返回投资研究台 2026-07-25
Market Context

报告对应赛道与标的

11

2026-07-25 研究记录02:Hyperscaler capex 能转化,但电力最晚兑现

截至研究所工作日 2026-07-25,本地 shell 日期命令已确认该日期 [S0];我对研究记录01采取综合而非反转:AI capex 仍在转化为供应商收入和已下单 racks,但只有具备电力保障的部分才应被资本化为近端可部署容量。7月29-30日财报窗口关键,是因为 Microsoft 将在 2026-07-29 发布业绩 [S17],Meta 将在 2026-07-29 发布业绩 [S18],Amazon 将在 2026-07-30 发布业绩 [S19]。最先断裂的不是 GPU 订单,而是 time-to-power:并网、强电设备、变压器交付和园区内电气集成。这支持 th_T01th_T11,也只部分支持 th_p_59893da8:ASIC 份额会重分配芯片栈收入,但物理建设并未表现为需求平台期。

核心判断

清晰的转换率应分三层。第一,供应商收入转换率高:Alphabet 表示 2026 年 capex 约 60% 用于服务器、40% 用于数据中心与网络 [S2];Microsoft 表示 FY2026 Q3 capex 约三分之二投向 GPU 和 CPU 等短寿命资产 [S3]。第二,rack 转换率中等:按 NVIDIA GB300 NVL72 参考形态,一个 rack 配 72 颗 Blackwell Ultra GPU [S6],单 rack 功耗最高可达 142 kW [S7],因此 headline capex 可以先买到大量 racks,但这些 racks 不等于已经具备通电容量。第三,电力保障容量转换率最低:JLL 给出的 hyperscale shell and core 成本为每 MW USD11.3mn,AI technology fit-out 最高每 MW USD25mn [S9];而美国并网周期的中位数以年为单位,不以季度为单位 [S10]。

配置含义是:不要在 7月29-30日窗口前简单做空整个 AI infrastructure 链 [S17][S18][S19]。仍应偏好能把订单转成收入的瓶颈供应商,但对平台公司和二线数据中心概念股维持折价,除非它们能同时证明 free cash flow。Alphabet 已经展示了矛盾:2026 Q2 property-and-equipment purchases 为 USD44.924bn,free cash flow 为负 USD5.855bn [S1],同时管理层把 2026 年 capex 指引上调至 USD195-205bn [S2]。

转换地图

capex 去向 截至 2026-07-25 的证据 转换判断
加速器、CPU、内存、网络和服务器收入 Alphabet 的 2026 年 capex 结构约 60% 为服务器 [S2];Microsoft FY2026 Q3 约三分之二 capex 为 GPU 和 CPU 等短寿命资产 [S3]。 最快转换。 对已经投向服务器和加速器的 capex,60-67% 的 compute-like mix 可作为近端收入传导区间 [自测算:以 Alphabet 60% 与 Microsoft 约 67% 为可观察锚点]。
实体 AI racks NVIDIA GB300 NVL72 每 rack 有 72 颗 GPU [S6],单 rack 功耗最高 142 kW [S7]。 先可下单,后可完全使用。 1 GW IT load 可支持约 7,042 个 GB300 NVL72 racks、约 507,000 颗 GPU [自测算:1,000,000 kW / 每 rack 142 kW;racks * 72 GPU]。若 PUE 为 1.15,1 GW 公用电力可支持约 870 MW IT load、约 6,126 个 racks、约 441,000 颗 GPU [自测算:1,000 MW / 1.15 / 0.142 MW;racks * 72 GPU]。
数据中心土建、冷却和电气 fit-out JLL 给出的 shell and core 为每 MW USD11.3mn,AI technology fit-out 最高每 MW USD25mn [S9]。 大额资金吸收项。 若把研究记录01引用的四大 hyperscaler 2026 年 USD725bn capex 计划中的 60% 视为 compute-oriented bucket,即 USD435bn [S21][自测算:USD725bn * 60%],则在每 GW USD35-49bn 成本下对应 8.9-12.4 GW 的 AI-capacity equivalent [S8][自测算:USD435bn / 每 GW USD49bn 至 USD35bn;USD35bn 下限也与 JLL 每 MW USD11.3mn + USD25mn、约等于每 GW USD36.3bn 的口径相近]。
已通电、具备 utility-backed 的容量 LBNL 报告称,截至 2023 年底,美国并网队列中等待接入的发电与储能容量超过 2,600 GW [S10]。DOE 把数据中心负荷增长列为国家层面的电网规划议题 [S11]。 最慢转换。 没有锁定电力、变压器和并网研究的 capex 应被视为 backlog,而不是 live capacity。物理队列对应多年电网流程,因此 2026 年 capex 不能直接假定为 2026 年已通电容量 [S10][S11]。

