 11: Large Model Function Calling Tool: Code Interpreter

    Large Model Function Calling Tool: Code Interpreter Question: Generate a population distribution map, using the national population data table and the map of China 11

  12. Slide 12: Rapid Evolution of Text-to-Image Model

    Rapid Evolution of Text-to-Image Model July 2023 April 2023 January 2023 SenseMirage 2.0 SenseMirage 3.0 SenseMirage 1.0 1.4bn 7bn 0.8bn parameters parameters SenseMirage 4.0 parameters Input prompts to generate Realistic lighting effects and Professional photography-level anime and manga images ric

  13. Slide 13: Rapid Evolution of Text-to-Image Model

    Rapid Evolution of Text-to-Image Model SenseMirage 1.0 SenseMirage 2.0 SenseMirage 3.0 Prompt: an astronaut Surfing in the space, highly detailed, high quality, digital painting, dramatic lighting, trending on artstation, masterpiece, illustration 13

  14. Slide 14: Rapid Evolution of Text-to-Image Model

    Rapid Evolution of Text-to-Image Model SenseMirage 1.0 SenseMirage 2.0 SenseMirage 3.0 Prompt: woman with short messy black hair, a square jaw, freckles, 20 years old, wearing streetwear clothes, standing in an alley with neon signs, blue and pink neon lighting 14

  15. Slide 15: Simple Prompting with Enhanced Image Generation

    Simple Prompting with Enhanced Image Generation “Chinese dragon, blue ornaments, jewelry style” 15

  16. Slide 16: Simple Prompting with Enhanced Image Generation

    Simple Prompting with Enhanced Image Generation “Chinese dragon, blue ornaments, jewelry style” “Chinese dragon, blue ornaments, jewelry style” SenseMirage 1.0 SenseMirage 2.0 16

  17. Slide 17: Simple Prompting with Enhanced Image Generation

    Simple Prompting with Enhanced Image Generation “Chinese dragon, blue ornaments, jewelry style” “Imagine a 3D rendering of a Chinese dragon with extreme details and detailed patterns. The dragon stands on a sea of blue onyx, adding depth and mystery. The dragon's body is decorated with exquisite jew

  18. Slide 18: Productive in Large Model R&D

    Productive in Large Model R&D First in the World First in the World First and Best in Best in China China High resolution, effective Remote sensing Large Model First domestic model that #2 globally forecasts of core atmospheric with the most comprehensive outperformed ChatGPT variables for more than

  19. Slide 19: SenseTime's Path to AGI

    SenseTime’s Path to AGI ·· · OpenAI 1026 Note: 1. Computational Complexity = Parameters x amount of tokens GPT-4 China’s GPT-4 2. Data from public sources estimation Computational Complexity 1025 · PaLM-2 · · Google · ·· 1024 Deepmind PaLM Google ·· · InternLM-123B Chinchilla 1023 ANTHROP\C Claude L

  20. Slide 20: Building the Ecosystem with Industry Models, Plug-ins and Applications

    Building the Ecosystem with Industry Models, Plug-ins and Applications ~USD 7 Trillion1 increase in global GDP driven by generative AI USD 2.6 - 4.4 Trillion2 generative AI productivity Gateway to realize ultimate Applications application value impact across industries Over 10,000 Industry-specific

  21. Slide 21: SenseNova Foundation Model Sets

    SenseNova Foundation Model Sets Natural Language AI Digital Avatar 3D Space 3D Object Interaction Text-to-images Generation Reconstruction Reconstruction Outstanding Photographic-level image Rapid generation Centimeter-level precision Detailed languages understanding quality with low cost reconstruc

  22. Slide 22: SenseNova Foundation Model Sets

    SenseNova Foundation Model Sets Natural Language AI Digital Avatar 3D Space 3D Object Interaction Text-to-images Generation Reconstruction Reconstruction Applied industries: Applied industries: Applied industries: Applied industries: Applied industries: healthcare, financial, code e-commerce, advert

  23. Slide 23: Cantonese Large Language Model to be Released

    Cantonese Large Language Model to be Released 1 Understanding of local context 2 Comprehension of verbalized questions 3 Verbalized response 4 Authentic cultural insights 23

