彭文生系中金公司首席经济学家 中国首席经济学家论坛副理事长 宏观视点 自2022年年底ChatGPT发布以来，美股AI龙头公司（所谓7姐妹）的股价大幅跑赢整体市场，2025年初 DeepSeek出现以来，中国的AI龙头公司（主要在港股）也大幅跑赢大市。就美国股市而言，尽管相关 企业盈利有较快增长，但风险溢价在极低水平，反映了投资者的乐观预期。股票高估值使得近期关于 AI相关资产价格泡沫的讨论多起来，本文不是如何定义泡沫和测算泡沫的技术分析，而是从资产价 格、创新与宏观经济的关系出发提供一些思考。 关键词| 人工智能 规模经济 科技创新 资产价格 研究员| 彭文生 正文 一、因与果 消化股票高估值的一个可能是利率下降，由此一些投资者把乐观的预期寄希望于美联储降息。传统的思 维是利率和风险资产价格是跷跷板的关系，无风险利率下降促使配置向风险资产倾斜，有利于股票估 值。这样的关系在当下是不是仍然成立，我们首先得解释为什么过去几年美元利率上升的环境下，股票 价格大幅上涨。 利率与股市的关系有三个可能。第一是传统的利率是因、股价是果的关系。第二个是反过来，股市是 因、利率是果。AI领衔的股市上涨是美国经济总需求的重要支 ...

Core Viewpoint - The article discusses the relationship between AI-driven stock valuations and macroeconomic factors, emphasizing the potential for both overvaluation and sustainable growth in the context of AI advancements and capital investments [4][5][17]. Group 1: Stock Valuation and Economic Impact - Since the launch of ChatGPT at the end of 2022, AI leading companies in the US and China have significantly outperformed the overall market, with AI-related capital expenditures contributing one-third to US GDP growth this year [4][5]. - The relationship between interest rates and stock prices can be viewed in three ways: traditional cause-and-effect, reverse causation, or both being influenced by external factors [5][6]. - The wealth effect from stock market gains, particularly among the wealthiest 10% of the population, has driven consumer spending, which in turn supports higher natural interest rates [5][6]. Group 2: Cost and Benefit Analysis of AI - The current AI technology development is characterized by low application maturity and high profit expectations, necessitating substantial capital market support [8][9]. - The shift from capital-light software distribution to capital-intensive hardware production is led by major tech companies, which are now primary supporters of AI startups [8][9]. - The economic potential of AI applications remains uncertain, with challenges in quantifying direct and indirect economic benefits [9][10]. Group 3: Economic Growth Projections - Different methodologies estimate AI's impact on economic growth, with projections suggesting an additional annual GDP growth of 0.8-1.3 percentage points over the next decade [10][11]. - The introduction of AI is expected to contribute approximately 9.8% to China's GDP by 2035, translating to an annual growth rate of about 0.8% [11]. Group 4: Scale Economics and Market Dynamics - The breakthrough of DeepSeek illustrates how algorithmic improvements can compensate for computational limitations, impacting the semiconductor industry [12][13]. - The distinction between scale economies in chip production and scale diseconomies in natural resources like coal highlights different market dynamics [13][14]. - The pricing power of large tech firms, driven by scale economies, raises questions about the sustainability of their monopolistic profits in the face of potential regulatory changes [14][15]. Group 5: Open Source and Competitive Landscape - The open-source model of AI development in China is reshaping the global competitive landscape, enhancing China's influence in AI and prompting adjustments in strategies by Western firms [15][16]. - The energy consumption of AI technologies poses a significant concern, with the contrasting approaches to energy sources in China and the US potentially impacting future AI applications [16]. Group 6: Creative Destruction and Market Risks - The high valuations of AI-related stocks may stem from overly optimistic long-term profit growth expectations, which could lead to unsustainable stock prices [17]. - The potential for a market correction exists, driven by changes in the semiconductor industry and the realization that AI applications may not meet current optimistic projections [17].