供应商收入快于容量兑现

NVIDIA FY2027 Q1 data-center revenue 为 USD75.2bn,同比增长 92% [S14]。Broadcom 表示 FY2026 Q2 AI semiconductor revenue 为 USD10.8bn,同比增长 143%,并预计 FY2026 Q3 AI semiconductor revenue 为 USD16.0bn [S15]。Micron FY2026 Q3 revenue 为 USD41.46bn,operating cash flow 为 USD25.39bn,Cloud Memory Business Unit revenue 为 USD13.769bn,Core Data Center Business Unit revenue 为 USD11.524bn [S16]。这些数据支持研究记录01的判断:capex 接收端仍在把建设周期转为收入。

需要警惕的是结构。如果 Microsoft 在 2026-07-29 确认 capex 中“约三分之二”仍是短寿命算力资产 [S3][S17],NVIDIA、HBM、网络和 ODM 的收入可见度仍强。如果 Meta 在 2026-07-29 维持 USD125-145bn 的 2026 年 capex 区间 [S4][S18],AI factory 的物理建设管线仍未断。如果 Amazon 在 2026-07-30 的更新显示 AWS 增长但现金转化继续承压 [S5][S19],则会确认 th_T11:AI 收入可以增长,但 free cash flow 仍会被折价。

最先断裂的瓶颈

最可能先断的是电力交付,而不是硅供给。硅仍紧,但证据显示供应商已在把 capex 转成收入 [S14][S15][S16]。电力基础设施弹性更弱:Grid Strategies 预计美国 2025-2029 年用电需求增长 15.8%,而其 2024 年预测的前一个五年期增长只有 4.7% [S12]。变压器约束也符合电力瓶颈判断:pv magazine 引用的 2025 年行业调查显示,变压器交期最高可达四年 [S13]。

因此,财报电话会的实际问题不是“是否继续花钱”,而是“多少支出已有电力保障”。Alphabet 2026 Q2 负 free cash flow 已说明,融资层先于收入层感受到压力 [S1]。NVIDIA 与 ERCOT、Utilidata 合作做 AI-powered grid management,也说明 AI factory 负荷平滑已成为问题 [S20];如果电网容量没有摩擦,这类工作就不重要。

投资映射

多头应集中在瓶颈收入接收方:先进算力、HBM、高速网络、液冷、电气设备、备用电源、变压器和电网软件。对只披露 capex、但不披露 power reservations、energized MW、GPU utilization 和 AI revenue per watt 的平台公司,不应给满估值。若 Microsoft、Meta 和 Amazon 维持 capex,7月29-30日财报窗口仍可能托住硬件复合体 [S17][S18][S19];但下一段行情应奖励能把美元转成通电 racks的公司,而不是只会宣布更大预算的公司。

对 A 股,最直接含义是优先具备可验证订单的电力、电网设备、液冷、光模块和先进封装标的;纯 AI 数据中心概念更像融资久期资产。这个结论细化了 th_T01:约束不再是“hyperscaler 能不能买芯片”,而是“它们能否足够早锁定电力,使芯片开始赚钱”。

交接

推荐下一位分析师:utilities-analyst

跟进主题:AI 数据中心部署的区域电力可得性与电网设备定价权。

跟进问题:哪些美国、中国与亚洲电力市场真正能在 2026-2028 年交付 AI clusters 所需的 energized MW,哪些上市公用事业或电网设备供应商具备最清晰的收入传导?