  24. Slide 24: Business Review

    02 Business Review - 24

  25. Slide 25: Foundation Models Driving the Growth of Our Business Segments

    Foundation Models Driving the Growth of Our Business Segments Leading to Generative AI Commercialization 20% Revenue Contribution from Generative AI Leading large model capabilities and access to exclusive AI infrastructure leading our clients to the AI 2.0 era Smart Business Smart Life Smart Auto S

  26. Slide 26: Smart Business: Empowering Intelligent Upgrades with Large Models

    Smart Business: Empowering Intelligent Upgrades with Large Models Large models promote intelligent Collaborating with industry-leading clients to transformation of enterprises build large models for various vertical industries RMB mn +50% 854 Energy I Finance I Real Estate I Human Resources 568 Medi

  27. Slide 27: Smart Life: Strong Demand for Multimodal Generative AI Capabilities

    Smart Life: Strong Demand for Multimodal Generative AI Capabilities Generative AI applications drive SenseMARS Mobile Devices significant growth of customers RMB mn  Collaborating with e-commerce, media and social media platforms clients to create industry- +7% specific large models AI Sensor ISP C

  28. Slide 28: Smart Auto: Initiating the Technological Transformation with Large Models

    Smart Auto: Initiating the Technological Transformation with Large Models Rapid growth of mass production Selected automobile companies and vehicle types served revenue and value per vehicle RMB mn -31% 121 84 NR- related revenue Mass 43% +5.7x production revenue 4% 1H 2022 1H 2023 % of Total 9% 6%

  29. Slide 29: SenseAuto Pilot and Cloud: Leading Innovation in Autonomous Driving

    SenseAuto Pilot and Cloud: Leading Innovation in Autonomous Driving Smart cabin products empowered by Large Industry-leading autonomous driving algorithms and Model, establishing traffic portal promoting the integration of Large Models and autonomous driving BEV Perception  First in China to succes

  30. Slide 30: Smart City: Strategic Repositioning on High-quality Revenue

    Smart City: Strategic Repositioning on High-quality Revenue Persistent strategic Large Model breakthroughs Visible results from transformation amid decelerating bring new forms of strategic transformation city intelligent upgrades city digital transformation RMB mn 434 -58% New 184 Infrastructure Fo

  31. Slide 31: Financial Review

    03 Financial Review - 31

  32. Slide 32: 2023 1H Financial Overview

    2023 1H Financial Overview Revenue Gross Profit Adjusted Net Loss(1) RMB million RMB million RMB million 1,608 2,393 934 1,415 1,433 (2,172) (2,393) 649 (2,564) FX gains/ (447) (6) (247) losses 2022 H1 2022 H2 2023 H1 2022 H1 2022 H2 2023 H1 2022 H1 2022 H2 2023 H1 (14.3%) (21.5%) 1.3% 66.0% 67.2% 4

  33. Slide 33: Operating Expenses

    Operating Expenses R&D Expenses G&A Expenses Selling Expenses Opex(1) Growth (excl. SBC) (RMB million) (RMB million) (RMB million) Opex (excl. SBC) Opex (excl. SBC) Opex (excl. SBC) SBC SBC SBC 2,035 1,979 152 66 1,783 39.3% 100 7.0% 1,883 1,913 799 797 6.4% 1,683 769 75 92 (0.6%) 102 499 (4.9%) 402

  34. Slide 34: Working Capital

    Working Capital Turnover Days – Period Ending Method(1) Provision for Impairment (as % of gross trade receivables) Cash Conversion Cycle 33.1% 37.8% 23.4% 388 434 413 2022 H1 2022 YE 2023 H1 500 459 Net Impairment Losses on Financial Assets 403 (RMB million) 38.3% 47.0% 27.7% 244 1,125 178 178 142 1

  35. Slide 35: Capex and Cash Balance

    Capex and Cash Balance Capex(1) Total Cash(2) Equity and Bond Investment(3) (RMB million) (RMB million) (RMB million) Other Capex (1) Net Cash Equity Investment Office Space Related Capex Total Borrowings Bond Investment 76.1% 33.0% 49.0% 6,806 6,676 6,456 4,057 18,048 1,461 1,523 1,608 1,065 15,080

  36. Slide 36: Q&A

    Q&A -

  37. Slide 37: THANKS!

    THANKS! TO CREATE AN AI-EMPOWERED FUTURE THROUGH INNOVATION 2023.08

SenseTime Group — H1 2023 Earnings Presentation on 6ix