The second half of AI investment: not only cloud vendors, but also an unknown giant. Bank of America gave a shot of CSP today, and it is expected that major companies such as Google and Amazon will increase capital expenditures in the future, and governments around the world will also increase investment. The market used to focus on the computing investment of Internet and cloud manufacturers such as Meta, Microsoft, Google and Amazon, but now more important increment comes from "sovereign AI" and AI search. Sovereign AI refers to data center investments promoted by governments and local enterprises of various countries for data security, industrial autonomy, and AI infrastructure construction; AI search comes from the continuous consumption of inference computing power by applications such as Gemini, Bedrock, ChatGPT, etc. This means that the demand for AI infrastructure is no longer solely driven by a few major American companies, but is beginning to spread to national level projects, enterprise level deployments, and new AI application scenarios. Bank of America predicts that global AI capital expenditures will continue to rise from 2026 to 2028. Alphabet, Meta, and Amazon have raised their Capex expectations for 2026 to $187 billion, approximately $130 billion to $145 billion, and approximately $159 billion, respectively. The reason behind the upgrade is very direct: AI training, inference, search, and enterprise AI services are all consuming more GPU and data center resources, and cloud vendors also need to build capacity in advance to avoid future supply shortages. From the perspective of data center costs, the report focuses on "capital expenditure per GW of computing power". The cost varies greatly among different companies: Amazon costs about $25 billion/GW, Microsoft costs about $37 billion/GW, and Meta costs about $45 billion/GW. Meta has the highest cost mainly because it undertakes more complete internal GPU construction and AI infrastructure, rather than simply renting external cloud resources. In the long run, the production capacity of Amazon, Google, and Meta is expected to significantly expand, with Meta increasing from about 5GW in 2025 to about 11.2GW in 2027. Amazon and Google will also enter the double-digit GW level. Bank of America believes that Meta's AI assets may be underestimated when measured solely by the ratio of enterprise value to data center capacity. The implied valuations given in the report are: Alphabet at approximately $110 billion/GW, Amazon at approximately $59 billion/GW, and Meta at approximately $4 billion/GW. In other words, the market is currently giving Meta a significant discount on its AI infrastructure, and may even view its computing power investment as a cost pressure rather than fully pricing it as a long-term asset. The core conclusion given by Bank of America is that Meta may be the "cheapest" target for AI capacity expansion. Although Capex is high and short-term profit margins are under pressure, it may release value in the long run through AI advertising, enterprise AI, internal efficiency improvement, and potential external computing power monetization. In contrast, Amazon benefits from improved cloud business profits and Bedrock demand, while Google benefits from Gemini and AI search, but both are more fully priced in the market. Reiterating Amazon's purchase, it is believed that Meta still has significant room for reassessment, and Google's long-term AI logic is valid, but the short-term response is more sufficient.