Group 1 - Microsoft is facing a significant issue with a surplus of GPUs that are currently idle due to a lack of power and space in data centers [4][5][6] - The primary challenge is not the supply of chips but rather the infrastructure needed to support their operation, particularly in terms of power supply and data center readiness [7][9] - The demand for electricity has surged in the past five years, driven by the rapid expansion of AI and cloud computing, outpacing the capacity of utility companies to meet this demand [11][12] Group 2 - Data center developers are increasingly opting for "behind-the-meter" power solutions to bypass public grids and meet their energy needs directly [12][14] - Solar energy is seen as a flexible and quick-to-deploy solution, but its implementation still takes considerable time, which does not align with the fast-paced demands of AI [14] - Concerns exist that if AI demand slows down, the investments in power plants and energy storage to support AI may become underutilized, although some industry leaders believe AI energy needs will continue to grow [15][17] Group 3 - Microsoft has decided not to stockpile single-generation GPUs due to the risk of obsolescence and depreciation if they cannot be powered in time [15][19] - The shift in focus from chip performance to energy efficiency is becoming more relevant as power shortages become a critical issue [17][19] - Microsoft announced plans to invest $8 billion in data centers and AI projects in the Middle East over the next four years, indicating a strategic move to regions with abundant energy resources [20]

Core Insights - AI is not leading to job losses but rather increasing the importance of radiologists, with demand for their services rising significantly [2][3][22] - The phenomenon of "Jevons Paradox" suggests that increased efficiency from AI can lead to greater consumption and job creation rather than reduction [4][12] - The "Baumol's Cost Disease" indicates that as some industries become more profitable due to efficiency gains, wages in other sectors must rise to retain talent, leading to increased costs across the board [16][18] AI in Radiology - Over 700 radiology AI models have received FDA approval, representing more than 75% of medical AI devices [22] - By 2025, the average salary for radiologists in the U.S. is projected to reach $520,000, making it the second-highest paid medical specialty [2][22] - The number of positions in diagnostic radiology is at a record high, with a vacancy rate also reaching historical levels [22] Economic Implications - The rise in AI efficiency is expected to increase service costs in unrelated sectors, as wages rise to compete for labor [18][20] - AI's impact on productivity may lead to a paradox where jobs that cannot be automated become more valuable, as they represent the final human touch in processes [21][23] - The overall economic growth driven by AI may lead to a situation where even low-skill jobs, like dog walking, become more expensive due to rising living costs [20][21] Future Workforce Dynamics - As AI takes over 99% of tasks, the remaining human roles may become highly specialized and valuable, leading to a unique labor market [21][23] - The potential for new job categories may emerge, focusing on tasks that require human presence or decision-making, despite the automation of many processes [23][24] - The ongoing challenge will be to enhance productivity while managing the societal changes that come with technological advancements [24]

Core Viewpoint - Microsoft is facing a significant challenge with a surplus of GPUs that are currently idle due to insufficient power supply and inadequate data center infrastructure [1][4][36]. Group 1: Power Supply Issues - The primary issue is not a surplus of chips but rather the lack of power capacity and the speed at which data centers can be built near power sources [2][4]. - Microsoft has a large number of NVIDIA AI chips that are currently unused due to power shortages [3][5]. - The demand for electricity has surged in the past five years, driven by the rapid expansion of AI and cloud computing, outpacing utility companies' capacity to generate additional power [15][16]. Group 2: Infrastructure Challenges - There is a significant lag in the supply side of electricity generation, with traditional power plants taking years to come online, while AI industry expansion occurs on a quarterly basis [17][18]. - Data center developers are increasingly opting for "behind-the-meter" power solutions to bypass public utilities and meet their energy needs [17]. - The construction of data centers and their associated power and cooling systems is not keeping pace with the actual demand for AI computing power [18][20]. Group 3: Future Outlook and Strategies - There is a belief among some industry leaders that AI's electricity demand will continue to grow, leading to more applications and increased overall demand [24][26]. - Microsoft has decided not to stockpile a single generation of GPUs due to the risk of obsolescence and depreciation over time [30][31][32]. - The industry is shifting focus towards energy-efficient chips as power supply becomes a more pressing concern than chip availability [39]. Group 4: Investment and Expansion - Microsoft has received approval to export NVIDIA chips to the UAE for building data centers necessary for AI model training [41]. - The company plans to invest $8 billion over the next four years in the Gulf region for data centers, cloud computing, and AI projects, indicating a shift of AI infrastructure from Silicon Valley to energy-rich emerging markets [42][43].

