In the past two years, my focus on AI has been electricity. Although the returns on power-related assets are not as exaggerated as those of storage, chips, and servers, I have always believed that electricity is the most fundamental, foundational, and irreplaceable demand for AI, as the core of all computing power ultimately goes back to electricity.
The United States is one of the countries where AI is developing the fastest and where the demand for electricity is most apparent. Data centers do not use electricity like ordinary residents; they require high-load power that operates stably 24 hours a day, needing fast grid connection, substations, large transformers, gas turbines, nuclear power, energy storage, backup power, and cooling systems.
The problem is that AI data centers can be built in 18 to 36 months, while the construction cycle for transmission lines, substations, large transformers, gas turbines, and dispatchable power sources often takes 4 to 8 years, or even longer. This means that the pace of capital expenditure for AI has already surpassed the physical expansion rate of the U.S. power system.
(This data has been shared in previous BOA reports.)
If AI data centers become the most certain increment in electricity demand over the next decade, then companies with power generation assets, transmission assets, substation equipment, grid connection capabilities, on-site power supply capabilities, and electrical engineering capabilities will no longer just be traditional utilities, but part of the AI infrastructure.
AI is not just about buying GPUs; AI is about rebuilding an energy system centered around computing power. A 100MW AI data center is not just a server room but a small power system. It requires upstream power generation, midstream transmission, downstream substations, campus distribution, UPS, PDU, busbars, backup generators, liquid cooling, and a large amount of electrical construction.
One can even imagine that in the future, a global electricity shortage could become a consensus because AI will not only happen in the U.S. Europe aims to build sovereign AI, the Middle East wants to develop AI data centers, China is also building computing power infrastructure, and Southeast Asia is taking on data center migrations.
Moreover, training models is merely a short-term demand; the demand for inference is the long-term load. If every search, every video generation, every autonomous driving judgment, every robot execution, and every enterprise AI agent invocation corresponds to actual computing power consumption, then the demand for electricity in AI is not a one-time construction demand but a long-term, continuous, and rigid energy demand.
Therefore, I believe that the end of computing power is electricity.
In the past two years, the main line of AI trading has been GPUs, HBM, servers, and optical modules; over the next few years, the main line of AI trading is likely to gradually shift down to electricity: power generation, transmission, substations, distribution, energy storage, liquid cooling, backup power, grid connection projects, and on-site energy systems.
#Bitget has arrived as VIP! Crypto, U.S. stocks, CFD, global opportunities consolidated in one stop.
免责声明：本文章仅代表作者个人观点，不代表本平台的立场和观点。本文章仅供信息分享，不构成对任何人的任何投资建议。用户与作者之间的任何争议，与本平台无关。如网页中刊载的文章或图片涉及侵权，请提供相关的权利证明和身份证明发送邮件到support@aicoin.com，本平台相关工作人员将会进行核查。