报告页脚:工作日 2026-07-25,Asia/Singapore [S0]。

资料来源 / Sources

[S0] Local shell command, 系统日期锚 returned 2026-07-25 — local workspace command, no URL

[S1] Alphabet Inc., "GOOG Exhibit 99.1 Q2 2026" — https://s206.q4cdn.com/479360582/files/doc_news/2026/Jul/22/attachments/2026q2-alphabet-earnings-release.pdf

[S2] Investing.com, "Alphabet Q2 2026 slides: 24% revenue growth, cloud surges despite capex" — https://www.investing.com/news/company-news/alphabet-q2-2026-slides-24-revenue-growth-cloud-surges-despite-capex-93CH-4807148

[S3] Microsoft, "Microsoft Fiscal Year 2026 Third Quarter Earnings Conference Call" — https://www.microsoft.com/en-us/investor/events/fy-2026/earnings-fy-2026-q3

[S4] Meta Platforms, "Meta Reports First Quarter 2026 Results" — https://investor.atmeta.com/investor-news/press-release-details/2026/Meta-Reports-First-Quarter-2026-Results/default.aspx

[S5] Amazon, "Amazon.com Announces First Quarter Results" — https://ir.aboutamazon.com/news-release/news-release-details/2026/Amazon-com-Announces-First-Quarter-Results/default.aspx

[S6] NVIDIA, "NVIDIA GB300 NVL72" — https://www.nvidia.com/en-us/data-center/gb300-nvl72/

[S7] NVIDIA Docs, "NVIDIA GB300 NVL72 Enterprise Reference Architecture: Components" — https://docs.nvidia.com/enterprise-reference-architectures/nvl72-ai-factory/latest/components.html

[S8] Business Insider, "Big Tech earnings season and the capex spiral" — https://www.businessinsider.com/big-tech-spending-capex-earnings-season-memory-prices-ai-2026-7

[S9] JLL, "2026 Market Outlook for Global Data Centers" — https://www.jll.com/en-us/insights/market-outlook/data-center-outlook

[S10] Lawrence Berkeley National Laboratory, "Queued Up: Characteristics of Power Plants Seeking Transmission Interconnection As of the End of 2023" — https://emp.lbl.gov/queues

[S11] U.S. Department of Energy, "Powering America's AI Future -- Data Center Resource Hub" — https://www.energy.gov/powering-americas-ai-future-data-center-resource-hub

[S12] Grid Strategies, "Strategic Industries Surging: Driving US Power Demand" — https://gridstrategiesllc.com/wp-content/uploads/National-Load-Growth-Report-2024.pdf

[S13] pv magazine USA, "U.S. transformer market faces severe supply constraints as lead times extend to four years" — https://pv-magazine-usa.com/2026/05/11/u-s-transformer-market-faces-severe-supply-constraints-as-lead-times-extend-to-four-years/

[S14] NVIDIA, "NVIDIA Announces Financial Results for First Quarter Fiscal 2027" — https://nvidianews.nvidia.com/news/nvidia-announces-financial-results-for-first-quarter-fiscal-2027

[S15] Broadcom, "Broadcom Inc. Announces Second Quarter Fiscal Year 2026 Financial Results" — https://investors.broadcom.com/news-releases/news-release-details/broadcom-inc-announces-second-quarter-fiscal-year-2026-financial

[S16] Micron, "Micron Technology, Inc. Reports Record Results for the Third Quarter of Fiscal 2026" — https://investors.micron.com/news-releases/news-release-details/micron-technology-inc-reports-record-results-third-quarter

[S17] Microsoft, "Microsoft Earnings Release Available on Investor Relations Website" — https://www.microsoft.com/en-us/Investor/events/FY-2026/earnings-fy-2026-q4

[S18] Meta Platforms, "Meta to Announce Second Quarter 2026 Results" — https://investor.atmeta.com/investor-news/press-release-details/2026/Meta-to-Announce-Second-Quarter-2026-Results/default.aspx

[S19] Amazon, "Amazon.com to Webcast Second Quarter 2026 Financial Results Conference Call" — https://ir.aboutamazon.com/news-release/news-release-details/2026/Amazon-com-to-Webcast-Second-Quarter-2026-Financial-Results-Conference-Call/default.aspx

[S20] NVIDIA, "NVIDIA, ERCOT and Utilidata Collaborate on AI-Powered Grid Management" — https://blogs.nvidia.com/blog/ercot-utilidata-ai-grid-management/

[S21] IG, "Big Tech Q2 2026 earnings -- the AI capex question and what UK investors should watch" — https://www.ig.com/uk/trading-strategies/big-tech-q2-2026-earnings---the-ai-capex-question-and-what-uk-in-260716