Core Viewpoint - Microsoft is facing an unprecedented issue with a surplus of GPUs that are idly stored due to insufficient power supply and data center infrastructure [1][2][3][4]. Group 1: Current Challenges - The primary challenge is not a surplus of computing power but rather the lack of electrical capacity and the inability to quickly build data centers close to power sources [2][4]. - Microsoft has a significant number of NVIDIA AI chips that are currently unused due to these infrastructure limitations [3][4]. - The industry is experiencing a collective challenge where energy and infrastructure must match the growing demand for AI computing power [10][11]. Group 2: Energy Supply Issues - The overall electricity demand in the U.S. has surged over the past five years, driven by the rapid expansion of AI and cloud computing data centers [15][16]. - Traditional power plants take years to develop, while the AI industry expands at a quarterly pace, leading to a mismatch in supply and demand [17][18]. - Many data center developers are opting for "behind-the-meter" power solutions to bypass public utilities and meet their energy needs [17]. Group 3: Future Outlook - There are concerns that if AI demand slows down, the investments in power plants and energy storage could become underutilized [22]. - However, some industry leaders believe that AI's energy demand will only continue to grow, leading to more applications and higher overall demand [24][25]. - The call for increased power generation capacity is seen as a strategic asset for AI development, with suggestions for the U.S. government to add 100 gigawatts of power annually [28]. Group 4: Strategic Adjustments - Microsoft has decided not to stockpile single-generation GPUs due to the risk of obsolescence and depreciation if they cannot be powered [30][32][33]. - The industry is shifting focus from peak performance to energy efficiency, as the limitations have transitioned from computing power to energy supply [39][40]. - Microsoft has announced plans to invest $8 billion in data centers and AI projects in the Gulf region over the next four years, indicating a shift in AI infrastructure towards energy-rich emerging markets [43][44].

Core Insights - The price of AI models has dramatically decreased, with costs dropping by up to 280 times for certain models over the past year, indicating a significant technological deflation in the AI sector [6][15][41] - As AI becomes cheaper, the cost of human labor in sectors that cannot be automated is rising, creating a paradox where technology deflation leads to human inflation [15][30][41] Price Dynamics - The cost of using models like GPT-3.5 has fallen from approximately $20 per million tokens to just $0.07, showcasing a free-fall in AI model pricing [6][15] - Different levels of models exhibit varying rates of price decline, with LLM inference costs decreasing at a rate comparable to Moore's Law, approximately tenfold annually [9][30] Economic Implications - The decline in AI model prices is leading to increased consumption and reliance on AI technologies, as businesses and individuals utilize these tools more frequently [21][24] - The rising costs of services that cannot be automated, such as home care and repairs, reflect a shift in labor value, where jobs requiring human presence are becoming more expensive [25][30] Jevons Paradox - The phenomenon where increased efficiency in AI leads to greater consumption aligns with Jevons Paradox, suggesting that cheaper AI will result in higher usage rather than savings [16][21] - As AI becomes a ubiquitous resource, it is transforming from a luxury service to a public utility, increasing dependency on AI technologies [21][32] Labor Market Changes - The labor market is experiencing a revaluation, where jobs that can be automated see their value decrease, while those that require human interaction become more valuable [31][41] - The rising costs of skilled labor in sectors like home repair and healthcare highlight the economic theory that productivity gains in one area can lead to increased costs in less efficient sectors [30][31] Power Dynamics - The decreasing costs of AI models are not leading to a democratization of technology but rather to a concentration of power among a few major companies that control the AI ecosystem [32][39] - As AI becomes cheaper, the dependency on major players like OpenAI and Google increases, leading to a more centralized control over AI technologies and data [33][39] Future Outlook - The ongoing price decline of AI models is not just a technological shift but a fundamental reorganization of value, where human creativity and emotional intelligence become the new high-value assets [41][42] - The future landscape may not be characterized by AI replacing humans but rather by AI redefining the value of human contributions in various sectors [41][42]

