The "AI Development Trends" report released by Mary Meeker's team provides a comprehensive analysis of the explosive development of AI technology and its global impact. The report points out that AI's user growth (such as ChatGPT reaching 800 million users in 17 months) and capital investment (AI spending by the six major U.S. tech companies exceeding $200 billion in 2025) far exceed any technology in history, and performance metrics (such as 89.8% accuracy on the MMLU test) have approached human level. China and the United States dominate global competition, with the United States leading in large language models (such as GPT-4.5), while China is rapidly catching up in industrial robots (accounting for 50% of the global total) and domestic models (such as Qwen 2.5). AI has penetrated deeply into fields such as finance (such as Bank of America's AI assistant Erica), healthcare (the FDA has approved nearly 700 AI devices), and education (such as Arizona State University's AI courses). Within the next 5 years, human-level text generation and personalized medicine may be achieved, but risks such as autonomous weapons and employment shocks must be guarded against. The report emphasizes that AI's compound growth (algorithms x computing power x data) is reshaping the economy, society, and the geopolitical landscape, and calls for balancing innovation with ethical safety.
Get Mary Meeker's《Trends – Artificial Intelligence》 PDF original file, scan the QR code to follow and reply:20250601
Is the pace of change faster than before? Yes, it indeed is
- Unprecedented pace of technological change: Comparing the pace of technology adoption between the internet era (1990s) and the AI era (2020s), AI's user growth and technological iteration speed far exceed any technology in history. ChatGPT reached 800 million users in 17 months, while the internet and the iPhone took 7 years and 3 years respectively to reach the same scale.
- Compound growth effect of technology: AI performance improvement is driven jointly by algorithmic improvements, growth in computing power, and expansion in data volume, forming a "compound growth" effect. MMLU test accuracy jumped from 30% in 2019 to 89.8% in 2024 (approaching human level).
- Explosion of the developer ecosystem: The number of NVIDIA developers grew 100% within 4 years (3 million in 2021 -> 6 million in 2025), reflecting the acceleration of AI technology deployment.
- Surge in capital investment: AI-related capital expenditure by the six major U.S. tech companies (Apple, NVIDIA, etc.) in 2025 exceeded $200 billion, far exceeding investment during the internet bubble period.
- Simultaneous global adoption: AI technologies (such as ChatGPT) have achieved simultaneous growth across multiple global regions since release, whereas the early internet spread only with Europe and the United States as the center.
AI user, usage, and capital expenditure growth = unprecedented
- Unprecedented speed of consumer/user AI adoption：
- ChatGPT user growth: Reached 800 million users in 17 months (P.52), 5.5 times the growth rate of the internet.
- Global adoption: ChatGPT's launch covered multiple regions, and the share of non-North American users quickly climbed to more than 50% (the internet took 20 years to reach the same proportion).
- AI adoption in the technology ecosystem：
- NVIDIA developer ecosystem: The number of developers grew 100% within 4 years (3 million in 2021 -> 6 million in 2025).
- Google ecosystem: The number of Gemini developers grew 5 times within 1 year to 7 million.
- Surge in capital expenditure by tech giants：
- Six major tech companies: AI-related capital expenditure by Apple, NVIDIA, etc. in 2025 exceeded $200 billion.
- Investment directions: AI infrastructure (such as data centers), large model training, and open-source ecosystems.
- Enterprise AI Application Acceleration：
- Bank of America Case: AI assistant Erica has accumulated 2 billion interactions, saving 40% of customer service costs.
- JP Morgan Practice: AI saves thousands of hours of manpower, optimizing risk management.
- AI Implementation in Healthcare and Government：
- FDA-approved AI devices: From 0 in 1995 to nearly 700 in 2023.
- Government policy: NVIDIA partners with 30+ countries on "Sovereign AI" initiatives.
High/rising AI model computing costs + declining per-inference costs = performance convergence + developer usage growth
- Training costs continue to climb: Large model training costs, GPT-4 training costs exceed $100 million (P.129), average annual growth of 200% (2019-2025).
- Per-inference costs decline rapidly: Per-inference cost from GPT-3 to GPT-4o decreased by 90%. Anthropic Claude 3.5's inference cost is 1/5 that of GPT-4.
- Performance convergence phenomenon: Mainstream models (GPT-4.5, Claude 3.7, Qwen 2.5) all exceed 85% accuracy on the MMLU test.
- Developer ecosystem explosion: NVIDIA developer count grew 100% in 4 years (3 million in 2021 → 6 million in 2025). Google Gemini developers grew 5x to 7 million within 1 year.
- Open-source models drive adoption: Meta's Llama 3 downloads exceed 50 million. China's Alibaba Qwen 2.5 series models approach GPT-4o in performance.
AI usage, costs, and losses growth = unprecedented
- AI usage surges: ChatGPT interactions, daily queries increased from 1 million in November 2022 to 1 billion in April 2025. Bank of America's AI assistant Erica has processed over 2 billion interactions cumulatively.
- Training and inference cost divergence：
- Training cost: GPT-4 training costs exceed $100 million, average annual growth of 200% (2019-2025).
- Inference cost: Per-token processing cost dropped from $0.01 for GPT-3 to $0.001 for GPT−4o.
- Tech giants' losses widen：
- TOpenAI losses: 2024 losses reached $5.3 billion (P.173), mainly due to computing investment and model training.
