高盛预计阿里未来数年资本开支将达4600亿元，远高于该公司此前3800亿元的目标。报告认为，AI推理需求激增将推动阿里、字节等巨头持续加码投资，其中阿里押注企业级AI 云，字节则发力消费级应用。尽管技术效率不断提升，但AI需求的爆发令资本开支继续扩张。 硬·AI 作者 | 张雅绮 编辑 | 硬 AI 高盛认为，需求的爆炸式增长将继续推动云服务厂商的资本开支（Capex）上行。中国互联网巨头在AI领域的战略路径正日益分化，阿里巴巴凭借其全栈能力押 注企业级AI云市场，而字节跳动则在消费级（To-C）应用上全力冲刺。 高盛在23日的报告中，上调了对中国头部云厂商的资本开支预测， 预计阿里巴巴在2026至2028财年的合计资本开支将达到4600亿元人民币，远高于该公司此前 3800亿元的目标。 该行认为，AI推理需求的激增是支撑这一判断的核心逻辑，更高的计算效率反而可能提升资本开支向云收入的转化率，从而加速收入增长。 与此同时，巨头们的战略差异愈发明显。高盛指出， 阿里巴巴在外部AI云收入规模和企业级（To-B）服务方面处于领先地位， 其商业化路径更为清晰。 而字节 跳动则凭借旗下聊天机器人"豆包"在To-C领域和 ...

近期，中国公司在AI基础设施和计算效率方面取得多项突破。例如，阿里云发布了新的GPU池化系统 Aegaeon，据称可节省82%的GPU资源；DeepSeek的新OCR模型能将文本输入消耗的token减少90%。 然而，高盛认为，这些效率提升并不意味着资本开支会相应缩减。报告引用的数据显示，AI推理需求 和token消耗量正呈指数级增长。字节跳动近期宣布，其日均token消耗量已在9月份突破30万亿，相比4- 5月实现翻番，该体量已接近谷歌的43万亿。阿里巴巴此前也曾在云栖大会上透露，其AI推理需求每2-3 个月就会翻一番。 高盛预测，中国云服务提供商（CSP）在2025年第三季度的资本开支将同比增长50%。该行分析称，强 劲的推理需求是驱动云厂商持续投入的关键，而技术效率的提升有助于将这些投资更有效地转化为收入 增长，形成正向循环。 巨头路径分化：阿里To-B，字节To-C 高盛认为，需求的爆炸式增长将继续推动云服务厂商的资本开支（Capex）上行。中国互联网巨头在AI 领域的战略路径正日益分化，阿里巴巴凭借其全栈能力押注企业级AI云市场，而字节跳动则在消费级 （To-C）应用上全力冲刺。 高盛在23日的报告 ...

万国数据管理层称，此前签署的152MW超大规模订单预计将推动2025年下半年客户入驻速度显著加 快，尽管芯片供应存在不确定性，但对未来今年AI推理需求充满信心。 8月21日，据追风交易台消息，美银美林在最新研报中称，万国数据2025年二季度电话会传递出积极信 号，管理层预计预计2025年下半年客户入驻步伐将明显加快，主要受惠于此前签订的152MW超大规模 订单交付，这将直接转化为收入增长。 美银美林称，尽管芯片供应不确定性导致AI相关新订单仍相对谨慎，但管理层对未来几年AI需求前景 保持乐观态度，预期推理需求将成为关键驱动力。 研报称，万国数据管理层在电话会议中明确表示，预计2025年三、四季度客户入驻速度将较前两个季度 明显提升。这一预期主要基于此前签署的152MW超大规模客户订单将在下半年开始交付。 据美银美林分析，万国数据目前拥有185,000平方米的积压项目(包括在服务和在建项目)，管理层计划在 2025年下半年交付其中35%，在2026年再交付35%。这一交付时间表显示公司具备稳定的收入增长基 础。 据见闻此前文章，最新财报显示，数据中心业务增长，万国数据2024年四季度收入增长9.1%，净利润 同 ...

Core Viewpoint - Morgan Stanley highlights that despite short-term financial pressure from the "H20 sales restrictions," the market may be underestimating the long-term potential of explosive AI inference demand, which, combined with improvements in Blackwell architecture supply, could lead to a performance acceleration for NVIDIA in the second half of the year [1][6]. Group 1: Short-term Impact of H20 Policy - The change in the U.S. H20 export policy has significantly impacted NVIDIA's revenue, with estimates suggesting a 10% quarter-over-quarter revenue decline risk for the second quarter of 2025 [2][3]. - This policy shift has forced NVIDIA to announce a record $5.5 billion inventory write-down, marking the largest in semiconductor industry history [2][3]. Group 2: Supply Chain Improvements - Despite the short-term impact of the H20 sales restrictions, improvements in the Blackwell supply chain and explosive inference demand are more noteworthy [5]. - The supply chain for Blackwell racks is improving, with three major ODM manufacturers reaching a monthly production of approximately 1,500 racks in April, expected to continue growing throughout the year [5]. - Analysts believe that the market has underestimated the importance of non-rack forms of Blackwell chips, as many customers are not ready to adopt liquid cooling technology or ARM processors [5]. Group 3: Long-term Growth Drivers - The explosive growth in AI inference demand is viewed as the fundamental driver of NVIDIA's long-term value [6]. - Major cloud providers like Microsoft, Amazon, and Google have reported "unexpected increases in Token usage," reflecting a surge in end-user demand for AI models [6][7]. - The demand for inference is driven by real business needs rather than venture-capital-supported training cluster construction, providing critical validation for NVIDIA's revenue sustainability [7]. Group 4: Performance Outlook - Morgan Stanley maintains an "overweight" rating on NVIDIA with a target price of $160, indicating a potential upside of approximately 21.87% from the current stock price of $131.29 [8]. - The path for a performance rebound in the second half of the year is clear, with recent issues being adequately warned and addressed [8]. - NVIDIA's revenue for the April quarter is estimated at approximately $42.2 billion, below the guidance of $43 billion, while the July quarter is projected at $43.5 billion, considering a $4-5 billion reduction in H20-related revenue [9].