Core Viewpoint - Microsoft is facing an unprecedented issue of having a surplus of GPUs that are idle due to power shortages and insufficient data center infrastructure to support their operation [2][5]. Group 1: Power Shortages and Infrastructure - The primary challenge is not the supply of chips but the availability of power and the speed at which data centers can be built near power sources [3][5]. - Microsoft has a significant number of Nvidia AI chips that are not being utilized because the existing infrastructure cannot support their operation due to power and cooling limitations [5][7]. - The overall power demand in the U.S. has surged in the past five years, driven by the rapid expansion of AI and cloud computing, which has outpaced the growth of power generation capacity [7][10]. Group 2: Industry Response and Future Outlook - Data center developers are increasingly opting for "behind-the-meter" power supply methods to bypass public grids and meet energy needs directly [7][8]. - There are concerns that if AI demand slows down, the investments in power plants and energy storage projects to support AI might become underutilized [8]. - However, some industry leaders believe that AI's power demand will continue to grow, leading to a call for increased power generation capacity as a strategic asset for AI [8][12]. Group 3: Strategic Shifts in Chip Production - Microsoft has decided not to stockpile single-generation GPUs due to the risk of obsolescence and depreciation if they remain unused for extended periods [8][10]. - The focus may shift from maximizing peak performance to developing energy-efficient chips, as the industry grapples with power limitations rather than chip shortages [10][11]. - Microsoft has announced plans to invest $8 billion in data centers and AI projects in the Middle East, indicating a shift of AI infrastructure to regions with abundant energy resources [12].

Core Insights - The focus of the artificial intelligence (AI) race is shifting from computing power to electricity supply, with industry leaders acknowledging the uncertainty surrounding future energy consumption for AI [2][4] - Microsoft CEO Satya Nadella highlighted that the biggest challenge is no longer chip shortages but rather the availability of electricity and the construction of data centers close to power sources [3][4] - OpenAI CEO Sam Altman emphasized the strategic dilemma faced by tech companies regarding energy contracts, as locking in long-term contracts could lead to losses if new energy technologies emerge [8][9] Group 1: Bottleneck Shift - The bottleneck in AI deployment has transitioned from acquiring advanced GPUs to securing adequate electricity supply [4] - The rapid increase in electricity demand for data centers in the U.S. has outpaced the capacity planning of utility companies, leading developers to seek alternative power solutions [4] Group 2: Energy Demand Uncertainty - There is significant uncertainty regarding the energy consumption required for AI, with both Altman and Nadella admitting they do not know the exact requirements [6][7] - Altman suggested a potential exponential growth in demand if the cost of AI units continues to decrease at a rapid pace, which could lead to a dramatic increase in energy needs [6][7] Group 3: Energy Gamble - The uncertainty in energy needs creates a dilemma for industry leaders, as they must decide whether to invest in current energy contracts or risk missing out on future opportunities [9] - Altman has invested in several energy startups to hedge against risks associated with energy supply and demand fluctuations [9] Group 4: Strategies for Adaptation - Tech companies are exploring solutions such as solar energy, which can be deployed more quickly and at lower costs compared to traditional natural gas plants [11] - The modular nature of solar technology allows for rapid assembly and deployment, aligning more closely with the fast-paced demands of the AI industry [11]

Core Viewpoint - Microsoft is facing an unprecedented issue with a surplus of GPUs that are idly stored due to insufficient power supply and space for data centers [1][3][4]. Group 1: Power Supply Issues - The primary challenge is not the surplus of chips but the lack of power capacity and the speed at which data centers can be built near power sources [2][5]. - Microsoft has a significant number of NVIDIA AI chips that are currently unused due to power shortages [3][4]. - The overall power demand has surged in the past five years, driven by the AI and cloud computing boom, outpacing utility companies' capacity planning [11][12]. Group 2: Infrastructure Development - The construction of traditional power plants takes several years, while the demand for AI capabilities is growing rapidly, leading data center developers to seek alternative power solutions [13][14]. - Many data center developers are adopting "behind-the-meter" power supply methods to bypass public grids and meet energy needs directly [13]. - The construction timelines for solar energy systems are also lengthy, making it challenging to keep pace with the rapid changes in AI demand [16][27]. Group 3: Strategic Adjustments - Microsoft has decided not to hoard single-generation GPUs due to the risk of obsolescence and depreciation over time [24][25]. - The company emphasizes the need for energy-efficient chips as power constraints become a more pressing issue than chip availability [31][32]. - The industry is shifting focus from peak performance to energy efficiency in chip production as power supply becomes the limiting factor [30][32]. Group 4: Future Investments - Microsoft has received approval to export NVIDIA chips to the UAE for building AI training data centers and plans to invest $8 billion in the Gulf region over the next four years [34]. - This move indicates a shift of AI infrastructure from Silicon Valley to emerging markets with abundant energy resources [34][35].