- Industry trend: Average loss rate of global AI startups rose from 40% in 2021 to 65% in 2025.
- Capital expenditure and returns imbalance: AI-related capital expenditure in 2025 exceeds $200 billion, but most projects have a payback period of over 5 years.
- Industry application cost pressure：
- Healthcare sector: Average deployment cost of FDA-approved AI medical devices increased by 120% (2020-2025).
- Financial sector: JP Morgan's AI risk control system annual maintenance cost reaches $250 million.
AI monetization threat = intensifying competition + open-source momentum + China's rise
- Competitive landscape, US LLM dominance challenged: US LLMs (such as GPT-4.5) account for 62% of global desktop user share, but China's Qwen 2.5 and other mod
- Open-source models disrupt commercial ecosystem：
- Key events：
- Alibaba releases 100 open-source Qwen 2.5 models, performance approaching Western competitors.
- DeepSeek launches R1-Zero open-source reasoning model.
- Impact: Open-source models reduce enterprises' willingness to pay, and the expected revenue growth rate for closed-source LLMs in 2025 is lowered to 35%.
- Key events：
- China's Technological Rise and Geopolitical Competition：
- Industrial Robots: China's deployment volume accounts for more than 50% of the global total.
- Policy Response：
- NVIDIA partners with 30+ countries on "Sovereign AI" initiatives.
- China's 14th Five-Year Plan sets a target of 70% self-sufficiency in AI chips.
- Corporate Response Strategies：
- Microsoft integrates Copilot into the 365 suite, strengthening the subscription model.
- Google Gemini's developer ecosystem grew 5x in 1 year to 7 million.
- Risks and Challenges：
- Commercial Risk: Open-source models cause LLM profit margins to fall from 45% in 2023 to 28% in 2025.
- Geopolitical Risk: The United States restricts exports of high-end GPUs to China, accelerating China's independent chip research and development.
The integration of AI and the physical world = fast + data-driven
- Accelerating Industrial Robot Deployment: China's industrial robot installations account for more than 50% of the global total.
- Breakthroughs in Autonomous Driving Technology: Waymo progresses, and in 2025 robotaxis cover 90% of San Francisco.
- Medical Robot Applications：
- Surgical Robots: The da Vinci system has completed more than 10 million surgeries worldwide.
- FDA Approval: Of nearly 700 AI medical devices in 2023, 23% were robot-assisted.
- Smart Homes and Consumer Robots: The 2025 robot vacuum achieves autonomous charging and waste disposal. Amazon Astro home robot sales increase 300% year over year.
- Data-Driven Urban Management：
- Traffic Optimization: Singapore uses AI to reduce congestion by 30%.
- Energy Networks: Google DeepMind improves grid efficiency by 15%.
- Agriculture and Logistics Automation：
- Agricultural Drones: XAG's drones cover more than 100 million mu of farmland in China.
- Logistics Robots: Amazon deploys more than 500,000 Kiva robots in its warehouses.
Global internet user growth driven by AI = unprecedented speed
- AI-Driven Global User Growth Curve: ChatGPT drives global internet user growth to a historic peak, with the share of non-North American users jumping from 35% in 2023 to 52% in 2025.
- AI Adoption in Emerging Markets：
- Case: Regions such as India and Indonesia skip the PC era directly through AI tools (such as voice assistants), with mobile AI usage increasing 200% annually.
- Data: ChatGPT's user share in Africa rises from 3% in 2024 to 12% in 2025.
- Multilingual and Low-Barrier Access: Meta's universal translation model supports real-time translation across 100+ languages, covering 80% of the global non-English-speaking population.
- Infrastructure and Policy Momentum: NVIDIA partners with 30+ countries to build local AI data centers (such as Saudi Arabia and Singapore), reducing dependence on computing power.
- Education and Inclusive Impact：
- Arizona State University's "AI Acceleration Program" increases course completion rates by 35%.
- The Indian government provides basic education to rural areas through AI voice assistants, reaching 120 million students.
AI and the Transformation of Work = Real + Rapid
- The Structural Impact of AI on the Job Market: In 2025, AI-related positions account for 35% of global IT jobs, while non-AI positions decline at an annual rate of 5%.
- Transformation of Skill Demands: Survey results show that 73% of companies require employees to master basic AI tools (such as Copilot). Demand for programming positions has fallen by 12%, but demand for "AI + domain expert" hybrid talent has grown by 200%.
- Industry Cases, Finance and Healthcare：
- Finance: Goldman Sachs uses AI to complete 60% of financial report analysis, improving analyst efficiency by 3x.
- Healthcare: Kaiser Permanente's AI medical record assistant saves doctors 50% of their paperwork time.
- Policy and Education Responses：
- Government Action: The U.S. Department of Labor launched the "AI Workforce Retraining Fund" in 2024.
- Educational Innovation: Arizona State University launched the "AI Acceleration Program," raising course completion rates by 35%. Oxford University is collaborating with enterprises to develop an AI ethics training module.
- Predictions for Future Work Scenarios：
- Short Term (2025-2030): 50% of repetitive work (such as data entry and basic programming) will be taken over by AI.
- Long Term (2030+): "Human-AI collaboration" will become mainstream, for example, AI managing project processes while humans handle creative decision-making.
Get Mary Meeker's《Trends – Artificial Intelligence》 Original PDF file, scan the QR code to follow and reply:20250601