AI 需求与供应链建模框架：资本支出峰值或在 2028 年-U.S. Internet & Semiconductors_ Framework for Modeling AI Demand & Supply – Capex 'Peak' Likely in 2028

Summary of Key Points from the Conference Call Industry Overview - The report focuses on the U.S. Internet & Semiconductors industry, particularly the AI sector and its capital expenditure (capex) trends [1][2]. Core Insights 1. AI Capex Peak: The analysis predicts that AI capital expenditure will peak around 2028, exceeding $1 trillion, which is approximately $300 billion above current consensus estimates [1][4]. 2. AI Adoption Acceleration: There is a notable acceleration in AI adoption, with leading labs reporting annual recurring revenues (ARRs) increasing by 20-35% in early 2026 [2][3]. 3. Recursive Self-Improvement (RSI): The emergence of RSI is expected to drive faster innovation, with AI models capable of self-optimization and continuous improvement starting in 2027 [3][12]. 4. Compute Capacity Requirements: By 2029, AI labs will require approximately 23 GWs of new compute capacity, with significant investments needed to support this demand [4][19]. Financial Projections 1. OpenAI and Anthropic Growth: OpenAI's ARR is projected to grow from $6 billion in early 2025 to $25 billion by early 2026, while Anthropic's ARR is expected to rise from $1 billion to $19 billion [7]. 2. Capex vs. Operating Cash Flow: In 2028, AI capex is anticipated to represent around 90% of hyperscaler operating cash flow, with some companies exceeding 100% [4][27]. 3. Training Compute Expense: The training compute expense for OpenAI and Anthropic is projected to peak at $155 billion in 2029, indicating a significant increase in required resources [29][30]. Market Dynamics 1. Hyperscaler Capex Underestimation: The report suggests that the market has consistently underestimated the capex required for AI, with Barclays' estimates being significantly higher than Bloomberg consensus [20]. 2. Inference vs. Training Compute: The report anticipates a shift where inference compute will become the majority of compute requirements in the 2030s, as training compute is expected to decline post-2029 [40]. Additional Considerations 1. AI Query Growth: The number of AI queries is expected to increase significantly, with estimates of around 4 trillion queries in 2027, doubling to 9 trillion by 2030 [45]. 2. Agentic AI Workflows: The transition to agentic AI is expected to disrupt various sectors, with AI systems becoming more autonomous and capable of handling complex tasks [12][15]. 3. Caveats on Forecasts: The report acknowledges potential discrepancies in compute forecasts between OpenAI and Anthropic, suggesting that these figures may converge over time [9][30]. Conclusion - The analysis indicates a robust growth trajectory for the AI sector, with significant capital investments required to meet the increasing demand for compute capacity. The anticipated advancements in AI capabilities, particularly through recursive self-improvement, are expected to further accelerate this growth, presenting both opportunities and challenges for investors and companies in the sector [1][3][4].